Cisco Data Intelligence Platform with Hortonworks Data Platform 3.1 and Cloudera Data Science Workbench 1.5

Deployment Guide for the Cisco Data Intelligence Platform with Hortonworks Data Platform 3.1.0

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

Last Updated: October 18, 2019

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)

© 2019 Cisco Systems, Inc. All rights reserved.

 

Table of Contents

Executive Summary. 8

Solution Overview.. 9

Cisco Data Intelligence Platform... 9

Audience. 11

Purpose of this Document. 11

What’s New in this Release?. 11

What’s Next?. 11

Reference Architecture. 12

Data Lake Reference Architecture. 12

AI Computing Farm Reference Architecture. 13

Scaling the Solution. 16

Scaled Architecture with 3:1 Oversubscription with Cisco Fabric Interconnects and Cisco ACI 17

Scaled Architecture with 2:1 Oversubscription with Cisco ACI 18

Technology Overview.. 20

Cisco UCS Integrated Infrastructure for Big Data and Analytics. 20

Cisco Unified Computing System... 20

Cisco UCS 6300 Series Fabric Interconnects. 20

Cisco UCS C-Series Rack-Mount Servers. 20

Cisco UCS Virtual Interface Cards. 25

Cisco UCS Manager 25

NVIDIA GPU.. 26

NVIDIA CUDA.. 26

Cloudera Enterprise Data Hub and Hortonworks Data Platform... 26

Cloudera (CDH 6.2.0) 26

Cloudera Data Science Workbench. 27

Hortonworks Data Platform... 29

Apache Ambari 29

HDP for Data Access. 30

Submarine. 31

Docker Containerization. 32

YARN Support For Docker 32

NVIDIA Docker 33

GPU Pooling and Isolation. 34

Red Hat Ansible Automation. 34

Solution Design. 35

Requirements. 35

Rack and PDU Configuration. 35

Port Configuration on Fabric Interconnect. 35

Cabling for Cisco UCS C240 M5. 35

Software Distributions and Versions. 36

Hortonworks Data Platform (HDP 3.1.0) 36

Red Hat Enterprise Linux (RHEL) 36

Software Versions. 36

Fabric Configuration. 37

Perform Initial Setup of Cisco UCS 6332 Fabric Interconnects. 38

Configure Fabric Interconnect A.. 38

Configure Fabric Interconnect B.. 39

Log Into Cisco UCS Manager 39

Upgrade Cisco UCS Manager Software to Version 4.0(4b) 39

Add a Block of IP Addresses for KVM Access. 39

Enable Uplink Ports. 41

Configure VLANs. 42

Enable Server Ports. 44

Create Pools for Service Profile Templates. 45

Create an Organization. 45

Create MAC Address Pools. 46

Create a Server Pool 48

Create Policies for Service Profile Templates. 50

Create Host Firmware Package Policy. 50

Create QoS Policies. 51

Create the Local Disk Configuration Policy. 53

Create the Server BIOS Policy. 53

Create the Boot Policy. 56

Create the Power Control Policy. 58

Create the Service Profile Template. 62

Configure the Storage Provisioning for the Template. 64

Configure Network Settings for the Template. 64

Configure the vMedia Policy for the Template. 68

Configure the Server Boot Order for the Template. 69

Configure the Server Assignment for the Template. 70

Configure the Operational Policies for the Template. 71

Install Red Hat Enterprise Linux 7.6. 73

Post OS Install Configuration. 91

Configure /etc/hosts. 91

Set Up the Passwordless Login. 92

Create the Red Hat Enterprise Linux (RHEL) 7.6 Local Repository. 93

Create the Red Hat Repository Database. 94

Set Up Ansible. 95

Install httpd. 97

Set Up All Nodes to Use the RHEL Repository. 97

Upgrade the Cisco Network Driver for VIC1387. 98

Install xfsprogs. 98

Set Up JAVA.. 99

Configure NTP. 100

Enable Syslog. 102

Set ulimit. 102

Disable SELinux. 103

Set TCP Retries. 103

Disable the Linux Firewall 104

Disable Swapping. 104

Disable Transparent Huge Pages. 104

Disable IPv6 Defaults. 105

Configure Data Drives on Name Node and Other Management Nodes. 105

Configure Data Drives on Data Nodes. 108

Configure the Filesystem for NameNodes and Datanodes. 109

Cluster Verification. 110

Install HDP 3.1.0. 112

Prerequisites for HDP Installation. 113

Hortonworks Repository. 113

Downgrade Snappy on All Nodes. 115

HDP Installation. 115

Install and Setup Ambari Server on rhel1. 115

Setup Ambari Server On Admin Node(Rhel1) 121

Launch the Ambari Server 123

Create the Cluster 124

Select Version. 124

Select Hosts. 125

Hostname Pattern Expressions. 127

Confirm Hosts. 128

Choose Services. 128

Assign Masters. 129

Assign Slaves and Clients. 130

Customize Services. 131

HDFS. 135

MapReduce2. 137

YARN.. 137

Configure the HDFS NameNode High Availability. 138

HBase. 139

Zookeeper 139

Storm... 139

Ambari Metrics. 139

Accumulo. 140

Atlas. 140

Kafka. 141

Knox. 141

SmartSense. 141

Spark. 142

Review.. 142

Deploy. 143

Summary of the Installation Process. 144

High Availability for HDFS NameNode and YARN ResourceManager 145

Configure the HDFS NameNode High Availability. 145

Configure the YARN ResourceManger HA.. 153

Bill of Materials. 158

Appendix – A.. 164

