Cisco Data Intelligence Platform with Cloudera Enterprise Data Hub 6.2 and Cloudera Data Science Workbench 1.5

Deployment Guide for the Cisco Data Intelligence Platform with Cloudera Enterprise Data Hub 6.2

 

Last Updated: January 8, 2020

NOTE: Works with document’s Advanced Properties “Last Updated” property. Click File | Properties | Advanced Properties | Custom.

About the Cisco Validated Design Program

The Cisco Validated Design (CVD) program consists of systems and solutions designed, tested, and documented to facilitate faster, more reliable, and more predictable customer deployments. For more information, go to:

http://www.cisco.com/go/designzone.

ALL DESIGNS, SPECIFICATIONS, STATEMENTS, INFORMATION, AND RECOMMENDATIONS (COLLECTIVELY, "DESIGNS") IN THIS MANUAL ARE PRESENTED "AS IS," WITH ALL FAULTS.  CISCO AND ITS SUPPLIERS DISCLAIM ALL WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE WARRANTY OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT OR ARISING FROM A COURSE OF DEALING, USAGE, OR TRADE PRACTICE.  IN NO EVENT SHALL CISCO OR ITS SUPPLIERS BE LIABLE FOR ANY INDIRECT, SPECIAL, CONSEQUENTIAL, OR INCIDENTAL DAMAGES, INCLUDING, WITHOUT LIMITATION, LOST PROFITS OR LOSS OR DAMAGE TO DATA ARISING OUT OF THE USE OR INABILITY TO USE THE DESIGNS, EVEN IF CISCO OR ITS SUPPLIERS HAVE BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.

THE DESIGNS ARE SUBJECT TO CHANGE WITHOUT NOTICE.  USERS ARE SOLELY RESPONSIBLE FOR THEIR APPLICATION OF THE DESIGNS.  THE DESIGNS DO NOT CONSTITUTE THE TECHNICAL OR OTHER PROFESSIONAL ADVICE OF CISCO, ITS SUPPLIERS OR PARTNERS.  USERS SHOULD CONSULT THEIR OWN TECHNICAL ADVISORS BEFORE IMPLEMENTING THE DESIGNS.  RESULTS MAY VARY DEPENDING ON FACTORS NOT TESTED BY CISCO.

CCDE, CCENT, Cisco Eos, Cisco Lumin, Cisco Nexus, Cisco StadiumVision, Cisco TelePresence, Cisco WebEx, the Cisco logo, DCE, and Welcome to the Human Network are trademarks; Changing the Way We Work, Live, Play, and Learn and Cisco Store are service marks; and Access Registrar, Aironet, AsyncOS, Bringing the Meeting To You, Catalyst, CCDA, CCDP, CCIE, CCIP, CCNA, CCNP, CCSP, CCVP, Cisco, the Cisco Certified Internetwork Expert logo, Cisco IOS, Cisco Press, Cisco Systems, Cisco Systems Capital, the Cisco Systems logo, Cisco Unified Computing System (Cisco UCS), Cisco UCS B-Series Blade Servers, Cisco UCS C-Series Rack Servers, Cisco UCS S-Series Storage Servers, Cisco UCS Manager, Cisco UCS Management Software, Cisco Unified Fabric, Cisco Application Centric Infrastructure, Cisco Nexus 9000 Series, Cisco Nexus 7000 Series. Cisco Prime Data Center Network Manager, Cisco NX-OS Software, Cisco MDS Series, Cisco Unity, Collaboration Without Limitation, EtherFast, EtherSwitch, Event Center, Fast Step, Follow Me Browsing, FormShare, GigaDrive, HomeLink, Internet Quotient, IOS, iPhone, iQuick Study,  LightStream, Linksys, MediaTone, MeetingPlace, MeetingPlace Chime Sound, MGX, Networkers, Networking Academy, Network Registrar, PCNow, PIX, PowerPanels, ProConnect, ScriptShare, SenderBase, SMARTnet, Spectrum Expert, StackWise, The Fastest Way to Increase Your Internet Quotient, TransPath, WebEx, and the WebEx logo are registered trademarks of Cisco Systems, Inc. and/or its affiliates in the United States and certain other countries.

All other trademarks mentioned in this document or website are the property of their respective owners. The use of the word partner does not imply a partnership relationship between Cisco and any other company. (0809R)

© 2020 Cisco Systems, Inc. All rights reserved.

 

Table of Contents

Executive Summary. 7

Solution Overview.. 8

Introduction. 8

Audience. 8

Purpose of this Document 8

What’s New in this Release?. 9

What’s Next?. 9

Solution Summary. 10

Cisco Data Intelligence Platform.. 10

Reference Architecture. 11

Data Lake Reference Architecture. 12

AI Computing Farm Reference Architecture. 12

Tiered Storage. 13

Scaling the Solution. 15

Technology Overview.. 19

Cisco UCS Integrated Infrastructure for Big Data and Analytics. 19

Cisco UCS Manager 19

Cisco UCS 6300 Series Fabric Interconnects. 19

Cisco UCS C-Series Rack-Mount Servers. 20

Cisco UCS C240 M5 Rack-Mount Server 20

Cisco UCS S3260 Storage Servers. 21

Cisco UCS Virtual Interface Cards (VICs) 24

NVIDIA GPU.. 25

NVIDIA CUDA. 25

Cloudera Enterprise Data Hub and Hortonworks Data Platform.. 25

Cloudera (CDH 6.2.0) 26

Hortonworks Data Platform.. 29

Apache Ambari 29

HDP for Data Access. 30

Submarine. 31

Docker Containerization. 32

YARN Support for Docker 33

Kubernetes. 34

NVIDIA Docker 34

GPU Pooling and Isolation. 35

Red Hat Ansible Automation. 35

Solution Design. 36

Requirements. 36

Physical Topology. 36

Port Configuration on Fabric Interconnect 37

Server Configuration and Cabling for Cisco UCS C240 M5. 37

Software Distributions and Versions. 38

Deployment Hardware and Software. 40

Cisco Unified Computing System Configuration. 40

Configure Cisco UCS Fabric Interconnect 40

Configure Fabric Interconnects for a Cluster Setup. 40

Configure Cisco UCS Manager 42

Create Service Profile Template. 71

Install Red Hat Enterprise Linux 7.6. 78

Post OS Install Configuration. 89

Configure /etc/hosts. 89

Set Up Passwordless Login. 90

Create a Red Hat Enterprise Linux (RHEL) 7.6 Local Repository. 91

Create the Red Hat Repository Database. 92

Set Up Ansible. 92

Install httpd. 95

Disable the Linux Firewall 95

Set Up All Nodes to use the RHEL Repository. 96

Upgrade the Cisco Network Driver for VIC1387. 96

Set Up JAVA. 97

Enable Syslog. 99

Set the ulimit 99

Disable SELinux. 100

Set TCP Retries. 100

Disable IPv6 Defaults. 101

Disable Swapping. 101

Disable Memory Overcommit 101

Disable Transparent Huge Pages. 102

NTP Configuration. 102

Install Megaraid StorCLI 104

Configure the Filesystem for NameNodes and DataNodes. 105

Delete Partitions. 106

Cluster Verification. 106

Install Cloudera. 109

Prerequisites for CDH Installation. 109

Set Up the Local Parcels for CDH 6.2.0. 110

Download Parcels. 110

Set Up the MariaDB Database for Cloudera Manager 113

Install the MySQL JDBC Driver 115

Cloudera Manager Installation. 116

Set Up the Cloudera Manager Server Database. 116

Install Cloudera Manager 116

Install Cloudera Enterprise Data Hub (CDH6) 118

Scale the Cluster 130

Enable High Availability. 131

HDFS High Availability. 131

Configure Hive Metastore to Use HDFS High Availability. 135

Configure Hue to Work with HDFS High Availability. 135

YARN High Availability. 136

Configure Yarn (MR2 Included) and HDFS Services. 138

Configure Spark. 139

Tune Resource Allocation for Spark. 139

Submit a Job. 140

Shuffle Performance Improvement 141

Improve Serialization Performance. 141

Spark SQL Tuning. 141

Compression for Hive. 142

Change the Log Directory for All Applications. 142

Summary. 145

For More Information. 145

Bill of Materials. 146

Appendix. 152

Configure Cisco Boot Optimized M.2 RAID Controller 152

Configure Disk Group Policy. 152

Configure Storage Profile. 155

Apply Storage Profile in Service Profile Template. 159

Install RHEL 7.6 on Cisco Optimized M.2 RAID Controller 160

Configure Data Drives on Name Node and Other Management Nodes. 170

Configure Data Drives on Data Nodes. 172

Cloudera Data Science Workbench (CDSW) 172

Install Prerequisites for CUDA. 172

Verify Supported Version of Linux. 173

GCC Installation. 174

Install Kernel Headers and Installation Packages. 174

Install dkms. 175

Install NVIDIA GPU Drivers. 175

Install CUDA. 176

Download and Setup NVIDIA CUDA Deep Neural Network Library (cuDNN) 178

Installation Prerequisites for CDSW.. 179

Set Up a Wildcard DNS Subdomain. 179

Supported JDK Version. 180

IP Tables and Security on CDSW Nodes. 180

Configure Block Devices. 181

Download and Install CDSW with Cloudera Manager 181

Add Apache Spark 2 Service. 184

Add the Cloudera Data Science Workbench Service. 186

Create the Administrator Account 189

Non-Kerberized Clusters. 190

Use GPUs for Cloudera Data Science Workbench Workloads. 191

Enable GPU with CDSW.. 191

Create a Custom CUDA-Capable Engine Image. 195

Designate a Server for Docker and Start the Registry. 195

Allocate GPUs for Sessions and Jobs. 198

About the Authors. 201

Acknowledgements. 201

 

 

Data scientists are constantly searching for newer techniques and methodologies that can unlock the value of big data and distill this data further to identify additional insights which could transform productivity and provide business differentiation.

One such area is Artificial Intelligence/Machine Learning (AI/ML), which has seen tremendous development with bringing in new frameworks and new forms of compute (CPU, GPU and FPGA) to work on data to provide key insights. While data lakes have historically been data intensive workloads, these advancements in technologies have led to a new growing demand of compute intensive workloads to operate on the same data.

While data scientists want to be able to use the latest and greatest advancements in AI/ML software and hardware technologies on their datasets, the IT team is also constantly looking at enabling these data scientists to be able to provide such a platform to a data lake. This has led to architecturally siloed implementations. When data, which is ingested, worked, and processed in a data lake, needs to be further operated by AI/ML frameworks, it often leaves the platform and must be on-boarded to a different platform to be processed. This would be fine if this demand is seen only on a small percentage of workloads. However, AI/ML workloads working closely on the data in a data lake are seeing an increase in adoption. For instance, data lakes in customer environment are seeing deluge of data from new use cases such as IoT, autonomous driving, smart cities, genomics and financials, who are all seeing more and more demand of AI/ML processing of this data.

IT is demanding newer solutions to enable data scientists to operate on both a data lake and an AI/ML platform (or a compute farm) without worrying about the underlying infrastructure. IT also needs this to seamlessly grow to cloud scale while reducing the TCO of this infrastructure and without affecting utilization. Thus, driving a need to plan a data lake along with an AI/ML platform in a systemic fashion.

Seeing this increasing demand by IT, and also envisioning this as a natural extension of a data lake, we announced the Cisco Data Intelligence Platform.  Cisco Data Intelligence Platform is discussed in detail here.

This CVD implements Cisco Data Intelligence Platform on Cisco UCS using Cloudera Enterprise Data Hub 6.2.0.

Introduction

Both Big Data and machine learning technology have progressed to the point where they are being implemented in production systems running 24x7. There exists a very clear need for a proven, dependable, high-performance platform for the ingestion, processing, storage and analysis of the data, as well as the seamless dissemination of the output, results and insights of the analysis.

This solution implements the Cisco UCS Integrated Infrastructure for Big Data and Analytics based on Cisco Data Intelligence Platform (CDIP) architecture, a world-class platform specifically designed for demanding workloads that is both easy to scale and easy to manage, even as the requirements grow to thousands of servers and petabytes of storage; and the Cloudera Data Science Workbench, an integrated set of tools designed to enable flexible, fast access to the entire data store.

Many companies, recognizing the immense potential of big data and machine learning technology, are gearing up to leverage these new capabilities, building out departments and increasing hiring. However, these efforts face a new set of challenges:

·       Making the data available to the diverse set of people who need it

·       Enabling access to high-performance computing resources, GPUs, that also scale with the data growth

·       Allowing people to work with the data using the environments in which they are familiar

·       Publishing their results so the organization can make use of it

·       Enabling the automated production of those results

·       Managing the data for compliance and governance

·       Scaling the system as the data grows

·       Managing and administering the system in an efficient, cost-effective way

This solution is based on the Cisco UCS Integrated Infrastructure for Big Data and Analytics and includes computing, storage, connectivity, and unified management capabilities to help companies manage the immense amount of data being collected. It is built on Cisco Unified Computing System (Cisco UCS) infrastructure, using Cisco UCS 6332 Series Fabric Interconnects, and Cisco UCS C-Series Rack Servers. This architecture is specifically designed for performance and linear scalability for big data and machine learning workload.

Audience

The intended audience of this document includes sales engineers, field consultants, professional services, IT managers, partner engineering and customers who want to deploy the Cloudera Distribution with Apache Hadoop (CDH 6.2.0) and Cloudera Data Science Workbench (CDSW 1.5.0) on the Cisco UCS Integrated Infrastructure for Big Data and Analytics (Cisco UCS M5 Rack mount servers).

Purpose of this Document

This document describes the architecture and deployment procedures for Cloudera 6.2.0 with Cloudera Data Science Workbench 1.5.0 on a 30-node Cisco UCS C240 M5 and Cisco UCS S3260 cluster based on Cisco UCS Integrated Infrastructure for Big Data and Analytics.

This document describes the architecture and step by step guidelines of deployment procedures for Cisco Data Intelligence Platform using Cloudera Enterprise Data Hub (CDH) 6.2.0 on Cisco UCS C240 M5.

This document explains the process of deploying the three independently scalable components of the architecture including data lake, AI/ML and Object stores:

·       Data Lake with Cloudera Enterprise Data Hub 6.2

·       Kubernetes / AI farm with Cloudera Data Science Workbench

·       Hadoop Tiered storage with S3260

What’s New in this Release?

This solution implements the Cisco UCS Integrated Infrastructure for Big Data and Analytics based on Cisco Data Intelligence Platform (CDIP) architecture , a world-class platform specifically designed for demanding workloads that is both easy to scale and easy to manage, even as the requirements grow to thousands of servers and petabytes of storage; and Hortonworks Data Platform, an integrated set of tools designed to enable flexible, fast access to the entire data store, while enabling AI workloads on servers with GPUs.

This CVD implements the following:

·       Data Lake with Cloudera Enterprise Data Hub 6.2.0 on Cisco UCS Integrated Infrastructure for Big Data and Analytics

·       Kubernetes / AI farm with Cloudera Data Science Workbench (CDSW) 1.5

-      Enable CUDA for the GPUs

-      Enable GPU as a resource to the Docker Containers through CDSW

-      Enable GPU isolation and scheduling (with Docker Containers) through YARN 2.0

-      Downloading a TensorFlow image from NVIDIA Cloud (NGC)

-      Adding trusted registries for Docker for YARN 2.0

-      Execute a sample TensorFlow job accessing data from Hadoop and running on a Docker container with GPU as a resource scheduled by YARN 2.0

·       Installation and setup of the above through Cloudera Manager (CM).

·       Tiered storage on Cisco UCS S3260 M5 storage server chassis.

What’s Next?

This CVD showcases Cisco UCS Manager (UCSM). This solution can also be deployed using Cisco Intersight. Additional Cisco UCS features will be added to the Appendix in the following months. Some of these include the following:

·       Cisco Intersight

·       Cloudera Data Science Workbench

·       Tiered Storage with HDFS on S3260

·       Cisco Boot optimized M.2 Raid Controller for hardware RAID

·       4^th Generation FI

·       Hadoop data offload to S3 compliant storage

Solution Summary

This CVD details the process of installing Cloudera 6.2.0 with Cloudera Data Science Workbench (CDSW) 1.5.0 and the configuration details of the cluster. The current version of Cisco UCS Integrated Infrastructure for Big Data and Analytics offers the following configurations depending on the compute and storage requirements.

Cisco Data Intelligence Platform

Cisco Data Intelligence Platform (CDIP) is a cloud scale architecture which brings together big data, AI/compute farm, and storage tiers to work together as a single entity while also being able to scale independently to address the IT issues in the modern data center. This architecture allows for:

·       Extremely fast data ingest, and data engineering done at the data lake

·       AI compute farm allowing for different types of AI frameworks and compute types (GPU, CPU, FPGA) to work on this data for further analytics

·       A storage tier, allowing to gradually retire data which has been worked on to a storage dense system with a lower $/TB providing a better TCO

·       Seamlessly scale the architecture to thousands of nodes with a single pane of glass management using Cisco Application Centric Infrastructure (ACI)

Cisco Data Intelligence Platform caters to the evolving architecture bringing together a fully scalable infrastructure with centralized management and fully supported software stack (in partnership with industry leaders in the space) to each of these three independently scalable components of the architecture including data lake, AI/ML and Object stores.

Figure 1        Cisco Data Intelligent Platform

Cisco has developed numerous industry leading Cisco Validated Designs (reference architectures) in the area of Big Data (CVDs with Cloudera, Hortonworks and MapR), compute farm with Kubernetes (CVD with RedHat OpenShift) and Object store (Scality, SwiftStack, Cloudian, and others).

This Cisco Data Intelligence Platform can be deployed in these variants:

·       CDIP with Cloudera with Data Science Workbench (powered by Kubernetes) and Tiered Storage with Hadoop

·       CDIP with Hortonworks with Apache Hadoop 3.1 and Data Science Workbench (powered by Kubernetes) and Tiered Storage with Hadoop

Figure 2        Cisco Data Intelligence Platform with Hadoop, Kubernetes and Object Store

This architecture can start from a single rack and scale to thousands of nodes with a single pane of glass management with Cisco Application Centric Infrastructure (ACI).

Figure 3        Cisco Data Intelligent Platform at Scale

Reference Architecture

Table 1, Table 2 , and Table 3  summarize the reference architecture configuration details for the data lake, AI/ML components of the data lake, and tiered storage.

Data Lake Reference Architecture

Table 1  lists the data lake reference architecture configuration details for Cisco UCS Integrated Infrastructure for Big Data and Analytics.

Table 1    Cisco UCS Integrated Infrastructure for Big Data and Analytics Configuration Options

AI Computing Farm Reference Architecture

Table 2  lists the AI computing farm reference architecture configuration details for high-density CPU cores and GPU nodes.

Table 2    High-density CPU Cores and GPU Nodes

Tiered Storage

Table 3  lists the tiered storage reference architecture configuration details for Cisco UCS Integrated Infrastructure for Big Data and Analytics.

Table 3    Cisco UCS Integrated Infrastructure for Big Data and Analytics Configuration Options

Figure 4 illustrates a 12-node starter cluster with all the 3 components in a single Rack. The top 8 node has Cisco UCS C240 M5 servers as a data lake. Each link in the figure represents a 40 Gigabit Ethernet link from each of the 12 servers directly connected to a Fabric Interconnect. The second 2 x Cisco UCS C480 ML M5 Servers and the last 4 servers illustrate a data tiering on 2xS3260 servers. Every server is connected to both Fabric Interconnects.

Figure 4        Single Rack Starter Cluster Topology

As illustrated in Figure 5, a 30-node starter cluster. Rack #1 has sixteen Cisco UCS C240 M5 servers. Each link in the figure represents a 40 Gigabit Ethernet link from each of the sixteen servers directly connected to a Fabric Interconnect. Rack #2 has six Cisco UCS C240 M5 and four Cisco UCS S3260 servers. Every server is connected to both Fabric Interconnects.

Figure 5        Cisco Data Intelligence Platform – 30 Node Configuration with CDH 6.2 and CDSW 1.5

An alternate configuration for cases where more GPU capacity is needed. Four of the Cisco UCS C240 M5 servers from the previous configuration in Figure 5 are replaced with Cisco UCS C480 M5 ML M5 server which support up to eight V100 MXM GPUs.

Scaling the Solution

Figure 6 illustrates how to scale the solution. Each pair of Cisco UCS 6332 Fabric Interconnects has 24 Cisco UCS C240 M5 servers connected to it. This allows for eight uplinks from each Fabric Interconnect to the Cisco Nexus 9332 switch. Six pairs of 6332 FI’s can connect to a single switch with four uplink ports each. With 24 servers per FI, a total of 144 servers can be supported. Additionally, the can scale to thousands of nodes with the Nexus 9500 series family of switches.

Figure 6        Scaling the Solution

In the reference architectures discussed here, each of the components is scaled separately, and for the purposes of this example, scaling is uniform. Two scale scenarios are as follows:

·       Scaled architecture with 3:1 oversubscription with Cisco fabric interconnects and Cisco ACI

·       Scaled architecture with 2:1 oversubscription with Cisco ACI

In the following scenarios, the goal is to populate up to a maximum of 200 leaf nodes in a Cisco ACI domain. Not all cases reach that number because they use the Cisco Nexus® 9508 Switch for this sizing and not the Cisco Nexus 9516 Switch.

Scaled Architecture with 3:1 Oversubscription with Cisco Fabric Interconnects and Cisco ACI

The architecture discussed here and shown in Figure 7 supports 3:1 network oversubscription from every node to every other node across a multidomain cluster (nodes in a single domain within a pair of Cisco fabric interconnects are locally switched and not oversubscribed).

From the viewpoint of the data lake, 24 Cisco UCS C240 M5 Rack Servers are connected to a pair of Cisco UCS 6332 Fabric Interconnects (with 24 x 40-Gbps throughput). From each fabric interconnect, 8 x 40-Gbps links connect to a pair of Cisco Nexus 9336 Switches. Three pairs of fabric interconnects can connect to a single pair of Cisco Nexus 9336 Switches (8 x 40-Gbps links per Fabric Interconnect to Nexus switch). Each of these Cisco Nexus 9336 Switches connects to a pair of Cisco Nexus 9508 Cisco ACI switches with 6 x 100-Gbps uplinks (connecting to a Cisco N9K-X9736C-FX line card). the Cisco Nexus 9508 Switch with the Cisco N9K-X9736C-FX line card can support up to 36 x 100-Gbps ports, each and 8 such line cards.

Figure 7        Scaled Architecture with 3:1 Oversubscription with Cisco Fabric Interconnects and Cisco ACI

Scaled Architecture with 2:1 Oversubscription with Cisco ACI

In the scenario discussed here and shown in Figure 8, the Cisco Nexus 9508 Switch with the Cisco N9K-X9736C-FX line card can support up to 36 x 100-Gbps ports, each and 8 such line cards.

Here, for the 2:1 oversubscription, 30 Cisco UCS C240 M5 Rack Servers are connected to a pair of Cisco Nexus 9336 Switches, and each Cisco Nexus 9336 connects to a pair of Cisco Nexus 9508 Switches with three uplinks each. A pair of Cisco Nexus 9336 Switches can support 30 servers and connect to a spine with 6 x 100-Gbps links on each spine. This single pod (pair of Cisco Nexus 9336 Switches connecting to 30 Cisco UCS C240 M5 servers and 6 uplinks to each spine) can be repeated 48 times (288/6) for a given Cisco Nexus 9508 Switch and can support up to1440 servers.

To reduce the oversubscription ratio (to get 1:1 network subscription from any node to any node), you can use just 15 servers under a pair of Cisco Nexus 9336 Switches and then move to Cisco Nexus 9516 Switches (the number of leaf nodes would double).

To scale beyond this number, multiple spines can be aggregated.

Figure 8        Scaled Architecture with 2:1 Oversubscription with Cisco ACI

Cisco UCS Integrated Infrastructure for Big Data and Analytics

The Cisco UCS Integrated Infrastructure for Big Data and Analytics solution for Cloudera is based on Cisco UCS Integrated Infrastructure for Big Data and Analytics, a highly scalable architecture designed to meet a variety of scale-out application demands with seamless data integration and management integration capabilities built using the components described in this section.

Cisco UCS Manager

Cisco UCS Manager (UCSM) resides within the Cisco UCS Fabric Interconnect. It makes the system self-aware and self-integrating, managing all the system components as a single logical entity. Cisco UCS Manager can be accessed through an intuitive graphical user interface (GUI), a command-line interface (CLI), or an XML application-programming interface (API). Cisco UCS Manager uses service profiles to define the personality, configuration, and connectivity of all resources within Cisco UCS, radically simplifying provisioning of resources so that the process takes minutes instead of days. This simplification allows IT departments to shift their focus from constant maintenance to strategic business initiatives.

Key Features

·       Supports Cisco UCS B-Series Blade and Cisco UCS C-Series Rack Servers, the Cisco UCS C3260 storage server, Cisco UCS Mini, and the Cisco HyperFlex hyperconverged infrastructure.

·       Programmatically controls server, network, and storage resources, with a unified, policy-driven management, so they can be efficiently managed at scale through software.

·       Works with HTML 5, Java, or CLI graphical user interfaces.

·       Can automatically detect, inventory, manage, and provision system components that are added or changed.

·       Facilitates integration with third-party systems management tools.

·       Builds on existing skills and supports collaboration across disciplines through role-based administration

Cisco UCS 6300 Series Fabric Interconnects

Cisco UCS 6300 Series Fabric Interconnects provide high-bandwidth, low-latency connectivity for servers, with integrated, unified management provided for all connected devices by Cisco UCS Manager. Deployed in redundant pairs, Cisco fabric interconnects offer the full active-active redundancy, performance, and exceptional scalability needed to support the large number of nodes that are typical in clusters serving big data applications. Cisco UCS Manager enables rapid and consistent server configuration using service profiles, automating ongoing system maintenance activities such as firmware updates across the entire cluster as a single operation. Cisco UCS Manager also offers advanced monitoring with options to raise alarms and send notifications about the health of the entire cluster.

The Cisco UCS 6300 series Fabric interconnects are a core part of Cisco UCS, providing low-latency, lossless 10 and 40 Gigabit Ethernet, Fiber Channel over Ethernet (FCoE), and Fiber Channel functions with management capabilities for the entire system. All servers attached to Fabric interconnects become part of a single, highly available management domain.

Figure 9        Cisco UCS 6332 UP 32 -Port Fabric Interconnect

Cisco UCS C-Series Rack-Mount Servers

Cisco UCS C-Series Rack-Mount Servers keep pace with Intel Xeon processor innovation by offering the latest processors with increased processor frequency and improved security and availability features. With the increased performance provided by the Intel Xeon Scalable Family Processors, Cisco UCS C-Series servers offer an improved price-to-performance ratio. They also extend Cisco UCS innovations to an industry-standard rack-mount form factor, including a standards-based unified network fabric, Cisco VN-Link virtualization support, and Cisco Extended Memory Technology.

It is designed to operate both in standalone environments and as part of Cisco UCS managed configuration, these servers enable organizations to deploy systems incrementally—using as many or as few servers as needed—on a schedule that best meets the organization’s timing and budget. Cisco UCS C-Series servers offer investment protection through the capability to deploy them either as standalone servers or as part of Cisco UCS. One compelling reason that many organizations prefer rack-mount servers is the wide range of I/O options available in the form of PCIe adapters. Cisco UCS C-Series servers support a broad range of I/O options, including interfaces supported by Cisco and adapters from third parties.

Cisco UCS C240 M5 Rack-Mount Server

The Cisco UCS C240 M5 Rack-Mount Server (Figure 10) is a 2-socket, 2-Rack-Unit (2RU) rack server offering industry-leading performance and expandability. It supports a wide range of storage and I/O-intensive infrastructure workloads, from big data and analytics to collaboration. Cisco UCS C-Series Rack Servers can be deployed as standalone servers or as part of a Cisco Unified Computing System (Cisco UCS) managed environment to take advantage of Cisco’s standards-based unified computing innovations that help reduce customers’ Total Cost of Ownership (TCO) and increase their business agility.

In response to ever-increasing computing and data-intensive real-time workloads, the enterprise-class Cisco UCS C240 M5 server extends the capabilities of the Cisco UCS portfolio in a 2RU form factor. It incorporates the 2^nd generation Intel® Xeon® Scalable and Intel® Xeon® Scalable processors, supporting up to 20 percent more cores per socket, twice the memory capacity, and five times more Non-Volatile Memory Express (NVMe) PCI Express (PCIe) Solid-State Disks (SSDs) compared to the previous generation of servers. These improvements deliver significant performance and efficiency gains that will improve your application performance. The Cisco UCS C240 M5 delivers outstanding levels of storage expandability with exceptional performance, along with the following:

·       Latest Intel Xeon Scalable CPUs with up to 28 cores per socket

·       Up to 24 DDR4 DIMMs for improved performance

·       Up to 26 hot-swappable Small-Form-Factor (SFF) 2.5-inch drives, including 2 rear hot-swappable SFF drives (up to 10 support NVMe PCIe SSDs on the NVMe-optimized chassis version), or 12 Large-Form- Factor (LFF) 3.5-inch drives plus 2 rear hot-swappable SFF drives

·       Support for 12-Gbps SAS modular RAID controller in a dedicated slot, leaving the remaining PCIe Generation 3.0 slots available for other expansion cards

·       Modular LAN-On-Motherboard (mLOM) slot that can be used to install a Cisco UCS Virtual Interface Card (VIC) without consuming a PCIe slot, supporting dual 10- or 40-Gbps network connectivity

·       Dual embedded Intel x550 10GBASE-T LAN-On-Motherboard (LOM) ports

·       Modular M.2 or Secure Digital (SD) cards that can be used for boot

Figure 10     Cisco UCS C240 M5 Rack-Mount Server

Cisco UCS S3260 Storage Servers

The Cisco UCS S3260 Storage Server is a modular storage server with dual M5 server nodes and is optimized to deliver efficient, industry-leading storage for data-intensive workloads. The Cisco UCS S3260 server with dual-node capability that is based on the 2^nd Gen Intel® Xeon® Scalable and Intel® Xeon® Scalable processors, the server features up to 840 TB of local storage in a compact 4-Rack-Unit (4RU) form factor. The drives can be configured with enterprise-class Redundant Array of Independent Disks (RAID) redundancy or with a pass-through Host Bus Adapter (HBA) controller. Network connectivity is provided with dual-port 40-Gbps nodes in each server, with expanded unified I/O capabilities for data migration between Network-Attached Storage (NAS) and SAN environments. This storage-optimized server comfortably fits in a standard 32-inch-depth rack, such as the Cisco® R 42610 Rack.

Figure 11     Cisco UCS S3260 Storage Server

The Cisco UCS S3260 Storage Server chassis has 56 top-load LFF HDDs option as shown above with a maximum capacity of 4 TB per HDD and can be mixed with up to 28 SSDs.

The modular Cisco UCS S3260 Storage Server chassis offers flexibility with more computing, storage, and PCIe expansion on the second slot in the chassis. This second slot can be used for:

·       An additional server node

·       Four additional LFF HDDs with up to 10 TB capacity per HDD

·       New PCIe expansion tray with up to two x8 half-height, half-width PCIe slots that can use any industry-standard PCIe card including Fibre Channel and Ethernet cards.

The Cisco UCS S3260 Storage Server Chassis includes a Cisco UCS Virtual Interface Card (VIC) 1300 platform chip onboard the system I/O controller, offering high-performance bandwidth with dual-port 40 Gigabit Ethernet and FCoE interfaces per system I/O controller.

Figure 12     Cisco UCS S3260 Storage Server: Rear View

 

Cisco UCS C480 M5 Rack-Mount Server

The Cisco UCS C480 M5 Rack-Mount Server is a storage and I/O-optimized enterprise-class rack-mount server that delivers industry-leading performance for in-memory databases, big data analytics, virtualization, Virtual Desktop Infrastructure (VDI), and bare-metal applications. The Cisco UCS C480 M5 (Figure 13) delivers outstanding levels of expandability and performance for standalone or Cisco Unified Computing System managed environments in a 4RU form-factor. Because of its modular design, you pay for only what you need. It offers these capabilities:

·       Latest 2^nd Gen Intel® Xeon® Scalable and Intel® Xeon® processors and support for two-or four-processor configurations

·       2666-MHz DDR4 memory and 48 DIMM slots for up to 6 Terabytes (TB) of total memory

·       12 PCI Express (PCIe) 3.0 slots

-      Six x 8 full-height, full length slots

-      Six x16 full-height, full length slots

·       Flexible storage options with support up to 32 Small-Form-Factor (SFF) 2.5-inch, SAS, SATA, and PCIe NVMe disk drives

·       Cisco 12-Gbps SAS Modular RAID Controller in a dedicated slot

·       Internal Secure Digital (SD) and M.2 boot options

·       Dual embedded 10 Gigabit Ethernet LAN-On-Motherboard (LOM) ports

Figure 13     Cisco UCS C480 M5 Rack-Mount Server – Front View

Figure 14     Cisco UCS C480 M5 Rack-Mount Server – Rear View

For more information about Cisco UCS C480 M5 Rack Server, go to: https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/datasheet-c78-739291.html

Cisco UCS C480 ML M5 Rack Server

The Cisco UCS C480 ML M5 Rack Server is a purpose-built server for Deep Learning. It is storage and I/O optimized to deliver an industry-leading performance for training Models. The Cisco UCS C480 ML M5 delivers outstanding levels of storage expandability and performance options for standalone or Cisco Unified Computing System managed environments in a 4RU form factor. Because of its modular design, you pay for only what you need. It offers these capabilities:

·       8 NVIDIA SXM2 V100 32G modules with NVLink interconnect

·       Dual socket latest 2^nd Gen Intel® Xeon® Scalable and Intel® Xeon® processors

·       2666-MHz DDR4 memory and 24 DIMM slots for up to 3 terabytes (TB) of total memory

·       4 PCI Express (PCIe) 3.0 slots for 100G Cisco UCS VIC 1495

·       Flexible storage options with support for up to 24 Small-Form-Factor (SFF) 2.5-inch, SAS/SATA Solid-State Disks (SSDs) and Hard-Disk Drives (HDDs)

·       Up to 6 PCIe NVMe disk drives

·       Cisco 12-Gbps SAS Modular RAID Controller in a dedicated slot

·       M.2 boot options

·       Dual embedded 10 Gigabit Ethernet LAN-On-Motherboard (LOM) ports

Figure 15     Cisco UCS C480 ML M5 Purpose Built Deep Learning Server – Front View

Figure 16     Cisco UCS C480 ML M5 Purpose Built Deep Learning Server – Rear View

For more information about Cisco UCS C480 ML M5 Server, review the specification sheet: https://www.cisco.com/c/dam/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/c480m5-specsheet-ml-m5-server.pdf

Table 4  lists the features and benefits of Cisco UCS C480 ML M5 Server.

Table 4    Features and Benefits of Cisco UCS C480 ML M5 Server

Cisco UCS Virtual Interface Cards (VICs)

Cisco UCS Virtual Interface Cards (VIC) are unique to Cisco. Cisco UCS Virtual Interface Cards incorporate next-generation converged network adapter (CNA) technology from Cisco and offer dual 10- and 40-Gbps ports designed for use with Cisco UCS servers. Optimized for virtualized networking, these cards deliver high performance and bandwidth utilization, and support up to 256 virtual devices.

The Cisco UCS Virtual Interface Card 1387 offers dual-port Enhanced Quad Small Form-Factor Pluggable (QSFP+) 40 Gigabit Ethernet and Fiber Channel over Ethernet (FCoE) in a modular-LAN-on-motherboard (mLOM) form factor. The mLOM slot can be used to install a Cisco VIC without consuming a PCIe slot providing greater I/O expandability.