Configure Data Drives on Name Node and Other Management Nodes. 164

Configure Data Drives on Data Nodes. 165

Cloudera Data Science Workbench (CDSW) 166

Install the Prerequisites for CUDA.. 167

Install GCC.. 168

Install Kernel Headers and Installation Packages. 169

Install DKMS. 170

Install NVIDIA GPU Drivers. 171

Install CUDA.. 172

Download and Setup NVIDIA CUDA Deep Neural Network Library (cuDNN) 174

Installation Prerequisites for CDSW... 175

Set Up a Wildcard DNS Subdomain. 175

Supported JDK Version. 176

IP Tables and Security on CDSW Nodes. 176

Configure Block Devices. 177

Download and Install CDSW with HDP 3.1.0. 177

Create HDFS User Directories. 178

Install Cloudera Data Science Workbench on the Master Host. 179

Install Cloudera Data Science Workbench on Worker Hosts. 179

Create the Administrator Account. 180

Non-Kerberized Clusters. 181

Use GPUs for Cloudera Data Science Workbench Workloads. 182

Enable GPU with CDSW... 182

Create a Custom CUDA-Capable Engine Image. 184

Configure CDSW to Run Docker Containers. 188

Set Up Docker Registry. 188

Create a Custom CUDA-capable Engine Image. 189

Allocate GPUs for Sessions and Jobs. 190

About the Authors. 193

Acknowledgements. 193

 

 

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 has to 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 Cisco Data Intelligence Platform.  Cisco Data Intelligence Platform is discussed in detail here.

This CVD implements Cisco Data Intelligence Platform on Cisco Unified Computing System (Cisco UCS) using Hortonworks Data Platform 3.1.

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.

Cisco has developed numerous industry leading Cisco Validated Designs (CVDs) 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 two 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 1  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 2  Solution Architecture

Audience

The intended audience of this document includes, but not limited to, sales engineers, field consultants, professional services, IT managers, partner engineering and customers who want to deploy Cisco Data Intelligence Platform using Hortonworks Data Platform (HDP 3.1.0) on Cisco UCS. You are assumed to have intermediate level of knowledge for Apache Hadoop and Cisco UCS based scale-out infrastructure.

Purpose of this Document

This document describes the architecture and step by step guidelines of deployment procedures for Cisco Data Intelligence Platform using Hortonworks Data Platform (HDP) 3.1.0 on Cisco UCS C240 M5.

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

·         Data Lake with Hortonworks Data Platform 3.1 or Cloudera Enterprise Data Hub 6.2

·         Kubernetes / AI farm with Cloudera Data Science Workbench

·         Hadoop Tiered storage with S3260

·         Distributed AI/ML with Apache Submarine

What’s New in this Release?

This CVD describes the deployment procedure for the following:

·         Data Lake with Cloudera Enterprise Data Hub 6.2.0 or Hortonworks Data Platform 3.1.0

·         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

·         Distributed Deep Learning with Submarine

What’s Next?

This CVD showcases Cisco UCS Manager (UCSM). This solution can also be deployed using Cisco Intersight. This along with other additional Cisco UCS features will be added to the appendix section in the following months. Some of these include,

·         Cisco Intersight

·         Cloudera Data Science Workbench

·         Tiered Storage with HDFS on Cisco UCS S3260

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

·         4^th Generation Fabric Interconnect

·         Hadoop data offload to S3 compliant storage

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

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

Table 3           Tiered Storage (Data Lake Reference Architecture)

 

Figure 3 illustrates a 12-node starter cluster with all the 3 components in a single rack. The top 8 nodes 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 3  Topology

As illustrated in Figure 4, 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 4  Cisco Data Intelligence Platform - 30 Node Configuration with Cloudera CDH 6.2 and CDSW 1.5

 

Figure 4 shows an alternate configuration for cases where more GPU capacity is needed. Four of the Cisco UCS C240 M5 servers from the previous configuraton in Figure 3 are replaced with Cisco UCS C480 M5 ML M5 server which support up to eight V100 MXM GPUs.

 

Scaling the Solution

Figure 5 illustrates how to scale the solution. Each pair of Cisco UCS 6332 Fabric Interconnects has 28 Cisco UCS C240 M5 servers connected to it. This allows for four 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 28 servers per FI, a total of 168 servers can be supported. Additionally, the can scale to thousands of nodes with the Nexus 9500 series family of switches.