Figure 17     Cisco UCS VIC 1387

NVIDIA GPU

Graphics Processing Units or GPUs are specialized processors designed to render images, animation and video for computer displays. They perform this task by running many operations simultaneously. While the number and kinds of operations they can do are limited, they make up for it by being able run many thousands in parallel. As the graphics capabilities of GPUs increased, it soon became apparent that the massive parallelism of GPUs could be put to other uses beside rendering graphics.

NVIDIA GPU used in this document, NVIDIA Tesla V100, is advanced data center GPU built to accelerate AI, HPC, and graphics. It is powered by NVIDIA Volta architecture, comes in 16 and 32 GB configurations.

NVIDIA GPUs bring two key advantages to the table. First, they make possible solutions that were simply not computationally possible before. Second, by providing the same processing power as scores of traditional CPUs they reduce the requirements for rack space, power, networking and cooling in the data center.

NVIDIA CUDA

GPUs are very good at running the same operation on different data simultaneously. This is often referred to as single instruction, multiple data, or SIMD. This is exactly what’s needed to render graphics, but many other computing problems can benefit from this approach. As a result, NVIDIA created CUDA. CUDA is a parallel computing platform and programming model that makes it possible to use a GPU for many general-purpose computing tasks via commonly used programming languages like C and C++.

In addition to the general-purpose computing capabilities that CUDA enables there is also a special CUDA library for deep learning called the CUDA Deep Neural Network library, or cuDNN. cuDNN makes it easier to implement deep machine learning architectures that take full advantage of the GPU’s capabilities.

Cloudera Enterprise Data Hub and Hortonworks Data Platform

The CVDs implement with Cloudera Enterprise Data Hub.

Cloudera (CDH 6.2.0)

Built on the transformative Apache Hadoop open source software project, Cloudera Enterprise is a hardened distribution of Apache Hadoop and related projects designed for the demanding requirements of enterprise customers. Cloudera is the leading contributor to the Hadoop ecosystem, and has created a rich suite of complementary open source projects that are included in Cloudera Enterprise.

All the integration and the entire solution is thoroughly tested and fully documented. By taking the guesswork out of building out a Hadoop deployment, CDH gives a streamlined path to success in solving real business problems.

Cloudera Enterprise with Apache Hadoop is:

·       Unified – one integrated system, bringing diverse users and application workloads to one pool of data on common infrastructure; no data movement required

·       Secure – perimeter security, authentication, granular authorization, and data-protection

·       Governed – enterprise-grade data auditing, data lineage, and data-discovery

·       Managed – native high-availability, fault-tolerance and self-healing storage, automated backup and disaster recovery, and advanced system and data management

·       Open – Apache-licensed open source to ensure both data and applications remain copy righted, and an open platform to connect with all the existing investments in technology and skills.

Figure 18     Cloudera Data Hub

Cloudera provides the following products and tools:

·       CDH—The Cloudera distribution of Apache Hadoop and other related open-source projects, including Spark. CDH also provides security and integration with numerous hardware and software solutions.

·       Apache Spark—An integrated part of CDH and supported with Cloudera Enterprise, Spark is an open standard for flexible in-memory data processing for batch, real time and advanced analytics. Via the one platform Cloudera is committed to adopting Spark as the default data execution engine for analytic workloads.

·       Cloudera Manager—A sophisticated application used to deploy, manage, monitor, and diagnose issues with CDH deployments. Cloudera Manager provides the Admin Console, a web-based user interface that makes administration of any enterprise data simple and straightforward. It also includes the Cloudera Manager API, which can be used to obtain cluster health information and metrics, as well as configure Cloudera Manager.

·       Cloudera Navigator—An end-to-end data management tool for the CDH platform. Cloudera Navigator enables administrators, data managers, and analysts to explore the large amounts of data in Hadoop. The robust auditing, data management, lineage management, and life cycle management in Cloudera Navigator allow enterprises to adhere to stringent compliance and regulatory requirements.

Cloudera Data Science Workbench

Cloudera Data Science Workbench (CDSW) is a web application that allows data scientists to use a variety of open source languages and libraries to directly and securely access the data in the Hadoop cluster. Direct access to the big data cluster means no more working with small subsets of the data on desktop systems; no sampling is required as the entire data set is available for use directly by the user. Further, users are not restricted to a single environment. Many popular open source libraries and languages are supported, including R, Python and Scala, as well as all the ML/DL frameworks such as TensorFlow, Theano, PyTorch, and so on. In addition, CDSW enables access to available GPU resources for deep learning workloads which means users become productive faster with no need for retraining and no time lost learning a new programming language.

CDSW is addressing the key challenge that every team or user may require a different language, library or framework in order to be productive while the organization requires reproducibility and collaboration. By making the entire set of data in the cluster available to the user, CSDW eliminates the problem that what works on small samples or extracts of the data on a user’s desktop computer may not scale across a large cluster. Cloudera Data Science Workbench gives data scientists the flexibility and simplicity they need to be productive and innovative at scale.

Figure 19     Cloudera Data Science Architecture

In addition, CDSW enables seamless access to high-performance processors in the form of GPUs. CSDW makes use of lightweight container architecture to rapidly and securely provide the environment and resources to the users.

Cloudera Data Science Workbench is directly aimed at helping data scientists build and test new analyses and analytics projects as quickly as possible in secure manner even in large scale environments. This flexibility improves the efficiency of the exploration process, a key requirement to meet in order to move rapidly from idea to answer. Most analytics problems, especially those with transformative power, are not standard analyses and require advanced models and iterative methods. Experimentation and innovation are the heart and soul of data science, but security is needed for compliance and governance.

Data has become one of the most strategic assets in the organization. Leveraging the data to drive the business forward is the primary motivation for building an enterprise data hub to support advanced analytics. Typically, when forced to make a choice between the security of the data and the flexibility to access it, security wins locking away the data from the people who most need it. CDSW address this issue by providing full authentication and access controls against data in the cluster, including complete Kerberos integration. It offers data science teams per-project isolation and reproducibility with no effort.

Cloudera Data Science Workbench allows you to automate analytics workloads with a built-in job and pipeline scheduling system that supports real-time monitoring, job history, and email alerts. Jobs are created and can be configured to run on a recurring schedule, as well as providing alerts for successful and failed runs. Multiple jobs can be scheduled together to create an automated pipeline; for example, the first job performs data acquisition, the next data cleansing, then analytics, and so on.

Collaboration and sharing of results are implemented via project sharing (either globally or to specific users, and project forking. To share results, CSDW enables publishing output for viewing via a browser, and even makes the console log itself available for viewing both during and after the run. Cloudera Data Science Workbench is a web application. It has no desktop footprint making it very easy to administer and maintain.

Cloudera Data Science Workbench's interactive mode allows you to launch a Spark application and work iteratively in R, Python, or Scala, rather than the standard workflow of launching an application and waiting for it to complete to view the results. Because of its interactive nature, Cloudera Data Science Workbench works with Spark on YARN's client mode, where the driver persists through the lifetime of the job and runs executors with full access to the CDH cluster resources. This architecture is illustrated in Figure 20.

Figure 20     Cloudera Data Science Workbench with Spark on Yarn

Data Science Workbench provides the following features:

·       CPU and GPU as a resource: Data Science Workbench provides basic support for the use of existing general-purpose CPUs for each stage of the workflow and, optionally, accelerates the math-intensive steps with the selective application of special-purpose GPUs all through a Docker container, with Kubernetes scheduling these resources in the back end.

·       Self-service portal: The Data Science Workbench web user interface console provides a self-service portal for data scientists to create an environment for their workloads (Figure 21). Currently, R, Python, and Scala are supported.

·       Jupyter Notebook: Most data scientists use Jupyter Notebooks for AI/ML analysis and development. Data Science Workbench provides a Jupyter Notebook environment when data scientists create a portal, and these notebooks can be shared or worked in a collaborative manner.

Figure 21     Example of Cloudera Data Science Workbench WebUI

Hortonworks Data Platform

The Hortonworks Data Platform (HDP 3.1.0) delivers essential capabilities in a completely open, integrated and tested platform that is ready for enterprise usage. With Hadoop YARN at its core, HDP provides flexible enterprise data processing across a range of data processing engines, paired with comprehensive enterprise capabilities for governance, security and operations.

All the integration of the entire solution is thoroughly tested and fully documented. By taking the guesswork out of building out a Hadoop deployment, HDP gives a streamlined path to success in solving real business problems.

Hortonworks Data Platform (HDP) 3.0 delivers significant new features, including the ability to launch apps in a matter of minutes and address new use cases for high-performance deep learning and machine learning apps. In addition, this new version of HDP enables enterprises to gain value from their data faster, smarter, in a hybrid environment.

Apache Ambari

Apache Ambari is a completely open source management platform. It performs provisioning, managing, securing, and monitoring Apache Hadoop clusters. Apache Ambari is a part of Hortonworks Data Platform and it allows enterprises to plan and deploy HDP cluster. It also provides ongoing cluster maintenance and management.

Ambari provides an intuitive Web UI as well as an extensive REST API framework which is very useful for automating cluster operations.

Below are the core benefits that Hadoop operators get with Ambari:

·       Simplified Installation, Configuration and Management. Easily and efficiently create, manage and monitor clusters at scale. Takes the guesswork out of configuration with Smart Configs and Cluster Recommendations.  Enables repeatable, automated cluster creation with Ambari Blueprints.

·       Centralized Security Setup. Reduce the complexity to administer and configure cluster security across the entire platform. Helps automate the setup and configuration of advanced cluster security capabilities such as Kerberos and Apache Ranger.

·       Full Visibility into Cluster Health. Ensure your cluster is healthy and available with a holistic approach to monitoring. Configures predefined alerts — based on operational best practices — for cluster monitoring. Captures and visualizes critical operational metrics — using Grafana — for analysis and troubleshooting. Integrated with Hortonworks SmartSense for proactive issue prevention and resolution.

·       Highly Extensible and Customizable. Fit Hadoop seamlessly into your enterprise environment. Highly extensible with Ambari Stacks for bringing custom services under management, and with Ambari Views for customizing the Ambari Web UI.

HDP for Data Access

With YARN at its foundation, HDP provides a range of processing engines that allow users to interact with data in multiple and parallel ways, without the need to stand up individual clusters for each data set/application. Some applications require batch while others require interactive SQL or low-latency access with NoSQL. Other applications require search, streaming or in-memory analytics. Apache Solr, Storm and Spark fulfill those needs respectively.

To function as a true data platform, the YARN-based architecture of HDP enables the widest possible range of access methods to coexist within the same cluster avoiding unnecessary and costly data silos.

As shown in Figure 22, HDP Enterprise natively provides for the following data access types:

·       Batch – Apache MapReduce has served as the default Hadoop processing engine for years. It is tested and relied upon by many existing applications.

·       Interactive SQL Query - Apache Hive is the de facto standard for SQL interactions at petabyte scale within Hadoop. Hive delivers interactive and batch SQL querying across the broadest set of SQL semantics.

·       Search - HDP integrates Apache Solr to provide high-speed indexing and sub-second search times across all your HDFS data.

·       Scripting - Apache Pig is a scripting language for Hadoop that can run on MapReduce or Apache Tez, allowing you to aggregate, join and sort data.

·       Low-latency access via NoSQL - Apache HBase provides extremely fast access to data as a columnar format, NoSQL database. Apache Accumulo also provides high-performance storage and retrieval, but with fine-grained access control to the data.

·       Streaming - Apache Storm processes streams of data in real time and can analyze and act on data as it flows into HDFS.

Figure 22     YARN

Submarine

Deep learning is useful for enterprises tasks in the field of speech recognition, image classification, AI chatbots, machine translation, just to name a few. In order to train deep learning/machine learning models, frameworks such as TensorFlow / MXNet / PyTorch / Caffe / XGBoost can be leveraged. Sometimes these frameworks are used together to solve different problems.

To make distributed deep learning/machine learning applications easily launched, managed and monitored, Hadoop community initiated the Submarine project along with other improvements such as first-class GPU support, Docker container support, container-DNS support, scheduling improvements, and so on.

These improvements make distributed deep learning/machine learning applications run on Apache Hadoop YARN as simple as running it locally, which can let machine-learning engineers focus on algorithms instead of worrying about underlying infrastructure. By upgrading to latest Hadoop, users can now run deep learning workloads with other ETL/streaming jobs running on the same cluster. This can achieve easy access to data on the same cluster and achieve better resource utilization.

Figure 23     Submarine Workflow

Docker Containerization

Cloudera Data Science Workbench makes use of container technology. Containers are conceptually like virtual machines, but instead of virtualizing the hardware, a container virtualizes the operating system. With a VM there is an entire operating system sitting on top of the hypervisor. Containers dispense with this time-consuming and resource hungry requirement by sharing the host system’s kernel. As a result, a container is far smaller, and its lightweight nature means they can be instantiated quickly. In fact, they can be instantiated so quickly that new application architectures are possible.

Docker is an open-source project based on Linux containers. It uses Linux kernel features like namespaces and control groups to create containers. These features are not new, but Docker has taken these concepts and improved them in the following ways:

·       Ease of use: Docker makes easier for anyone — developers, systems admins, architects and others — to take advantage of containers in order to quickly build and test portable applications. It allows anyone to package an application on their development system, which can then run unmodified on any cloud or bare metal server. The basic idea is to create a “build once, run anywhere” system.

·       Speed: Docker containers are very fast with a small footprint. Ultimately, containers are just sandboxed environments running on the kernel, so they take up few resources. You can create and run a Docker container in seconds. Compare this to a VM which takes much longer because it must boot up a full virtual operating system every time.

·       Modularity: Docker makes it easy to take an application and breaks its functionality into separate individual containers. These containers can then be spun up and run as needed. This is particularly useful for cases where an application needs to hold and lock a resource, like a GPU, and then release it once it’s done using it. Modularity also enables each component, i.e., container to be updated independently.

·       Scalability: modularity enables scalability. With different parts of the system running in different containers it becomes possible, and with Docker, it becomes easy to connect these containers together to create an application, which can then be scaled out as needed.

YARN Support for Docker

Containerization provides YARN support for Docker containers, which makes it easier to bundle libraries and dependencies along with their application, allowing third-party applications to run on Apache Hadoop (for example, containerized applications), enabling:

·       Faster time to deployment by enabling third-party apps.

·       The ability to run multiple versions of an application, enabling users to rapidly create features by developing and testing new versions of services without disrupting old ones.

·       Improved resource utilization and increased task throughput for containers, yielding faster time to market for services.

·       Orchestration of stateless distributed applications.

·       Packaging libraries for Spark application, eliminating the need for operations to deploy those libraries cluster wide.

Figure 24     Containerized Application on Apache Hadoop YARN 3.1

As shown in Figure 24, YARN Services Framework in addition with Docker containerization, it is now possible to run both existing Hadoop frameworks, such as Hive, Spark, and so on, and new containerized workloads on the same underlying infrastructure. Apache Hadoop 3.1 further improved these capabilities to enable advanced use cases such as TensorFlow and HBase.

Kubernetes

Applications built using container technology provide a great deal of flexibility in terms of their architecture, deployment and scaling. Since containers provide VM-like separation of concerns but with far less overhead they allow system developers to package different services of the same application into separate containers. These containers can then be deployed in a very flexible manner including across clusters of physical and virtual machines. This builds the ability to scale directly into the application architecture. This ability in turn requires a tool to aid in deploying, managing and scaling container-based applications.

Kubernetes is an open source project specifically designed for deploying and managing multi-container applications at scale. Kubernetes automates and simplifies the following tasks:

Deploying multi-container applications. With the application split into separate containers for different services, Kubernetes manages the deployment of the containers both at initial startup and in real-time as needed by the application.

Scaling containers. Applications need to spin up and down containers to suit demand, to balance incoming load, and make better use of physical resources. Kubernetes provides the mechanisms for doing these things in a completely automated way.

Updating applications. One advantage of container-based application development is the ability to independently change, improve and fix individual containers. Kubernetes has mechanisms for allowing graceful updates to new versions of container images, including rollbacks if something does not go as planned.

Kubernetes manages application status and any replication and load balancing needs. It also handles hardware resource allocation including GPUs. Kubernetes also has facilities for maximizing the use of hardware resources including memory, storage I/O, and network bandwidth. Applications can have soft and hard limits set on their resource usage. For example, many small applications that use minimal resources can be run together on the same hardware while resource hungry applications can be placed on different hardware and scale out as needed.

NVIDIA Docker

Docker containers are platform-agnostic, but also hardware-agnostic. This presents a problem when using specialized hardware such as NVIDIA GPUs which require kernel modules and user-level libraries to operate.  As a result, Docker does not natively support NVIDIA GPUs within containers.

One of the early workarounds to this problem was to fully install the NVIDIA drivers inside the container and map in the character devices corresponding to the NVIDIA GPUs (for example, /dev/nvidia0) on launch. This solution is brittle because the version of the host driver must exactly match the version of the driver installed in the container. This requirement drastically reduced the portability of these early containers, undermining one of Docker’s more important features.

To enable portability in Docker images that leverage NVIDIA GPUs, NVIDIA developed nvidia-docker, an open-source project hosted on GitHub that provides the two critical components needed for portable GPU-based containers:

·       driver-agnostic CUDA images; and a Docker command line wrapper that mounts the user mode components of the driver and the GPUs (character devices) into the container at launch.

·       nvidia-docker is essentially a wrapper around the docker command that transparently provisions a container with the necessary components to execute code on the GPU.

GPU Pooling and Isolation

GPU pooling and isolation allows GPU to be a first-class resource type in Hadoop, making it easier for customers to run machine learning and deep learning workloads.

·       Compute-intensive analytics require not only a large compute pool, but also a fast and expensive processing pool with GPUs in tandem

·       Customers can share cluster-wide GPU resources without having to dedicate a GPU node to a single tenant or workload

·       GPU isolation dedicates a GPU to an application so that no other application has access to that GPU

When it comes to resource scheduling, it is important to recognize GPU as a resource. YARN extends the resource model to more flexible mode which makes it easier to add new countable resource-types. When GPU is added as resource type, YARN can schedule applications on GPU machines. Furthermore, by specifying the number of requested GPU to containers, YARN can find machines with available GPUs to satisfy container requests.

Red Hat Ansible Automation

Red Hat Ansible Automation is a powerful IT automation tool. It is capable of provisioning numerous types of resources and deploying applications. It can configure and manage devices and operating system components. Due to its simplicity, extensibility, and portability, this solution extensively utilizes Ansible for performing repetitive deployment steps across the nodes.

Requirements

This CVD describes architecture and deployment procedures for Cloudera Enterprise Data Hub (CDH 6.2.0) and Cloudera Data Science Workbench (CDSW 1.5.0) on a 30-node cluster based on Cisco UCS Integrated Infrastructure for Big Data and Analytics. The solution goes into detail configuring CDH 6.2.0 on the infrastructure, as well as the complete installation and configuration of CDSW 1.5.0 and all its dependencies.

The cluster configuration consists of the following:

·       2 Cisco UCS 6332UP Fabric Interconnects

·       22 Cisco UCS C240 M5 Rack-Mount servers

·       4 Cisco UCS S3260 Storage Server

·       8 NVIDIA Tesla T4 GPU

·       2 Cisco R42610 standard racks

·       4 Vertical Power distribution units (PDUs) (Country Specific)

Physical Topology

Each rack consists of two vertical PDUs. The first rack consists of two Cisco UCS 6332UP Fabric Interconnects, 16 Cisco UCS C240 M5 Rack Servers connected to each of the vertical PDUs for redundancy; thereby, ensuring availability during power source failure. The second rack consists of 6 Cisco UCS C240 M5 Servers and 4 Cisco UCS S3260 Modular Storage Server connected to each of the vertical PDUs for redundancy; thereby, ensuring availability during power source failure, like the first rack.

As illustrated in Figure 25, a 30-node starter cluster. Rack #1 has sixteen Cisco UCS C240 M5 servers. Each link in the figure represents a 40 Gigabit Ethernet link from each of the sixteen servers directly connected to a Fabric Interconnect. Rack #2 has six Cisco UCS C240 M5 and four Cisco UCS S3260 servers. Every server is connected to both Fabric Interconnects.

Figure 25     Cisco Data Intelligence Platform - 30 Node Configuration with Cloudera CDH 6.2 and CDSW 1.5

Port Configuration on Fabric Interconnect

Table 5  lists the port configuration on Cisco UCS FI 6332 Fabric Interconnect.

Table 5    Port Configuration on Fabric Interconnect

Server Configuration and Cabling for Cisco UCS C240 M5

The Cisco UCS C240 M5 rack server is equipped with 2 x Intel Xeon Scalable Family Processor 6230 (2 x 20 cores, 2.1 GHz), 384 GB of memory, Cisco UCS Virtual Interface Card 1337, Cisco 12-Gbps SAS Modular Raid Controller with 4-GB FBWC, 26 x 1.8 TB 10K rpm SFF SAS HDDs or 12 x 1.6 TB Enterprise Value SATA SSDs, M.2 with 2 x 240-GB SSDs for Boot.

Figure 26 illustrates the port connectivity between the Cisco UCS FI 6332 and Cisco UCS C240 M5 Rack Server. Twenty-two Cisco UCS C240 M5 servers are installed in this configuration.

Figure 26     Fabric Topology for Cisco UCS C240 M5 Rack Server

Figure 27 illustrates the port connectivity between the Cisco UCS FI 6332 and Cisco UCS S3260 Storage Server. Four Cisco UCS S3260 M5 servers are installed in this configuration.

Figure 27     Fabric Topology for Cisco UCS S3260 Storage Server

Figure 27 shows the connectivity between Cisco UCS S3260 Storage Server chassis and Cisco UCS 6300 Fabric Interconnects. Each chassis has two Cisco UCS C3260 M5 server nodes. Each link in Figure 27 represents a 40 Gigabit Ethernet link from the Cisco UCS S3260 Storage Server chassis connecting to a Fabric Interconnect. Every chassis is connected to both Fabric Interconnects represented with dual links. Since each chassis will have two server nodes, the top server node works with the left SIOC and the bottom server node works with right SIOC. Similarly, for the boot drives, the top two SSD slots are assigned for server node 1 and the bottom two SSD slots are assigned for server node 2.

For information on physical connectivity and single-wire management, see:

https://www.cisco.com/c/en/us/td/docs/unified_computing/ucs/c-series_integration/ucsm4-0/b_C-Series-Integration_UCSM4-0/b_C-Series-Integration_UCSM4-0_chapter_01.html

Software Distributions and Versions

The software distributions required versions are listed in Table 6. 

Table 6    Software distribution and Version

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Cisco Unified Computing System Configuration

This section details the Cisco Unified Computing System (Cisco UCS) configuration that was done as part of the infrastructure build out. The racking, power, and installation of the Cisco UCS Rack Server is described in the physical topology section earlier in this document. Please refer to the Cisco UCS Manager Getting Started Guide. For more information about each step, see the Cisco UCS Manager - Configuration Guides.

Configure Cisco UCS Fabric Interconnect

This document assumes you are using Cisco UCS Manager Software version 4.0(4c). To upgrade the Cisco UCS Manager software and the Cisco UCS 6332 Fabric Interconnect software to a higher version of the firmware, see the Cisco UCS Manager Install and Upgrade Guides.

Alternatively, if you intend to clear the existing Cisco UCS Manager configuration, follow these steps:

1.     Connect a console cable to the console port on what will become the primary fabric interconnect.

2.     If the fabric interconnects were previously deployed and you want to erase it to redeploy, follow these steps:

a.     Login with the existing username and password.

#connect local-mgmt

#erase config

#yes (to confirm)

3.     After the fabric interconnect restarts, the out-of-box first time installation prompt appears, type “console” and press Enter.

4.     Follow the Initial Configuration steps as outlined in Cisco UCS Manager Getting Started Guide. When configured, log into UCSM IP Address via the web interface to perform the base Cisco UCS configuration.

Configure Fabric Interconnects for a Cluster Setup

To configure the Cisco UCS Fabric Interconnects, follow these steps:

1.     Verify the following physical connections on the fabric interconnect:

-      The management Ethernet port (mgmt0) is connected to an external hub, switch, or router.

-      The L1 ports on both fabric interconnects are directly connected to each other.

-      The L2 ports on both fabric interconnects are directly connected to each other

Configure Fabric Interconnect A

To configure Fabric Interconnect A, follow these steps:

1.     Connect to the console port on the first Cisco UCS 6332 Fabric Interconnect.

2.     Enter the password for the admin user.

3.     Enter the same password again to confirm the password for the admin user.

4.     Enter the cluster name for the system name.

5.     Enter the Mgmt0 IPv4 address.

6.     Enter the Mgmt0 IPv4 netmask.

7.     Enter the IPv4 address of the default gateway.

8.     Enter the cluster IPv4 address.

9.     Enter the DNS IPv4 address.

10.  Enter the default domain name.

11.  Wait for the login prompt to make sure the configuration has been saved.

Configure Fabric Interconnect B

To configure Fabric Interconnect B, follow these steps:

1.     Connect to the console port on the second Cisco UCS 6332 Fabric Interconnect.

2.     Enter the admin password that was configured for the first Fabric Interconnect.

3.     Enter the Mgmt0 IPv4 address.

4.     Answer yes to save the configuration.

5.     Wait for the login prompt to confirm that the configuration has been saved.

For more information about configuring Cisco UCS 6332 Series Fabric Interconnect, go to: https://www.cisco.com/c/en/us/td/docs/unified_computing/ucs/ucs-manager/GUI-User-Guides/Getting-Started/4-0/b_UCSM_Getting_Started_Guide_4_0.html

Log into Cisco UCS Manager

To log into Cisco UCS Manager, follow these steps:

1.     Open a Web browser and navigate to the Cisco UCS 6332 Fabric Interconnect cluster address.

2.     Click the Launch link to download the Cisco UCS Manager software.

3.     If prompted to accept security certificates, accept as necessary.

4.     When prompted, enter admin for the username and enter the administrative password.

5.     Click Login to log in to the Cisco UCS Manager.

Upgrade Cisco UCS Manager Software to Version 4.0(4c)

This document assumes you’re using Cisco UCS 4.0(4c). Refer to the Cisco UCS 4.0 Release  (upgrade Cisco UCS Manager software and Cisco UCS 6332 Fabric Interconnect software to version 4.0(2a) ). Also, make sure the Cisco UCS C-Series version 4.0(4c) software bundles are installed on the Fabric Interconnects.

Configure Cisco UCS Manager

The following are the high-level steps involved for a Cisco UCS Manager configuration:

1.     Configure Fabric Interconnects for a Cluster Setup.

2.     Set Fabric Interconnects to Fibre Channel End Host Mode.

3.     Synchronize Cisco UCS to NTP.

4.     Configure Fabric Interconnects for Rack or Chassis and Blade Server Discovery.

5.     Configure Global Policies.

6.     Configure Server Ports.

7.     Configure LAN on Cisco UCS Manager.

8.     Configure Ethernet LAN Uplink Ports.

9.     Set QoS system class and Jumbo Frames in both the Cisco Fabric Interconnect.

10.  Create Uplink Port Channels to Cisco Nexus Switches.

11.  Configure FC SAN Uplink Ports

12.  Configure VLAN

13.  Configure IP, UUID, Server, MAC Pool and policy:

a.     IP Pool Creation

b.     UUID Suffix Pool Creation

c.     Server Pool Creation

d.     Configure Server BIOS Policy.

e.     Create Adapter Policy.

f.      Configure Default Maintenance Policy.

g.     Configure vNIC Template

h.     Create Server Boot Policy

Details for each step are discussed in the following sections.

Synchronize Cisco UCSM to NTP

To synchronize the Cisco UCS environment to the NTP server, follow these steps:

1.     In Cisco UCS Manager, in the navigation pane, click the Admin tab.

2.     Select All > Time zone Management.

3.     In the Properties pane, select the appropriate time zone in the Time zone menu.

4.     Click Save Changes and then click OK.

5.     Click Add NTP Server.

6.     Enter the NTP server IP address and click OK.

7.     Click OK to finish.

8.     Click Save Changes.

Figure 28     Synchronize Cisco UCS Manager to NTP

Configure Global Policies

The rack server and chassis discovery policy determine how the system reacts when you add a new rack server or chassis. We recommend using the platform max value as shown. Using platform max helps ensure that Cisco UCS Manager uses the maximum number of IOM uplinks available.

To configure the global policies, follow this step:

1.     In Cisco UCS Manager; Go to Equipment > Policies (right pane) > Global Policies as shown in Figure 29.

Figure 29     Global Policies in UCSM

Configure Server Ports

Configure Server Ports to initiate Chassis and Blade discovery. To configure server ports, follow these steps:

1.     Go to Equipment > Fabric Interconnects > Fabric Interconnect A > Fixed Module > Ethernet Ports.

2.     Select the ports (for this solution ports are 1-28) which are connected to the Cisco UCS VIC 1387 on Cisco UCS C240 M5 rack server.

3.     Right-click and select Configure as Server Port.

Figure 30     Configure Server Port on Cisco UCS Manager Fabric Interconnect for Server/Chassis Discovery

Figure 31     Ports Status after the Server Discovery

Configure Uplink Ports

Configure Network Ports to connect to the datacenter network switch.

To configure Network ports, follow these steps:

1.     Go to Equipment > Fabric Interconnects > Fabric Interconnect A > Fixed Module > Ethernet Ports.

2.     Select the ports (for this solution ports are 29-32) which are connected to the Cisco Nexus 9000 series switch for northbound network connectivity.

3.     Right-click and select Configure as Network Port.

Figure 32     Configure Network Port on Cisco UCS Manager Fabric Interconnect

After the Server port and network port configuration on Cisco UCS FI 6332, Ports 1-30 are utilized for server management and data traffic and 31-32 will be a Network Port.

Create New Organization

To configure the necessary Organization for the Cisco UCS environment, follow these steps:

1.     In Cisco UCS Manager, click the Servers tab in the navigation pane.

2.     Select root > Sub-Organization.

3.     Right-click Sub-Organization.

4.     Enter the name of the Organization.

5.     Click Ok.

Figure 33     Create New Organization

 

Configure IP, UUID, Server and MAC Pools

IP Pool Creation

An IP address pool on the out of band management network must be created to facilitate KVM access to each compute node in the Cisco UCS domain. To create a block of IP addresses for server KVM access in the Cisco UCS environment, follow these steps:

1.     In Cisco UCS Manager, in the navigation pane, click the LAN tab.

2.     Select Pools > root > Sub-Organizations > UCS-HDP > IP Pools > click Create IP Pool.

3.     Enter name for the IP Pool, select option Sequential to assign IP in sequential order then click Next.

4.     Click Add IPv4 Block.

5.     Enter the starting IP address of the block and the number of IP addresses required, and the subnet and gateway information as shown below.

UUID Suffix Pool Creation

To configure the necessary universally unique identifier (UUID) suffix pool for the Cisco UCS environment, follow these steps:

1.     In Cisco UCS Manager, click the Servers tab in the navigation pane.

2.     Select Pools > root > Sub-Organization > UCS-HDP.

3.     Right-click UUID Suffix Pools and then select Create UUID Suffix Pool.

4.     Enter the name of the UUID name.

5.     Optional: Enter a description for the UUID pool.

6.     Keep the prefix at the derived option and select Sequential in as Assignment Order then click Next.

Figure 34     UUID Suffix Pool Creation

Figure 35     Create a Block of UUID Suffixes

Server Pool Creation

To configure the necessary server pool for the Cisco UCS environment, follow these steps:

1.     In Cisco UCS Manager, click the Servers tab in the navigation pane.

2.     Select Pools > root > Sub-Organization > UCS-HDP> right-click Server Pools > Select Create Server Pool.

3.     Enter name of the server pool.

4.     Optional: Enter a description for the server pool then click Next.

Figure 36     Create Server Pool

5.     Select servers to be used for the deployment and click > to add them to the server pool. In our case we added thirty servers in this server pool.

6.     Click Finish and then click OK.

Figure 37     Add Server in the Server Pool

7.     Once the added Servers are in the Pooled servers, click Finish.

MAC Pool Creation

To configure the necessary MAC address pools for the Cisco UCS environment, follow these steps:

1.     In Cisco UCS Manager, click the LAN tab in the navigation pane.

2.     Select Pools > root > Sub-Organization > UCS-HDP > right-click MAC Pools.

3.     Select Create MAC Pool to create the MAC address pool.

4.     Enter name for MAC pool. Select Assignment Order as “Sequential”.

5.     Enter the seed MAC address and provide the number of MAC addresses to be provisioned.

6.     Click OK and then click Finish.

7.     In the confirmation message, click OK.

Figure 38     Creating a Block of MAC Addresses

 

Configure VLAN

To configure the necessary virtual local area networks (VLANs) for the Cisco UCS environment, follow these steps:

1.     In Cisco UCS Manager, click the LAN tab in the navigation pane.

2.     Select LAN > LAN Cloud.

3.     Right-click VLANs

4.     Select Create VLANs

5.     Enter the name of the VLAN to be used.

6.     Keep the Common/Global option selected for the scope of the VLAN.

7.     Enter <VLAN Number> as the ID of the VLAN ID.

8.     Keep the Sharing Type as None.

Figure 39     Create VLAN

The NIC will carry the data traffic from VLAN13. A single vNIC is used in this configuration and the Fabric Failover feature in Fabric Interconnects will take care of any physical port down issues. It will be a seamless transition from an application perspective.

Figure 40     Create VLANs

Set System Class QoS and Jumbo Frame in Both Cisco Fabric Interconnects

To configure jumbo frames and enable quality of service in the Cisco UCS fabric, follow these steps:

1.     In Cisco UCS Manager, click the LAN tab in the navigation pane.

2.     Select LAN > LAN Cloud > QoS System Class.

3.     In the right pane, click the General tab.

4.     On the Platinum row, enter 9216 in the box under the MTU column.

5.     Click Save Changes.

6.     Click OK.

Figure 41     Configure System Class QoS on Cisco UCS Fabric Interconnects

Create QoS Policies

To create the QoS policy to assign priority based on the class using the Cisco UCS Manager GUI, follow these steps:

1.     Select LAN tab in the left pane in the Cisco UCS Manager GUI.

2.     Select LAN > Policies > root > UCS-HDP > QoS Policies.

3.     Right-click QoS Policies.

4.     Select Create QoS Policy.

Figure 42     Create QoS Policy

Figure 43     Platinum QoS Policy

Create vNIC Templates

To create multiple virtual network interface card (vNIC) templates for the Cisco UCS environment, follow these steps:

1.     In Cisco UCS Manager, click the LAN tab in the navigation pane.

2.     Select Policies > root > Sub-Organization > UCS-HDP > vNIC Template.

3.     Right-click vNIC Templates.

4.     Select Create vNIC Template.

5.     Enter name for vNIC template.

6.     Keep Fabric A selected. Select the Enable Failover checkbox.

7.     Select Updating Template as the Template Type.

8.     Under VLANs, select the checkboxes for desired VLANs to add as part of the vNIC Template.

9.     Set Native-VLAN as the native VLAN.

10.  For MTU, enter 9000.

11.  In the MAC Pool list, select MAC Pool configured.

12.  Select QOS policy created earlier.

13.  Select default Network Control Policy.

14.  Click OK to create the vNIC template.

Figure 44     Create the vNIC Template

Create Host Firmware Package

Firmware management policies allow the administrator to select the corresponding packages for a given server configuration. These policies often include packages for adapter, BIOS, board controller, FC adapters, host bus adapter (HBA) option ROM, and storage controller properties.

To create a firmware management policy for a given server configuration in the Cisco UCS environment, follow these steps:

1.     In Cisco UCS Manager, click the Servers tab in the navigation pane.

2.     Select Policies > root > Sub-Organization > UCS-HDP > Host Firmware Packages.

3.     Right-click Host Firmware Packages.

4.     Select Create Host Firmware Package.

5.     Enter name of the host firmware package.

6.     Leave Simple selected.

7.     Select the version.

8.     Click OK to create the host firmware package.

Figure 45     Host Firmware Package

Create Power Control Policy

To create a power control policy for the Cisco UCS environment, follow these steps:

1.     In Cisco UCS Manager, click the Servers tab in the navigation pane.

2.     Select Policies > root > Sub-Organization > UCS-HDP > Power Control Policies.

3.     Right-click Power Control Policies.

4.     Select Create Power Control Policy.

5.     Select Fan Speed Policy as “Max Power”.

6.     Enter NoPowerCap as the power control policy name.

7.     Change the power capping setting to No Cap.

8.     Click OK to create the power control policy.

Figure 46     Create Power Control Policy

Create Server BIOS Policy

To create a server BIOS policy for the Cisco UCS environment, follow these steps:

1.     In Cisco UCS Manager, click the Servers tab in the navigation pane.

2.     Select Policies > root > Sub-Organization > UCS-HDP > BIOS Policies.

3.     Right-click BIOS Policies.

4.     Select Create BIOS Policy.

5.     Enter the BIOS policy name.

Figure 47     BIOS Configuration

 

Configure Maintenance Policy

To update the default Maintenance Policy, follow these steps:

1.     In Cisco UCS Manager, click the Servers tab in the navigation pane.

2.     Select Policies > root > Sub-Organization > UCS-HDP > Maintenance Policies.

3.     Right-click Maintenance Policies to create a new policy.

4.     Enter name for Maintenance Policy

5.     Change the Reboot Policy to User Ack.

6.     Click Save Changes.

7.     Click OK to accept the change.

Figure 48     Create Server Maintenance Policy

Create the Local Disk Configuration Policy

To create local disk configuration in the Cisco UCS Manager GUI, follow these steps:

1.     Select the Servers tab on the left pane in the Cisco UCS Manager GUI.

2.     Select Policies > root > Sub-Organization > UCS-HDP > Local Disk Config Policies.

3.     Right-click Local Disk Config Policies and Select Create Local Disk Config Policies.

4.     Enter UCS-Boot as the local disk configuration policy name.

5.     Change the Mode to Any Configuration. Check the Protect Configuration box.

6.     Keep the FlexFlash State field as default (Disable).

7.     Keep the FlexFlash RAID Reporting State field as default (Disable).

8.     Click OK to complete the creation of the Local Disk Configuration Policy.

9.     Click OK.

Figure 49     Create the Local Disk Configuration Policy

Create Boot Policy

To create boot policies within the Cisco UCS Manager GUI, follow these steps:

1.     Select the Servers tab in the left pane in the Cisco UCS Manager GUI.

2.     Select Policies > root.

3.     Right-click the Boot Policies.

4.     Select Create Boot Policy.

5.     Enter ucs for the boot policy name.

6.     (Optional) enter a description for the boot policy.

7.     Keep the Reboot on Boot Order Change check box unchecked.

8.     Keep Enforce vNIC/vHBA/iSCSI Name check box checked.

9.     Keep Boot Mode Default (Legacy).

10.  Expand Local Devices > Add CD/DVD and select Add Local CD/DVD.

11.  Expand Local Devices and select Add Local Disk.

12.  Expand vNICs and select Add LAN Boot and enter eth0.

13.  Click OK to add the Boot Policy.

14.  Click OK.

Figure 50     Create Boot Policy for Cisco UCS Server(s)

Create Storage Profile for Individual RAID0

To create the storage profile for the individual RAIDP, follow these steps:

1.     On the UCSM navigation page on the left-hand side select Storage tab.

2.     From the Storage Profiles drop-down list, right-click and select Create Storage Profile.

Figure 51     Create Storage Profile

3.     Enter a name for the Storage Profile and click the LUN Set tab.

4.     Click Add.

5.     Select the properties for the LUN set:

a.     Enter a name for LUN set.

b.     Disk Slot Range – 1 – 24/26/56 (Depends on number of drives installed in a server).

c.     Enter Virtual Drive configuration:

                           i.    Strip Size(kb) – 1024KB

                          ii.    Access Policy – Read Write

                         iii.    Read Policy – Read Ahead

                        iv.    Write Cache Policy – Write Back Good Bbu

                         v.    IO Policy – Direct

                        vi.    Drive Cache – Disable

Figure 52     Set JBOD Disks to Unconfigured Good

Create Storage Policy and Storage Profile

To create a Storage Profile with multiple RAID LUNs, create Storage Policies and attach them to a Storage Profile.

To create a Storage Policy and attach them to a Storage Profile, follow these steps:

1.     Go to the Storage tab on the left side panel selection, select “Storage Policies”.

2.     From the Storage Policies drop-down list, select and right-click “Disk Group Policies”. Select “Create Disk Group Policy”.

3.     Enter name for Disk Group Policy, Select RAID level.

4.     Select “Disk Group Configuration” (Automatic/Manual).

5.     Disk Group Configuration.

6.     Virtual Drive Configuration.

7.     Select Storage Profiles, right-click and select Create Storage Profile.

8.     Enter a name for the Storage profile and click Add.

9.     Enter a Local LUN name and select Auto Deploy.

10.  Check the box for Expand to Available and from the drop-down list select the storage policy you want to attach with the Storage Profile. Click OK.

Figure 53     Storage Policy to Configure Boot LUN for S3260 Server Node(s)

Figure 54     Storage Profile to Configure Individual RAID 0 on Server Node 1: Disk Slot 1-28

Figure 55     Storage Profile to Configure Individual RAID 0 on Server Node 1: Disk Slot 29-56

Create Service Profile Template

To create a service profile template, follow these steps:

1.     In the Cisco UCS Manager, go to Servers > Service Profile Templates > root Sub Organization > UCS-HDP > and right-click “Create Service Profile Template” as shown below.

2.     Enter the Service Profile Template name, Updating Template as type of template and select the UUID Pool that was created earlier. Click Next.

3.     Select Local Disk Configuration Policy tab and select Local Storage policy from the drop-down list.

4.     On Storage Profile Policy; select Storage Profile to attach with the server.

5.     In the networking window, select “Expert” and click “Add” to create vNICs. Add one or more vNICs that the server should use to connect to the LAN.

6.     In the create vNIC menu as vNIC name.

7.     Select vNIC Template as vNIC0 and Adapter Policy as Linux.

8.     In the SAN Connectivity menu, select no vHBAs.

9.     Click Next on the Zoning tab.

10.  Select Let System Perform Placement for vNIC/vHBA Placement. Click Next.

11.  Click Next on the vMedia Policy tab.

12.  Select Boot Policy in the Server Boot Order tab.

13.  Select UserAck maintenance policy, which requires user acknowledgement prior rebooting server when making changes to policy or pool configuration tied to a service profile.

14.  Select the Server Pool policy to automatically assign a service profile to a server that meets the requirements for server qualification based on the pool configuration. Select Power state when the Service Profile is associated to server

15.  On the same page you can configure “Host firmware Package Policy” which helps to keep the firmware in sync when associated to server.

16.  Click Finish to create the Service Profile template.

Create Service Profiles from Template

To create a Service Profile from a template, follow these steps:

1.     Right-click the Service Profile Template and select Create Service profile from Template.

Figure 56     Create Service Profile from Template

2.     Repeat the steps above to create service profile template(s) and service profile(s) for Cisco UCS S3260, Cisco C240 M5 according to different deployment scenario.

Install Red Hat Enterprise Linux 7.6

This section provides detailed procedures for installing Red Hat Enterprise Linux Server using Software RAID (OS based Mirroring) on Cisco UCS C240 M5 servers. There are multiple ways to install the RHEL operating system. The installation procedure described in this deployment guide uses KVM console and virtual media from Cisco UCS Manager.

To install the Red Hat Enterprise Linux 7.6 operating system, follow these steps:

1.     Log into the Cisco UCS Manager.

2.     Select the Equipment tab.

3.     In the navigation pane expand Rack-Mounts and then Servers.

4.     Right-click the server and select KVM console.

5.     In the right pane, click the KVM Console >>.

6.     Click the link to launch the KVM console.

7.     Point the cursor over the top right corner and select the Virtual Media tab.

8.     Click the Activate Virtual Devices found in Virtual Media tab.

9.     Click the Virtual Media tab to select CD/DVD.

10.  Select Map Drive in the Virtual Disk Management windows.

11.  Browse to the Red Hat Enterprise Linux 7.6 installer ISO image file.

12.  Click Open to add the image to the list of virtual media.

13.  Select the Installation option from Red Hat Enterprise Linux 7.6.

14.  Select the language for the installation and click Continue.

15.  Select date and time, which pops up another window as shown below.

16.  Select the location on the map, set the time, and click Done.

17.  Click Installation Destination.

18.  This opens a new window with the boot disks. Make the selection and choose “I will configure partitioning”. Click Done. We selected two M.2 SATA SSDs.

19.  This opens a window to create the partitions. Click the + sign to add a new partition as shown below with a boot partition size 2048 MB.

20.  Click Add Mount Point to add the partition.

21.  Change the device type to RAID and make sure the RAID level is RAID1 (redundancy) and click Update Settings to save the changes.

22.  Click the + sign to create the swap partition of size 2048 MB. Click Add Mount Point.

23.  Change the Device type to RAID and RAID level to RAID1 (Redundancy) and click Update Settings.

24.  Click + to add the / partition. The size can be left empty so it will use the remaining capacity. Click Add Mountpoint.

25.  Change the Device type to RAID and RAID level to RAID1 (Redundancy). Click Update Settings.

26.  Click Done to go back to the main screen and continue the Installation.

27.  Click Software Selection.

28.  Select Infrastructure Server and select the Add-Ons as noted below then click Done:

a.     Network File System Client

b.     Performance Tools

c.     Compatibility Libraries

d.     Development Tools

e.     Security Tools

29.  Click Network and Hostname and configure Hostname and Networking for the Host.

30.  Type in the hostname as shown below.

31.  Click Configure to open the Network Connectivity window. Click IPv4 Settings.

32.  Change the Method to Manual and click Add to enter the IP Address, Netmask and Gateway details.

33.  Click Save, update the hostname, and turn Ethernet ON. Click Done to return to the main menu.

34.  Click Begin Installation in the main menu.

35.  Select Root Password in the User Settings.

36.  Enter the Root Password and click Done.

37.  Once the installation is complete reboot the system.

38.  Repeat steps 1 to 37 to install Red Hat Enterprise Linux 7.6 on Servers 2 through 30.

The hostnames and their corresponding IP addresses are shown in Table 7  .

Table 7    Hostname and IP address

Post OS Install Configuration

Choose one of the nodes of the cluster or a separate node as the Admin Node for management, such as CDH installation, Ansible, creating a local Red Hat repo, and others. In this document, we used rhel01 for this purpose.

Configure /etc/hosts

Setup /etc/hosts on the Admin node; this is a pre-configuration to setup DNS as shown in the next section.

Below are the sample A records for DNS configuration within Linux environment.

To create the host file on the admin node, follow these steps:

1.     Login to the Admin Node (rhel01).

#ssh 10.13.1.31

2.     Populate the host file with IP addresses and corresponding hostnames on the Admin node (rhel01) and other nodes as follows:

3.     On Admin Node (rhel01):

Set Up Passwordless Login

To manage all the nodes in a cluster from the admin node password-less login needs to be setup. It assists in automating common tasks with Ansible, and shell-scripts without having to use passwords.

To enable password-less login across all the nodes when Red Hat Linux is installed across all the nodes in the cluster, follow these steps:

1.     Log into the Admin Node (rhel01).

2.     Run the ssh-keygen command to create both public and private keys on the admin node.

Figure 57     Ssh-keygen

3.     Run the following command from the admin node to copy the public key id_rsa.pub to all the nodes of the cluster.  ssh-copy-id appends the keys to the remote-hosts  .ssh/authorized_keys.

4.     Enter yes for Are you sure you want to continue connecting (yes/no)?

5.     Enter the password of the remote host.

Create a Red Hat Enterprise Linux (RHEL) 7.6 Local Repository

To create a repository using RHEL DVD or ISO on the admin node (in this deployment rhel01 is used for this purpose), create a directory with all the required RPMs, run the “createrepo” command and then publish the resulting repository.

To create a RHEL 7.6 local repository, follow these steps:

1.     Log on to rhel01. Create a directory that would contain the repository.

2.     Copy the contents of the Red Hat DVD to /var/www/html/rhelrepo

3.     Alternatively, if you have access to a Red Hat ISO Image, Copy the ISO file to rhel01.

4.     Log back into rhel01 and create the mount directory.

5.     Copy the contents of the ISO to the /var/www/html/rhelrepo directory.

6.     On rhel01 create a .repo file to enable the use of the yum command.

7.     Copy rheliso.repo file from /var/www/html/rhelrepo to /etc/yum.repos.d on rhel01.

8.     To make use of repository files on rhel01 without httpd, edit the baseurl of repo file /etc/yum.repos.d/rheliso.repo to point repository location in the file system.

Create the Red Hat Repository Database

To create the Red Hat repository database, follow these steps:

1.     Install the “createrepo” package on admin node (rhel01). Use it to regenerate the repository database(s) for the local copy of the RHEL DVD contents.

2.     Run “createrepo” on the RHEL repository to create the repo database on admin node

Figure 58     createrepo

 

Set Up Ansible

To set up Ansible, follow these steps:

1.     Download Ansible rpm from the following link: https://releases.ansible.com/ansible/rpm/release/epel-7-x86_64/ansible-2.7.11-1.el7.ans.noarch.rpm

2.     Run the following command to install ansible:

3.     Verify Ansible installation by running the following commands:

4.     Prepare the host inventory file for Ansible as shown below. Various host groups have been created based on any specific installation requirements of certain hosts.

5.     Verify host group by running the following commands. Figure 59 shows the outcome of the ping command.

Figure 59     Ansible – Ping Hosts

Install httpd

Setting up the RHEL repository on the admin node requires httpd. To set up RHEL repository on the admin node, follow these steps:

1.     Install httpd on the admin node to host repositories:

2.     Add ServerName and make the necessary changes to the server configuration file:

3.     Start httpd:

Disable the Linux Firewall

Set Up All Nodes to use the RHEL Repository

To set up all noes to use the RHEL repository, follow these steps:

1.     Copy the rheliso.repo to all the nodes of the cluster:

2.     Copy the /etc/hosts file to all nodes:

3.     Purge the yum caches:

Upgrade the Cisco Network Driver for VIC1387

The latest Cisco Network driver is required for performance and updates. The latest drivers can be downloaded from the link below:

https://software.cisco.com/download/home/283862063/type/283853158/release/4.0(4)

1.     Copy the rpm on all nodes of the cluster using the following Ansible commands. For this example, the rpm is assumed to be in present working directory of rhel01:

2.     Use the yum module to install the enic driver rpm file on all the nodes through Ansible:

3.     It is recommended to download the kmod-megaraid driver for higher performance. The RPM can be found in the same package at: \Storage\LSI\Cisco_Storage_12G_SAS_RAID_controller\RHEL\RHEL7.6\kmod-megaraid_sas-07.708.03.00_el7.6-2.x86_64.rpm

4.     Copy the rpm on all nodes of the cluster using the following Ansible commands. For this example, the rpm is assumed to be in present working directory of rhel01:

5.     Use the yum module to install the enic driver rpm file on all the nodes through Ansible:

Set Up JAVA

To setup JAVA, follow these steps:

1.     Download jdk-8u211-linux-x64.rpm and src the rpm to admin node (rhel01) from the link: https://www.oracle.com/technetwork/java/javase/downloads/jdk8-downloads-2133151.html

2.     Copy JDK rpm to all nodes:

3.     Extract and Install JDK on all nodes:

4.     Create the following files java-set-alternatives.sh and java-home.sh on admin node (rhel01):

5.     Make the two java scripts created above executable:

6.     Copying java-set-alternatives.sh to all nodes.

7.     Setup Java Alternatives

8.     Make sure correct java is setup on all nodes (should point to newly installed java path).

9.     Setup JAVA_HOME on all nodes.

10.  Display JAVA_HOME on all nodes.

 

11.  Display current java –version.

 

Enable Syslog

Syslog must be enabled on each node to preserve logs regarding killed processes or failed jobs. Modern versions such as syslog-ng and rsyslog are possible, making it more difficult to be sure that a syslog daemon is present.

Use one of the following commands to confirm that the service is properly configured:

Set the ulimit

On each node, ulimit -n specifies the number of inodes that can be opened simultaneously. With the default value of 1024, the system appears to be out of disk space and shows no inodes available. This value should be set to 64000 on every node.

Higher values are unlikely to result in an appreciable performance gain.

To set ulimit, follow these steps:

1.     For setting the ulimit on Redhat, edit “/etc/security/limits.conf” on admin node rhel01 and add the following lines:

2.     Copy the /etc/security/limits.conf file from admin node (rhel01) to all the nodes using the following command:

3.     Make sure that the /etc/pam.d/su file contains the following settings:

Disable SELinux

SELinux can be disabled by editing /etc/selinux/config and changing the SELINUX line to SELINUX=disabled.

To disable SELinux, follow these steps:

1.     The following command will disable SELINUX on all nodes:

2.     Reboot the machine, if needed for SELinux to be disabled in case it does not take effect. It can be checked using the following command:

Set TCP Retries

Adjusting the tcp_retries parameter for the system network enables faster detection of failed nodes. Given the advanced networking features of UCS, this is a safe and recommended change (failures observed at the operating system layer are most likely serious rather than transitory).

To set TCP retries, follow these steps:

1.     Edit the file /etc/sysctl.conf and on admin node rhel01 and add the following lines:

2.     Load the settings from default sysctl file /etc/sysctl.conf by running the following command:

Disable IPv6 Defaults

To disable IPv6 defaults, follow these steps:

1.     Run the following command:

2.     Load the settings from default sysctl file /etc/sysctl.conf:

Disable Swapping

To disable swapping, follow these steps:

1.     Run the following on all nodes. Variable vm.swappiness defines how often swap should be used, 60 is default:

2.     Load the settings from default sysctl file /etc/sysctl.conf and verify the content of sysctl.conf:

Disable Memory Overcommit

To disable Memory Overcommit, follow these steps:

1.     Run the following on all nodes. Variable vm.overcommit_memory=0

2.     Load the settings from default sysctl file /etc/sysctl.conf and verify the content of sysctl.conf:

 

Disable Transparent Huge Pages

Disabling Transparent Huge Pages (THP) reduces elevated CPU usage caused by THP.

To disable Transparent Huge Pages, follow these steps:

1.     You must run the following commands for every reboot; copy this command to /etc/rc.local so they are executed automatically for every reboot:

2.     On the Admin node, run the following commands:

3.     Copy file to each node:

4.     Append the content of file thp_disable to /etc/rc.local:

NTP Configuration

The Network Time Protocol (NTP) is used to synchronize the time of all the nodes within the cluster. The Network Time Protocol daemon (ntpd) sets and maintains the system time of day in synchronism with the timeserver located in the admin node (rhel01). Configuring NTP is critical for any Hadoop Cluster. If server clocks in the cluster drift out of sync, serious problems will occur with HBase and other services.

1.     Configure /etc/ntp.conf on the admin node only with the following contents:

2.     Create /root/ntp.conf on the admin node and copy it to all nodes:

3.     Copy ntp.conf file from the admin node to /etc of all the nodes by executing the following commands in the admin node (rhel01):

4.     Run the following to syncronize the time and restart NTP daemon on all nodes:

5.     Make sure to restart of NTP daemon across reboots:

6.     Verify NTP is up and running in all nodes by running the following commands:

 

7.     Install the Chrony service:

8.     Activate the Chrony service at boot:

9.     Start the Chrony service:

Install Megaraid StorCLI

This section explains the steps needed to install StorCLI (Storage Command Line Tool) which is a command line interface designed to be easy to use, consistent, and script. For more details, go to: https://docs.broadcom.com/docs/12352476

To install StorCLI, follow these steps:

1.     Download StorCLI:  https://www.broadcom.com/support/download-search/?pg=&pf=&pn=&po=&pa=&dk=storcli.

2.     Extract the .zip file and copy storcli-1.23.02-1.noarch.rpm from the linux directory.

3.     Download StorCLI and its dependencies and transfer to Admin node:

4.     Copy storcli rpm to all the nodes using the following commands:

5.     Run this command to install storcli on all the nodes:

6.     Run this command to copy storcli64 to root directory:

7.     Run this command to check the state of the disks:

Configure the Filesystem for NameNodes and DataNodes

The following script formats and mounts the available volumes on each node whether it is NameNode or Data node. OS boot partition will be skipped. All drives are mounted based on their UUID as /data/disk1, /data/disk2, etc. To configure the filesystem for NameNodes and DataNodes, follow these steps:

1.     On the Admin node, create a file containing the following script:

#vi /root/driveconf.sh

2.     To create partition tables and file systems on the local disks supplied to each of the nodes, run the following script as the root user on each node:

3.     Run the following command to copy driveconf.sh to all the nodes:

4.     Run the following command from the admin node to run the script across all data nodes:

5.     Run the following from the admin node to list the partitions and mount points:

Delete Partitions

To delete a partition, follow these steps:

1.     Run the mount command (‘mount’) to identify which drive is mounted to which device /dev/sd<?>

2.     umount the drive for which partition is to be deleted and run fdisk to delete as shown below.

Cluster Verification

This section explains the steps to create the script cluster_verification.sh that helps to verify the CPU, memory, NIC, and storage adapter settings across the cluster on all nodes. This script also checks additional prerequisites such as NTP status, SELinux status, ulimit settings, JAVA_HOME settings and JDK version, IP address and hostname resolution, Linux version and firewall settings.

To verify a cluster, follow these steps:

1.     Create the script cluster_verification.sh as shown, on the Admin node (rhel01).

2.     Change permissions to executable:

3.     Run the Cluster Verification tool from the admin node. This can be run before starting Hadoop to identify any discrepancies in Post OS Configuration between the servers or during troubleshooting of any cluster / Hadoop issues:

Install Cloudera

Cloudera’s Distribution including Apache Hadoop (CDH) is an enterprise grade, hardened Hadoop distribution. CDH offers Apache Hadoop and several related projects into a single tested and certified product. It offers the latest innovations from the open source community with the testing and quality expected from enterprise quality software.

Prerequisites for CDH Installation

This section details the prerequisites for the CDH installation, such as setting up CDH Repo.

Cloudera Manager Repository

1.     From a host connected to the Internet, download the Cloudera’s repositories as shown below and transfer it to the admin node:

2.     Download Cloudera Manager Repository:

3.     Run the following command to move the RPMs:

4.     Copy the repository directory to the admin node (rhel1):

5.     On admin node (rhel1) run create repo command:

6.     Create the Cloudera Manager repo file with following contents:

 

7.     Copy the file cloudera-manager.repo into /etc/yum.repos.d/ on the admin node to enable it to find the packages that are locally hosted:

Set Up the Local Parcels for CDH 6.2.0

From a host connected the internet, download the appropriate CDH 6.2.0 parcels that are meant for RHEL7.6 from the URL: http://archive.cloudera.com/cdh6/6.2.0/parcels/ and place them in the directory "/var/www/html/CDH6.2.0parcels" of the Admin node.

The following are the required files for RHEL7.6:

·       CDH-6.2.0-1.cdh6.2.0.p0.967373-el7.parcel

·       CDH-6.2.0-1.cdh6.2.0.p0.967373-el7.parcel.sha1

·       manifest.json

Download Parcels

To download parcels, follow these steps:

1.     From a host connected to the Internet, download the Cloudera’s parcels as shown below and transfer it to the admin node:

2.     Download parcels:

3.     Now edit the /tmp/clouderarepo/CDH6.2.0parcels/manifest.json file and remove the scripts that are not meant for RHEL7.6. Below is that script which can be copy and pasted.

4.     Copy /tmp/clouderarepo/CDH6.2.0parcels to the admin node (rhel01):

5.     Verify that these files are accessible by visiting the URL http://10.13.1.31/CDH6.2.0parcels/ in admin node.

Set Up the MariaDB Database for Cloudera Manager

You will set up the following for Cloudera Manager:

·       Install the MariaDB Server

·       Configure and Start the MariaDB Server

·       Install the MariaDB/MySQL JDBC Driver

·       Create Databases for Activity Monitor, Reports Manager, Hive Metastore Server, Sentry Server, Cloudera Navigator Audit Server, and Cloudera Navigator Metadata Server

Install the MariaDB Server

To use a MariaDB database, follow these steps:

1.     In the admin node where Cloudera Manager will be installed, use the following command to install the mari-adb/mysql server.

2.     To configure and start the MySQL Server, stop the MariaDB server if it is running.

3.     Move the old InnoDB log if exists.

4.     Move files /var/lib/mysql/ib_logfile0 and /var/lib/mysql/ib_logfile1 out of/var/lib/mysql/ to a backup location.

5.     Determine the location of the option file, my.cnf and edit/add following lines:

6.     Make sure MySQL Server starts at boot:

7.     Start the MySQL Server:

8.     Set the MySQL root password on admin node (rhel01)

Install the MySQL JDBC Driver

Install the JDBC driver on the Cloudera Manager Server host, as well as any hosts that run the Activity Monitor, Reports Manager, Hive Metastore Server, Sentry Server, Cloudera Navigator Audit Server, and Cloudera Navigator Metadata Server roles.

To install the MySQL JDBC driver, follow these steps:

1.     From a host connected to the Internet, download the MySQL JDBC Driver and transfer it to the admin node. Download the MySQL JDBC driver from the URL http://www.mysql.com/downloads/connector/j/5.1.html.

2.     Copy mysql-connector-java-5.1.47.tar.gz to admin node(rhel01)

3.     Log in to the admin node and extract the file:

4.     Create the /usr/share/java/ directory on the admin node (rhel01):

5.     Go to the mysql-connector-java-5.1.47 directory on the admin node (rhel1) and copy mysql-connector-java-5.1.47-bin.jar to /usr/share/java/:

6.     Creating Databases for Activity Monitor, Reports Manager, Hive Metastore Server, Navigator Audit Server, and Navigator Metadata Server:

7.     In the admin node Log into MySQL as the root user:

8.     Enter the password that was previously supplied.

9.     Create databases for the Activity Monitor, Reports Manager and Hive Metastore Server using the command below:

Cloudera Manager Installation

The following sections describe how to install Cloudera Manager and then using Cloudera Manager to install CDH 6.2.0.

Set Up the Cloudera Manager Server Database

The Cloudera Manager Server Database stores information about service and host configurations.

Install Cloudera Manager

Cloudera Manager, an end to end management application, is used to install and configure CDH. During CDH Installation, Cloudera Manager's Wizard will help to install Hadoop services on all nodes using the following procedure:

·       Discovery of the cluster nodes

·       Configure the Cloudera parcel or package repositories

·       Install Hadoop, Cloudera Manager Agent (CMA) and Impala on all the cluster nodes.

·       Install the Oracle JDK if it is not already installed across all the cluster nodes.

·       Assign various services to nodes.

·       Start the Hadoop services

To install Cloudera Manager, follow these steps:

1.     Update the repo files to point to local repository.

2.     Install the Oracle Java Development Kit on the Cloudera Manager Server host.

 

3.     Install the Cloudera Manager Server packages either on the host where the database is installed, or on a host that has access to the database:

 

Prepare a Cloudera Manager Server External Database

To prepare a Cloudera Manager Server external database, follow these steps:

1.     Run the scm_prepare_database.sh script on the host where the Cloudera Manager Server package is installed (rhel1) admin node:

2.     Verify the database connectivity using the following command.

The MySQL External database setup is complete.

Start the Cloudera Manager Server

To start the Cloudera Manager Server, follow these steps:

1.     Start the Cloudera Manager Server:

2.     Once the installation of Cloudera Manager is complete, install CDH5 using the Cloudera Manager Web interface.

Install Cloudera Enterprise Data Hub (CDH6)

To install the Cloudera Enterprise Data Hub, follow these steps:

1.     Log into the Cloudera Manager. Enter "admin" for both the Username and Password fields.

2.     Click Continue.

3.     If you do not have a Cloudera license, select Cloudera Enterprise Data Hub Trial Edition. If you do have a Cloudera license, Click “Upload License” and select your license.

4.     Based on your requirements, choose the appropriate Cloudera Editions for the Installation.

Figure 60     Installing Cloudera Enterprise

5.     Click Continue.

6.     Enter a name for the Cluster.

7.     Specify the hosts that are part of the cluster using their IP addresses or hostname. The figure below shows a pattern that specifies the IP addresses range.

8.     After the IP addresses or hostnames are entered, click Search.

9.     Cloudera Manager will "discover" the nodes in the cluster. Verify that all desired nodes have been found and selected for installation.

Edit the Cloudera Enterprise Parcel Settings to Use the CDH 5.13.0 Parcels

To edit the Cloudera Enterprise Parcel settings, follow these steps:

1.     Open another tab in the same browser window and visit the URL: http://10.13.1.31:7180/cmf/parcel/status for modifying the parcel settings.

2.     Click Configuration.

3.     Click to remove the entire remote repository URLs and add the URL to the location where we kept the CDH 6.2.0 parcels i.e. http://10.13.1.31/CDH6.2.0parcels/.

4.     Click Save Changes to finish the configuration.

5.     Navigate back to the Cloudera installation home page, http://10.13.1.31:7180.

6.      Click Continue on the confirmation page.

7.     For the method of installation, select the Use Parcels (Recommended) radio button.

8.     For the CDH version, select the CDH-6.2.0-1.cdh6.2.0.p0.967373 radio button.

9.     For the specific release of Cloudera Manager, select the Custom Repository radio button.

10.  Enter the URL for the repository within the admin node. http://10.13.1.50/clouderarepo/cloudera-manager and click Continue.

11.  Select Install Oracle Java SE Development Kit 9JDK 8). Click Continue.

12.  Provide SSH login credentials for the cluster and click Continue.

Installation using parcels begins.

13.  Run inspect the hosts and network performance test through Cloudera Manager on which it has just performed the installation.

14.  Review and verify the summary. Click Continue.

15.  Select services that need to be started on the cluster.

16.  This is a critical step in the installation. Inspect and customize the role assignments of all the nodes based on your requirements and click Continue.

17.  Reconfigure the service assignment to match Table 8. 

Table 8    Service/Role Assignment

Set Up the Database

The role assignment recommendation above is for clusters of up to 64 servers. For clusters larger than 64 nodes, use the HA recommendation defined in Table 8. 

To set up the database, follow these steps:

1.     In the Database Host Name sections use port 3306 for TCP/IP because connection to the remote server always uses TCP/IP.

2.     Enter the Database Name, username and password that were used during the database creation stage earlier in this document.

3.     Click Test Connection to verify the connection and click Continue.

4.     Review and customize the configuration changes based on your requirements.

5.     Click Continue to start running the cluster services.

6.     Hadoop services are installed, configured, and now running on all the nodes of the cluster. Click Finish to complete the installation.

7.     Cloudera Manager now displays the status of all Hadoop services running on the cluster.

Scale the Cluster

The role assignment recommendation above is for cluster with at least 64 servers and in High Availability. For smaller cluster running without High Availability the recommendation is to dedicate one server for NameNode and a second server for secondary name node and YARN Resource Manager. For larger clusters larger than 28 nodes the recommendation is to dedicate one server each for name node, YARN Resource Manager and one more for running both NameNode (High Availability) and Resource Manager (High Availability) as in the table (no Secondary NameNode when in High Availability).

This implies that there will be at least 3 master nodes, running the NameNode, YARN Resource manager, the failover counter-part being designated to run on another node and a third node that would have similar capacity as the other two nodes.

All the three nodes will also need to run zookeeper and quorum journal node services. It is also recommended to have a minimum of 7 DataNodes in a cluster. Please refer to the next section for details on how to enable HA.

Enable High Availability

HDFS High Availability

The HDFS High Availability feature provides the option of running two NameNodes in the same cluster, in an Active/Passive configuration. These are referred to as the Active NameNode and the Standby NameNode. Unlike the Secondary NameNode, the Standby NameNode is a hot standby, allowing a fast failover to a new NameNode in the case that a machine crashes, or a graceful administrator-initiated failover for the purpose of planned maintenance. There cannot be more than two NameNodes.

For more information go to: https://www.cloudera.com/documentation/enterprise/latest/topics/cdh_hag_hdfs_ha_enabling.html

Set Up HDFS High Availability

The Enable High Availability workflow leads through adding a second (standby) NameNode and configuring JournalNodes. During the workflow, Cloudera Manager creates a federated namespace. To set up HDFS High Availability, follow these steps:

1.     Log into the admin node (rhel01) and create the Edit directory for the JournalNode:

2.     Log into the Cloudera manager and go to the HDFS service.

3.     Select Actions> Enable High Availability. A screen showing the hosts that are eligible to run a standby NameNode and the JournalNodes displays.

4.     Specify a name for the nameservice or accept the default name nameservice1 and click Continue.

5.     In the NameNode Hosts field, click Select a host. The host selection dialog displays.

6.     Check the checkbox next to the hosts (rhel2) where the standby NameNode is to be set up and click OK.

7.     In the JournalNode Hosts field, click Select hosts. The host selection dialog displays.

8.     Check the checkboxes next to an odd number of hosts (a minimum of three) to act as JournalNodes and click OK. We used the same nodes for the Zookeeper nodes.

9.     Click Continue.

10.  In the JournalNode Edits Directory property, enter a directory location created earlier in step 1 for the JournalNode edits directory into the fields for each JournalNode host.

11.  Extra Options: Decide whether Cloudera Manager should clear existing data in ZooKeeper, Standby NameNode, and JournalNodes. If the directories are not empty (for example, re-enabling a previous HA configuration), Cloudera Manager will not automatically delete the contents—select to delete the contents by keeping the default checkbox selection. The recommended default is to clear the directories.

12.  Click Continue.

13.  Cloudera Manager executes a set of commands that will stop the dependent services, delete, create, and configure roles and directories as appropriate, create a nameservice and failover controller, and restart the dependent services and deploy the new client configuration.

14.  In the next screen additional steps are suggested by the Cloudera Manager to update the Hue and Hive metastore. Click Finish.

The following subsections explain configuring Hue and Hive for High Availability as needed.

Configure Hive Metastore to Use HDFS High Availability

To configure the Hive Megastore to use HDFS High Availability, follow these steps:

1.     Go the Hive service.

2.     Select Actions > Stop.

3.     Click Stop to confirm the command.

4.     Back up the Hive metastore database (if any existing data is present).

5.     Select Actions> Update Hive Metastore NameNodes and confirm the command.

6.     Select Actions> Start.

7.     Restart the Hue and Impala services if stopped prior to updating the Metastore.

Configure Hue to Work with HDFS High Availability

To configure Hue to work with HDFS High Availability, follow these steps:

1.     Go to the HDFS service.

2.     Click the Instances tab.

3.     Click Add Role Instances.

4.     Select the text box below the HttpFS field. The Select Hosts dialog displays.

5.     Select the host on which to run the role and click OK.

6.     Click Continue.

7.     Check the checkbox next to the HttpFS role and select Actions for Selected> Start.

8.     After the command has completed, go to the Hue service.

9.     Click the Configuration tab.

10.  Locate the HDFS Web Interface Role property or search for it by typing its name in the Search box.

11.  Select the HttpFS role that was just created instead of the NameNode role and save your changes.

12.  Restart the Hue service.

YARN High Availability

The YARN Resource Manager (RM) is responsible for tracking the resources in a cluster and scheduling applications (for example, MapReduce jobs). Before CDH 5, the RM was a single point of failure in a YARN cluster. The RM high availability (HA) feature adds redundancy in the form of an Active/Standby RM pair to remove this single point of failure. Furthermore, upon failover from the Standby RM to the Active, the applications can resume from their last check-pointed state; for example, completed map tasks in a MapReduce job are not re-run on a subsequent attempt. This allows events such the following to be handled without any significant performance effect on running applications.

·       Unplanned events such as machine crashes.

·       Planned maintenance events such as software or hardware upgrades on the machine running the ResourceManager.

For more information, go to: https://www.cloudera.com/documentation/enterprise/latest/topics/cdh_hag_rm_ha_config.html#xd_583c10bfdbd326ba--43d5fd93-1410993f8c2--7f77

Set Up YARN High Availability

To set up YARN high availability, follow these steps:

1.     Log into the Cloudera manager and go to the YARN service.

2.     Select Actions> Enable High Availability.

3.     A screen showing the hosts that are eligible to run a standby ResourceManager displays.

4.     The host where the current ResourceManager is running is not available as a choice.

5.     Select the host (rhel3) where the standby ResourceManager is to be installed and click Continue.

6.     Cloudera Manager proceeds to execute a set of commands that stop the YARN service, add a standby ResourceManager, initialize the ResourceManager high availability state in ZooKeeper, restart YARN, and redeploy the relevant client configurations.

7.     Click Finish once the installation is completed successfully.

Configure Yarn (MR2 Included) and HDFS Services

The parameters in Table 9  and Table 10  are used for Cisco UCS Integrated Infrastructure for Big Data and Analytics Performance Optimized cluster configuration described in this document. These parameters are to be changed based on the cluster configuration, number of nodes and specific workload.

Table 9    YARN

Table 10      HDFS

Configure Spark

The two main resources that Spark (and YARN) are dependent on are CPU and memory. Disk and network I/O, play a part in Spark performance as well, but neither Spark nor YARN currently can actively manage them. Every Spark executor in any application has the same fixed number of cores and same fixed heap size. The number of cores can be specified with the executor-cores flag when invoking spark-submit, spark-shell, and pyspark from the command line, or by setting the spark.executor.cores property in the spark-defaults.conf file or in the SparkConf object.

And the heap size can be controlled with the executor-memory flag or the spark.executor.memory property. The cores property controls the number of concurrent tasks an executor can run, executor-cores = 5 mean that each executor can run a maximum of five tasks at the same time. The memory property impacts the amount of data Spark can cache, as well as the maximum sizes of the shuffle data structures used for grouping, aggregations, and joins.

The num-executors command-line flag or spark.executor.instances configuration property control the number of executors requested. Dynamic Allocation can be enabled from CDH5.4 instead setting the spark.dynamicAllocation.enabled to true. Dynamic allocation enables a Spark application to request executors when there is a backlog of pending tasks and free up executors when idle.

Asking for five executor cores will result in a request to YARN for five virtual cores. The memory requested from YARN is a little more complex for a couple reasons:

·       executor-memory/spark.executor.memory controls the executor heap size, but JVMs can also use some memory off heap, for example for VM overhead, interned Strings and direct byte buffers. The value of the spark.yarn.executor.memoryOverhead property is added to the executor memory to determine the full memory request to YARN for each executor. It defaults to max (384, 0.10 * spark.executor.memory).

·       YARN may round the requested memory up a little. YARN’s yarn.scheduler.minimum-allocation-mb and yarn.scheduler.increment-allocation-mb properties control the minimum and increment request values respectively.

·       The application master is a non-executor container with the special capability of requesting containers from YARN, takes up resources of its own that must be budgeted in. In yarn-client mode, it defaults to a 1024MB and one vcore. In yarn-cluster mode, the application master runs the driver, so it’s often useful to add its resources with the –driver-memory and –driver-cores properties.

·       Running executors with too much memory often results in excessive garbage collection delays. 64GB is a rough guess at a good upper limit for a single executor.

·       A good estimate is that at most five tasks per executor can achieve full write throughput, so it’s good to keep the number of cores per executor around that number.

·       Running tiny executors (with a single core and just enough memory needed to run a single task, for example) throws away the benefits that come from running multiple tasks in a single JVM. For example, broadcast variables need to be replicated once on each executor, so many small executors will result in many more copies of the data.

Tune Resource Allocation for Spark

Below is an example of configuring a Spark application to use as much of the cluster as possible, we are using an example cluster with 16 nodes running NodeManagers, each equipped with 56 cores and 256GB of memory. yarn.nodemanager.resource.memory-mb and yarn.nodemanager. resource.cpu-vcores should be set to 180 * 1024 = 184320 (megabytes) and 48 respectively.

This configuration results in four executors on all nodes except for the one with the AM, which will have three executors.

Submit a Job

In yarn-cluster mode, the local directories used by the Spark executors and the Spark driver will be the local directories configured for YARN (Hadoop YARN config yarn.nodemanager.local-dirs). If the user specifies spark.local.dir, it will be ignored.

In yarn-client mode, the Spark executors will use the local directories configured for YARN while the Spark driver will use those defined in spark.local.dir. The Spark driver does not run on the YARN cluster in yarn-client mode, only the Spark executors do.

Every Spark stage has several tasks, each of which processes data sequentially. In tuning Spark jobs, this parallelism number is the most important parameter in determining performance. The number of tasks in a stage is the same as the number of partitions in the last RDD in the stage. The number of partitions in an RDD is the same as the number of partitions in the RDD on which it depends, with a couple exceptions: the coalesce transformation allows creating an RDD with fewer partitions than its parent RDD, the union transformation creates an RDD with the sum of its parents’ number of partitions, and Cartesian creates an RDD with their product.

The primary concern is that the number of tasks will be too small. If there are fewer tasks than slots available to run them in, the stage won’t be taking advantage of all the CPU available.

If the stage in question is reading from Hadoop, your options are:

·       Use the repartition transformation, which will trigger a shuffle.

·       Configure your InputFormat to create more splits.

·       Write the input data out to HDFS with a smaller block size.

If the stage is getting its input from another stage, the transformation that triggered the stage boundary will accept a numPartitions argument.

The most straightforward way to tune the number of partitions is experimentation: Look at the number of partitions in the parent RDD and then keep multiplying that by 1.5 until performance stops improving.

In contrast with MapReduce for Spark when in doubt, it is almost always better to be on the side of a larger number of tasks (and thus partitions).

Shuffle Performance Improvement

spark.shuffle.compress true (compress map output files)

spark.broadcast.compress true (compress broadcast variables before sending them)

spark.io.compression.codec org.apache.spark.io.SnappyCompressionCodec (codec used to compress internal data such as RDD partitions, broadcast variables and shuffle outputs)

spark.shuffle.spill.compress true (Whether to compress data spilled during shuffles.)

spark.shuffle.io.numConnectionsPerPeer 4 (Connections between hosts are reused in order to reduce connection buildup for large clusters. For clusters with many hard disks and few hosts, this may result in insufficient concurrency to saturate all disks, and so users may consider increasing this value.)

spark.shuffle.file.buffer 64K (Size of the in-memory buffer for each shuffle file output stream. These buffers reduce the number of disk seeks and system calls made in creating intermediate shuffle file)

Improve Serialization Performance

Serialization plays an important role in the performance of any distributed application. Often, this will be the first thing that should be tuned to optimize a Spark application.

spark.serializer org.apache.spark.serializer.KryoSerializer (when speed is necessary)

spark.kryo.referenceTracking false

spark.kryoserializer.buffer 2000 (If the objects are large, may need to increase the size further to fit the size of the object being deserialized).

SparkSQL is ideally suited for mixed procedure jobs where SQL code is combined with Scala, Java, or Python programs. In general, the SparkSQL command line interface is used for single user operations and ad hoc queries.

For multi-user SparkSQL environments, it is recommended to use a Thrift server connected via JDBC.

Spark SQL Tuning

Below are some guidelines for Spark SQL tuning:

To compile each query to Java bytecode on the fly, turn on sql.codegen. This can improve performance for large queries but can slow down very short queries.

spark.sql.codegen true

spark.sql.unsafe.enabled true

Configuration of in-memory caching can be done using the setConf method on SQLContext or by running SET key=value commands using SQL.

spark.sql.inMemoryColumnarStorage.compressed true (will automatically select a compression codec for each column based on statistics of the data)

spark.sql.inMemoryColumnarStorage.batchSize 5000 (Controls the size of batches for columnar caching. Larger batch sizes can improve memory utilization and compression, but risk OOMs when caching data)

The columnar nature of the ORC format helps avoid reading unnecessary columns, but it is still possible to read unnecessary rows. ORC avoids this type of overhead by using predicate push-down with three levels of built-in indexes within each file: file level, stripe level, and row level. This combination of indexed data and columnar storage reduces disk I/O significantly, especially for larger datasets where I/O bandwidth becomes the main bottleneck for performance.

By default, ORC predicate push-down is disabled in Spark SQL. To obtain performance benefits from predicate push-down, enable it explicitly, as follows:

spark.sql.orc.filterPushdown=true

In SparkSQL to automatically determine the number of reducers for joins and groupbys, use the parameter,

spark.sql.shuffle.partitions 200, (default value is 200)

This property can be put into hive-site.xml to override the default value.

Set log to WARN in log4j.properties to reduce log level.

Compression for Hive

Set the following Hive parameters to compress the Hive output files using Snappy compression:

Change the Log Directory for All Applications

To change the default log from the /var prefix to /data/disk1, follow these steps:

1.     Log into the cloudera home page and click My Clusters.

2.     From the configuration drop-down list select “All Log Directories.”

3.     Click Save.

When building an infrastructure to enable this modernized architecture which could scale to thousands of nodes, operational efficiency can’t be an afterthought.

To bring in seamless operation of the application at this scale, you need:

·       Infrastructure automation of Cisco UCS servers with service profiles and Cisco Data Center network automation with application profiles with Cisco ACI

·       Centralized Management and Deep telemetry and Simplified granular trouble-shooting capabilities and • Multi-tenancy allowing application workloads including containers, micro-services, with the right level of security and SLA for each workload.

·       Cisco UCS with Cisco Intersight and Cisco ACI can enable this cloud scale architecture deployed and managed with ease

For More Information

For additional information, see the following resources:

·       To find out more about Cisco UCS big data solutions, see http://www.cisco.com/go/bigdata.

·       To find out more about Cisco UCS big data validated designs, see http://www.cisco.com/go/bigdata_design

·       To find out more about Cisco UCS AI/ML solutions, see http://www.cisco.com/go/ai-compute

·       To find out more about Cisco ACI solutions, see http://www.cisco.com/go/aci

·       To find out more about Cisco validated solutions based on Software Defined Storage, see https://www.cisco.com/c/en/us/solutions/data-center-virtualization/software-defined-storage-solutions/index.html

This section provides the BOM for the 16 Nodes Hadoop Base Rack, 6 Nodes Data Science (AI/ML) Rack and 8 Nodes Hadoop tiered storage. See Table 11  for BOM for the Hadoop Base rack, Table 12  for BOM for Data Science(AI/ML) Expansion Rack, Table 13   for BOM for Hadoop Tiered Storage. Table 14  , Table 15  , and Table 16   for software components. Table 17   lists Cloudera SKUs available from Cisco.

Table 11      Bill of Materials for Cisco UCS C240M5SX Hadoop Nodes Base Rack

Table 12      Bill of Materials for Hadoop Nodes Expansion Rack

Table 13      Bill of Materials for Hadoop Nodes Tiered Storage Rack

Table 14      Bill of Materials for CDSW Nodes Rack

Table 15      Red Hat Enterprise Linux License

Table 16      Cloudera Data Science Workbench Software

Table 17      Cloudera SKU’s Available at Cisco

Configure Cisco Boot Optimized M.2 RAID Controller

Beginning with 4.0(4a), Cisco UCS Manager supports Cisco boot optimized M.2 RAID controller (UCS-M2-HWRAID), which is based on Marvell® 88SE92xx PCIe to SATA 6Gb/s controller.

The following M.2 drives are managed by the Cisco boot optimized M.2 RAID controller:

·       240GB M.2 6G SATA SSD

·       960GB M.2 6G SATA SSD

The following are the limitations of the Cisco boot optimized M.2 RAID controller:

·       Existing LUN migration is not supported.

·       Local Disk Configuration policy is not supported.

·       Entire disk capacity is used while creating single LUN.

·       LUN is created using the Local LUN tab (see Configuring Local LUNs) under storage profile and not using the controller definitions.

·       You cannot mix different capacity M.2 drives.

To create a Disk Group Policy and Storage Profile Policy to be attach with Service Profile for Cisco Optimized M.2 RAID Controller follow the steps in the following sections.

Configure Disk Group Policy

To configure the disk group policy, follow these steps:

1.     In the UCSM WebUI, Go to storage tab. In the Storage Policy section, right-click Disk Group Policies. Click Create Disk Group Policy.

2.     Enter a name and description for the new Disk Group Policy. Select Manual Disk Group Configuration. Click Add.

3.     Enter Slot Number 253 for the first disk. Click OK.

4.     Click Add to add second disk, enter Slot Number 254.

5.     In Virtual Drive Configuration section leave all option as Platform Default. Click OK.

Configure Storage Profile

To configure the storage profile, follow these steps:

1.     In the Storage Profiles section, select Storage Profiles. Right-click and select Create Storage Profile.

2.     Enter a name for the Storage Profile. Click Add.

3.     Enter a name for the Local LUN to be created, Click Auto Deploy, check the box for Expand to Available, and from the drop-down list for Disk Group Configuration, select RAID 1 Disk Group Policy created for M.2 SATA Disks. Click OK.

 

4.     Attach a Storage Profile created to a Service profile or create a new Service Profile.

5.     Go to the Storage tab on the Service Profile, select Storage Profile. Select Modify Storage Profile. Select Storage Profile created for M.2 SATA Disks.

Figure 61     Example of the Service Profile Associated to a Cisco UCS C240 M5 server with Cisco UCS-M2-HWRAID and 2 240GB M.2 SATA SSD Installed

Figure 62     Example of Virtual Drive Created from 2 M.2 SATA SSD

Apply Storage Profile in Service Profile Template

To create a new Service Profile template or update an existing template for Service Profile to attach a newly created Storage Profile for Cisco Boot Optimized RAID Controller, follow these steps:

1.     Go to Service Profile Template.

2.     Select Storage tab in Service Profile Template.

3.     Select Storage Profile tab. Click Modify Storage Profile.

4.     From the Storage Profile drop-down list, select Storage Profile for Cisco Boot Optimized RAID Controller.

5.     If updating a Service Profile Template, once saved the changes in the configuration change in the Service Profile Template and are automatically applied to all Service Profile binded with the template.

Install RHEL 7.6 on Cisco Optimized M.2 RAID Controller

To install Red Hat Enterprise Linux 7.6 OS on Cisco UCS server with Virtual Drive created from Cisco Optimized M.2 RAID Controller (UCS-M2-HWRAID) in UEFI Boot Mode, follow these steps:

1.     On the Welcome screen, select a language and click Continue.

2.     Select DATE & TIME.

3.     Select region and City.

4.     Select SOFTWARE SELECTION.

5.     Select Infrastructure Server in Base Environment. For Add-Ons for the selected environment, select:

-      Network File System Client

-      Performance Tools

-      Compatibility Libraries

-      Development Tools

-      Security Tools

6.     Select Installation Destination.

 

7.     Select the Virtual Drive created from M.2 SATA SSDs. Select I will Configure Partitioning. Click DONE.

8.     Click the + button to add manual configuration to install Red Hat Enterprise Linux 7.6. Enter /boot/efi as mount point and 2048mb as Desired Capacity.

9.     Click the + button for the following mountpoint and desired capacity as shown below.

Table 18      Mount Point and Desired Capacity for RHEL Installation

10.  Verify the mount points and desired capacity. Click DONE.

11.  Click Accept Changes.

12.  Click NETWORK & HOST NAME.

13.  Enter Host name, click Apply. Select Configure.

14.  Select IPv4 Settings, Enter IP Address, Netmask and Gateway. Click Save.

15.  Click OFF to turn ON the network adapter. Click Done.

16.  Click Begin Installation.

17.  Click ROOT PASSWORD.

18.  Enter Root Password. Click Done.

19.  Reboot when installation process completes.

Configure Data Drives on Name Node and Other Management Nodes

This section describes the steps needed to configure non-OS disk drives as RAID1 using the StorCli command. All drives are part of a single RAID1 volume. This volume can be used for staging any client data to be loaded to HDFS. This volume will not be used for HDFS data.

To configure data drive on the Name node and other management nodes, follow these steps:

1.     If the drive state shows up as JBOD, it can be converted into Unconfigured Good using Cisco UCSM or storcli64 command. The following steps should be performed if the state is JBOD.

2.     Get the enclosure id as follows:

 

 

3.     Convert to unconfigured good:

4.      Verify status by running the following command:

5.     Run this script as root user on rhel01 to rhel3 to create the virtual drives for the management nodes:

6.     Run the following command to get enclosure id:

7.     Run MegaCli script:

8.     Run the following command. State should change to Online:

9.     State can also be verified in UCSM as show below in Equipment>Rack-Mounts>Servers>Server # under Inventory/Storage/Disk tab:

Configure Data Drives on Data Nodes

To configure non-OS disk drives as individual RAID0 volumes using StorCli command, follow this step. These volumes will be used for HDFS Data.

1.     Issue the following command from the admin node to create the virtual drives with individual RAID 0 configurations on all the data nodes:

Cloudera Data Science Workbench (CDSW)

CDSW has prerequisites, one of which is CUDA. CUDA itself, also has prerequisites. The order of installation is:

1.     CUDA prerequisites

2.     CUDA

3.     CDSW prerequisites

4.     CDSW

Install Prerequisites for CUDA

 

List GPUs and CPUs installed:

lspci | grep -i nvidia

#lscpu

# ansible gpunodes -m shell -a "lscpu"

Verify Supported Version of Linux

Make sure Red Hat Enterprise Linux OS running is supported.

# ansible gpunodes -m shell -a "uname -m && cat /etc/*release"

GCC Installation

Make sure gcc is installed in the system:

# gcc –version

# ansible gpunodes -m shell -a "gcc --version"

Install Kernel Headers and Installation Packages

The CUDA Driver requires that the kernel headers and development packages for the running version of the kernel are installed at the time of the driver installation, as well as whenever the driver is rebuilt. For example, if your system is running kernel version 3.17.4-301, the 3.17.4-301 kernel headers and development packages must also be installed.

# ansible gpunodes -m shell -a "uname -r"

Run the following command to Install:

# ansible gpunodes -m  yum -a "name=kernel-devel-$(uname -r) state=present"

Install dkms

The NVIDIA driver RPM packages depend on other external packages, such as DKMS. Those packages are only available on third-party repositories, such as EPEL.

http://rpmfind.net/linux/rpm2html/search.php?query=dkms  for RHEL 7.x

RHEL 7.x http://rpmfind.net/linux/epel/7/x86_64/Packages/d/dkms-2.7.1-1.el7.noarch.rpm

Copy dkms rpm to all the GPU servers

Install dkms with yum install

Install NVIDIA GPU Drivers

To install the NVIDIA GPU drivers, follow these steps:

1.     Download this NVIDIA GPU driver from http://www.nvidia.com/Download/index.asp?lang=en-us

2.     For the NVIDIA driver download, select the product type, Series, Product, OS, and CUDA toolkit.

3.     Click SEARCH. The selected driver is shown below.

4.     Click Download. The download link can be captured by right-clicking AGREE & DOWNLOAD as shown below.

5.     Copy the .rpm file in all the GPU nodes as shown below.

6.     Install the driver by running the following command:

Install CUDA

To install CUDA, follow this step:

1.     Download CUDA 9.2

CUDA version 9.2 is available here:

https://developer.nvidia.com/cuda-92-download-archive?target_os=Linux&target_arch=x86_64&target_distro=RHEL&target_version=7&target_type=rpmlocal

Download and Setup NVIDIA CUDA Deep Neural Network Library (cuDNN)

Download cuDNN 7.2.1

Download cuDNN from https://developer.nvidia.com/cudnn  for the same CUDA version.

Copy it into all the GPU servers.

Post Installation Steps

1.     Add CUDA in PATH and LD_LIBRARY_PATH variable in all the GPU nodes.

2.     The PATH and LD_LIBRARY_PATH variable needs to include /usr/local/<cuda 8 or 9 path>/bin:

3.     Reboot the GPU nodes.

Verify Drivers

Restart the servers to verify the drivers have been installed:

# ansible rhel20.hdp3.cisco.local -m shell -a "nvidia-smi"

# ansible gpunodes -m shell -a "/usr/local/cuda-9.2/bin/nvcc --version"

# ansible gpunodes -m shell -a "cat /proc/driver/nvidia/version"

ansible gpunodes -m shell -a "export NVIDIA_DRIVER_VERSION=418.67"

Installation Prerequisites for CDSW

Follow the steps in this section on all CDSW nodes:

Set Up a Wildcard DNS Subdomain

Cloudera Data Science Workbench uses subdomains to provide isolation for user-generated HTML and JavaScript, and routing requests between services. To set up subdomains for Cloudera Data Science Workbench, configure your DNS server with an A record for a wildcard DNS name such as *.cdsw.<your_domain>.com for the master host, and a second A record for the root entry of cdsw.<your_domain>.com.

You can also use a wildcard CNAME record if it is supported by your DNS provider.

 

To set up a wildcard DNS subdomain, follow these steps:

1.     With dnsmasq, to add a wildcard DNS subdomain, complete the following steps only for the master host.

2.     Update /etc/dnsmasq.conf to enable Wildcard entry:

address=/cdsw/10.13.1.50

#service dnsmasq restart

3.     Test the working of wildcard DNS with dig or nslookup:

#nslookup *.cdsw.<your domain>.com

#dig cdsw.<your_domain>.com

#dig *.cdsw.<your domain>.com

Supported JDK Version

The entire CDH cluster, including Cloudera Data Science Workbench gateway nodes, must use Oracle JDK. OpenJDK is not supported by CDH, Cloudera Manager, or Cloudera Data Science Workbench.

Spark 2.2 which is needed by CDSW to run Spark jobs (on Hadoop nodes) requires JDK 1.8. On CSD-based deployments, Cloudera Manager automatically detects the path and version of Java installed on Cloudera Data Science Workbench gateway hosts.

To upgrade your entire CDH cluster to JDK 1.8, see Upgrading to Oracle JDK 1.8.

IP Tables and Security on CDSW Nodes                                                      

Disable all pre-existing iptables rules. While Kubernetes makes extensive use of iptables, it is difficult to predict how pre-existing iptables rules will interact with the rules inserted by Kubernetes. Therefore, Cloudera recommends you use the following commands to disable all pre-existing rules before you proceed with the installation.

 

Configure Block Devices

Docker Block Device

The Cloudera Data Science Workbench installer will format and mount Docker on each gateway host. Make sure there is no important data stored on these devices. Do not mount these block devices prior to installation.

Application Block Device or Mount Point

The master host on Cloudera Data Science Workbench requires at least 500 GB for database and project storage. This recommended capacity is contingent on the expected number of users and projects on the cluster. While large data files should be stored on HDFS, it is not uncommon to find gigabytes of data or libraries in individual projects. Running out of storage will cause the application to fail. Make sure you continue to carefully monitor disk space usage and I/O using Cloudera Manager.

Cloudera Data Science Workbench will store all application data at /var/lib/cdsw. In a CSD-based deployment, this location is not configurable. Cloudera Data Science Workbench will assume the system administrator has formatted and mounted one or more block devices to /var/lib/cdsw.

Regardless of the application data storage configuration you choose, /var/lib/cdsw must be stored on a separate block device (the RAID1 of SSDs created for this).

Download and Install CDSW with Cloudera Manager

To download and install CDSW, follow these steps:

1.     Download the Cloudera Data Science Workbench Custom Service Descriptor (CSD) CLOUDERA_DATA_SCIENCE_WORKBENCH-1.5.0.jar

 

 

2.     Restart cloudera-management service

3.     Log into the Cloudera Manager Admin Console and restart the Cloudera Management Service.

a.     Select Clusters > Cloudera Management Service.

b.     Select Actions > Restart.

c.     From the machine with internet access, download Cloudera CDSW parcel from https://archive.cloudera.com/cdsw1/1.5.0/parcels/

d.     Copy downloaded CDSW parcels into the server hosting Cloudera parcels copy all these files to CDSW folder and move it to /var/www/html/CDSW

e.     Log into the Cloudera Manager Admin Console.

4.     Click Configuration to configure the Cloudera Data Science Workbench parcel. Add URL to access CDSW local repository.

5.     Click Download to download parcel to all nodes in the cluster.

6.     Distribute the parcel to the hosts in your cluster, and then activate the parcel.

7.     Once distributed, click Activate.

8.     Click OK.

Add Apache Spark 2 Service

From CDH version 6 and later, Apache Spark 2 is packaged within CDH and can no longer be installed separately.

To find out which version of Spark 2 ships with your version of CDH 6, see the CDH 6 Packaging Information guide.

To add Apache Spark 2 Service, follow these steps:

1.     Select host(s) to add Spark gateway, click Continue.

2.     Select Dependencies.

3.     If Spark was not added part of the cluster creation process. Click Add Service and Assign Spark History Server and gateway roles as shown in the screenshot below and complete adding Spark service to cluster.

4.     Cloudera Manager runs the command to configure history server and Spark gateway role and start them on nodes assigned.

5.     Review the Summary. Click Finish.

6.     Select Deploy Client Configuration.

Add the Cloudera Data Science Workbench Service

To install CDSW, follow these steps:

1.     Log into the Cloudera Manager Admin Console.

2.     On the Home > Status tab, right-click the cluster name and select Add a Service to launch the wizard. A list of services will be displayed.

3.     Select the Cloudera Data Science Workbench service and click Continue.

4.      Select Dependencies. Click Continue.

5.     Assign the Master and Worker roles to the gateway hosts. You must assign the Cloudera Data Science Workbench Master role to one gateway host, and optionally, assign the Worker role to one or more gate-way hosts.

6.     Configure the following parameters and click Continue.

7.     The wizard will now begin a First Run of the Cloudera Data Science Workbench service. This includes deploying client configuration for HDFS, YARN and Spark 2, installing the package dependencies on all hosts, and formatting the Docker block device. The wizard will also assign the Application role to the host running Master, and the Docker Daemon role to all the Cloudera Data Science Workbench gateway hosts.

8.     Once the First Run command has completed successfully, click Finish to go back to the Cloudera Manager home page.

The Cloudera Manager Web Console reporting CDSW service configured on Cluster is shown below:

Create the Administrator Account

After your installation is complete, set up the initial administrator account. Go to the Cloudera Data Science Workbench web application or via CDSW WebUI from Cloudera Manager.

You must access Cloudera Data Science Workbench from the Cloudera Data Science Workbench Domain configured when setting up the service, and not the hostname of the master node. Visiting the hostname of the master node will result in a 404 error.

The first account that you create becomes the site administrator. You may now use this account to create a new project and start using the workbench to run data science workloads.

Non-Kerberized Clusters

 

1.     For a non-kerberized cluster, by default, your Hadoop username will be set to your Cloudera Data Science Workbench login username. Override this default and set an alternative HADOOP_USER_NAME (“admin” in this case which was the admin user created).

2.     Create User admin in hdfs:

 

3.     Go to the Settings on Cloudera Data Science WorkBench, Click Hadoop Authentication > Hadoop Username Override. Enter admin. Click Authenticate.

4.     Restart CDSW from Cloudera Manager.

Use GPUs for Cloudera Data Science Workbench Workloads

A GPU is a specialized processor that can be used to accelerate highly parallelized computationally intensive workloads. Because of their computational power, GPUs have been found to be particularly well-suited to deep learning workloads. Ideally, CPUs and GPUs should be used in tandem for data engineering and data science workloads. A typical machine learning workflow involves data preparation, model training, model scoring, and model fitting. You can use existing general-purpose CPUs for each stage of the workflow, and optionally accelerate the math-intensive steps with the selective application of special-purpose GPUs. For example, GPUs allow you to accelerate model fitting using frameworks such as TensorFlow, PyTorch, Keras, MXNet, and Microsoft Cognitive Toolkit (CNTK).

By enabling GPU support, data scientists can share GPU resources available on Cloudera Data Science Workbench nodes. Users can request a specific number of GPU instances, up to the total number available on a node, which are then allocated to the running session or job for the duration of the run. Projects can use isolated versions of libraries, and even different CUDA and cuDNN versions via Cloudera Data Science Workbench's extensible engine feature.

Enable GPU with CDSW

·       Cloudera Data Science Workbench only supports CUDA-enabled NVIDIA GPU cards.

·       Cloudera Data Science Workbench does not support heterogeneous GPU hardware in a single deployment.

·       Cloudera Data Science Workbench does not include an engine image that supports NVIDIA libraries. Create your own custom CUDA-capable engine image using the instructions described in this topic.

·       Cloudera Data Science Workbench does not install or configure the NVIDIA drivers on the Cloudera Data Science Workbench gateway hosts. These depend on your GPU hardware and will have to be installed by your system administrator. The steps provided in this topic are generic guidelines that will help you evaluate your setup.

·       The instructions described in this topic require Internet access. If you have an airgapped deployment, you will be required to manually download and load the resources onto your hosts.

·       For a list of known issues associated with this feature, refer Known Issues - GPU Support.

This section provides instructions about creating your own custom CUDA-capable engine image.

To enable Docker containers to use the GPUs, the previously installed NVIDIA driver libraries must be consolidated in a single directory named after the <driver_version> and mounted into the containers. This is done using the nvidia-docker package, which is a thin wrapper around the Docker CLI and a Docker plugin.

1.     Download and install nvidia-docker:

2.     Run a small container to create the Docker volume structure:

3.     Use the following Docker command to verify that Cloudera Data Science Workbench can access the GPU:

#vi nvidia-gpu-access.sh

mkdir /var/lib/nvidia-docker/

mkdir /var/lib/nvidia-docker/volumes/

mkdir /var/lib/nvidia-docker/volumes/nvidia_driver/

mkdir /var/lib/nvidia-docker/volumes/nvidia_driver/418.67

mkdir /var/lib/nvidia-docker/volumes/nvidia_driver/418.67/bin

mkdir /var/lib/nvidia-docker/volumes/nvidia_driver/418.67/lib64

cp /usr/bin/nvidia* /var/lib/nvidia-docker/volumes/nvidia_driver/418.67/bin

cp /usr/lib64/libcuda* /var/lib/nvidia-docker/volumes/nvidia_driver/418.67/lib64

cp /usr/lib64/libnvidia* /var/lib/nvidia-docker/volumes/nvidia_driver/418.67/lib64

# docker run --net host \

--device=/dev/nvidiactl \

--device=/dev/nvidia-uvm \

--device=/dev/nvidia0 \

-v /var/lib/nvidia-docker/volumes/nvidia_driver/418.67/:/usr/local/nvidia/ \

-it nvidia/cuda \

usr/local/nvidia/bin/nvidia-smi

Enable GPU Support in Cloudera Data Science Workbench

To enable GPU support for a CSD deployment, follow these steps:

1.     Go to the CDSW service in Cloudera Manager. Click Configuration.

2.     Search for the Enable GPU Support property and click the check box to enable it.

3.     Click Save Changes.

4.     Restart the CDSW service in Cloudera Manager.

Test Cloudera Data Science Workbench for GPUs

Once Cloudera Data Science Workbench has successfully restarted, if NVIDIA drivers have been installed on the Cloudera Data Science Workbench hosts, Cloudera Data Science Workbench will now be able to detect the GPUs available on its hosts.

Create a Custom CUDA-Capable Engine Image

The base engine image (docker.repository.cloudera.com/cdsw/engine:8) that ships with Cloudera Data Science Workbench will need to be extended with CUDA libraries to make it possible to use GPUs in jobs and sessions.

Designate a Server for Docker and Start the Registry

To designate a server for Docker and start the registry, follow these steps:

Designate a server in the cluster for the Docker registry. Minimal resources are required, but sufficient disk space is needed to store the images and metadata. Docker must be installed and running.

Optional: By default, data will only be persisted within the container. If you would like to persist the data on the host, you can customize the bind mounts using the -v option.

1.     Create /var/lib/registry folder:

# mkdir /var/lib/registry

2.     Configure Docker to allow pulling from this insecure registry. Modify /etc/docker/daemon.json on all nodes in the cluster to include the following configuration options:

# vi /etc/docker/daemon.json

{

 

    "live-restore" : true,

 

    "debug" : true,

 

    "insecure-registries" : ["rhel20.hdp3.cisco.local:5000"]

}

3.     Restart Docker on all nodes.

Run a Local Registry

To start the registry container, run the following command:

# docker run -d -p 5000:5000 --restart=always --name registry -v /mnt/registry:/var/lib/registry registry:2

Verify the registry container by running docker ps command:

Test the Docker Registry

To test the docker registry, follow the steps:

1.     Pull the docker image.

2.     Tag the image

3.     Push the image into private registry.

The following sample Dockerfile illustrates an engine on top of which machine learning frameworks such as TensorFlow and PyTorch can be used. This Dockerfile uses a deep learning library from NVIDIA called NVIDIA CUDA Deep Neural Network (cuDNN). Make sure you check with the machine learning framework that you intend to use in order to know which version of cuDNN is needed. As an example, TensorFlow 1.14.0 uses CUDA 10.0 and requires cuDNN 7.4.

To create the cuda.Dockerfile, run the following command:

 

docker images

1.     Tag the image as localhost:5000/<image name>. This creates an additional tag for the existing image. When the first part of the tag is a hostname and port, Docker interprets this as the location of a registry, when pushing.

 # docker tag rhel20.hdp3.cisco.local:5000/cdip-cuda:9.2 rhel20.hdp3.cisco.local:5000/cdip-cuda

2.     Push the image to the local registry running at localhost:5000:

# docker push rhel20.hdp3.cisco.local:5000/cdip-cuda

Allocate GPUs for Sessions and Jobs

Once Cloudera Data Science Workbench has been enabled to use GPUs, a site administrator must whitelist the CUDA-capable engine image created in the previous step. Site administrators can also set a limit on the maximum number of GPUs that can be allocated per session or job.

To allocate GPUs for sessions and jobs, follow these steps:

1.     Sign into Cloudera Data Science Workbench as a site administrator.

2.     Click Admin.

3.     Go to the Engines tab.

4.     From the Maximum GPUs per Session/Job drop-down list, select the maximum number of GPUs that can be used by an engine.

5.     Under Engine Images, add the custom CUDA-capable engine image created in the previous step. This whitelists the image and allows project administrators to use the engine in their jobs and sessions. Enter description and Tag created for repository. Click Add.

Project administrators can now whitelist the CUDA engine image to make it available for sessions and jobs within a project by following these steps:

1.     Go to Projects Overview.

2.     Click any existing project or create a new project and choose Settings.

3.     Go to the Engines tab.

4.     Under Engine Image, add the CUDA-capable engine image.

5.     Go to the Sessions tab and click an existing session or create a new session and allocate resources.

6.     Click Terminal Access.

Hardik Patel, Big Data Solutions Architect, Cisco Systems, Inc.

Hardik Patel is a Big Data Solutions Architect at Computing Systems Product Group. He is part of the solution engineering team focusing on big data infrastructure, solutions, and performance.

Acknowledgements

For their support and contribution to the design, validation, and creation of this Cisco Validated Design, the author would like to thank:

·       Karthik Kulkarni, Architect, Computing Systems Product Group, Cisco Systems, Inc.

·       Muhammad Afzal, Architect, Computing Systems Product Group, Cisco Systems, Inc.

NOTE: Available paragraph styles are listed in the Quick Styles Gallery in the Styles group on the Home tab. Alternatively, they can be accessed via the Styles window (press Alt + Ctrl + Shift + S).