Figure 5  Scaling the Solution

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

·         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 in this document and shown in Figure 6 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 32 x 40-Gbps throughput). From each fabric interconnect, 8 x 40-Gbps links connect to a pair of Cisco Nexus 9336 Switches. Two pairs of fabric interconnects can connect to a single pair of Cisco Nexus 9336 Switches (8 x 2 40-Gbps links). 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).

Figure 6  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 7, 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.

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 7  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 Hortonworks Data Platform on Cisco UCS Integrated Infrastructure for Big Data and Analytics, is 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 Unified Computing System

Cisco Unified Computing System (Cisco UCS) is a next-generation solution for blade and rack server computing. Cisco UCS integrates a low-latency; lossless 10 and 40 Gigabit Ethernet unified network fabric with enterprise-class, x86-architecture servers. Cisco UCS is an integrated, scalable, multi-chassis platform in which all resources participate in a unified management domain. Cisco UCS accelerates the delivery of new services simply, reliably, and securely through end-to-end provisioning and migration support for both virtualized and non-virtualized systems. Cisco UCS fuses access layer networking and servers. This high-performance, next-generation server system provides a data center with a high degree of workload agility and scalability.

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 (UCSM). 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 8  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 9) 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 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 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 2^nd Gen 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 9  Cisco UCS C240 M5 Rack-Mount Server – Front View

Figure 10           Cisco UCS C240 M5 Rack-Mount 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 11) 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 Intel Xeon Scalable processors with up to 28 cores per socket and support for two-or four-processor configurations

·         2933-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 11           Cisco UCS C480 M5 Rack-Mount Server – Front View

Figure 12           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

·         Latest Intel Xeon Scalable processors with up to 28 cores per socket and support for two processor configurations

·         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 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 13           Cisco UCS C480 ML M5 Purpose Built Deep Learning Server – Front View

Figure 14           Cisco UCS C480 ML M5 Purpose Built Deep Learning Server – Rear View

For more information about Cisco UCS C480 ML M5 Server, go to: 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          Feature and Benefits for Cisco UCS C480 ML M5 Server

Cisco UCS Virtual Interface Cards

Cisco UCS Virtual Interface Cards (VICs) 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 15           Cisco UCS VIC 1387

For more information about Cisco UCS Adapters, go to: https://www.cisco.com/c/en/us/products/interfaces-modules/unified-computing-system-adapters/index.html

Cisco UCS Manager

Cisco UCS Manager (UCSM) resides within the Cisco UCS 6300 Series Fabric Interconnect. It makes the system self-aware and self-integrating, managing all of the system components as a single logical entity. Cisco UCS Manager can be accessed through an intuitive GUI, a CLI, or an XML 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.

For more information about Cisco UCS Manger, go to: https://www.cisco.com/c/en/us/products/servers-unified-computing/ucs-manager/index.html

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

This CVD can be implemented with Cloudera Enterprise Data Hub and also with Hortonworks Data Platform

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 of the existing investments in technology and skills.

Figure 16           Cloudera Data Hub

Cloudera provides a scalable, flexible, integrated platform that makes it easy to manage rapidly increasing volumes and varieties of data in any enterprise. Industry-leading Cloudera products and solutions enable to deploy and manage Apache Hadoop and related projects, manipulate and analyze data, and keep that data secure and protected.

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 of the ML/DL frameworks such as TensorFlow, Theano, PyTorch, and so on. Additionally, 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.

Additionally, 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.

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 17). 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 17           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.

The following 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 18, 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 take action on data as it flows into HDFS.

Figure 18           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. And 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 19           Submarine Workflow

Docker Containerization

Hortonworks Data Platform (HDP 3.0) makes use of container technology. Containers are conceptually similar to 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 that performs operating-system-level virtualization, also known as "containerization." 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 has to 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 particular 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 20          Containerized Application on Apache Hadoop YARN 3.1

As shown in Figure 20, YARN Services Framework in addition with Docker containerization, it is now possible to run both existing Hadoop frameworks, such as Hive, Spark, etc., 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.

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 the architecture and deployment procedures for Hortonworks Data Platform (HDP) 3.1.0 on a 31 node cluster based on Cisco UCS Integrated Infrastructure for Big Data and Analytics. The solution goes into detail configuring HDP 3.1.0 on the Cisco UCS Integrated infrastructure for Big Data. In addition, it also details the configuration for Hortonworks Dataflow for various use cases.

The cluster configuration consists of the following:

·         2 Cisco UCS 6332UP Fabric Interconnects

·         22 Cisco UCS C240 M5 Rack-Mount servers

·         8 Cisco UCS C3260 M5 Storage Server

·         12 NVIDIA T4 GPUs

·         2 Cisco R42610 standard racks

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

Rack and PDU Configuration

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 Chassis connected to each of the vertical PDUs for redundancy; thereby, ensuring availability during power source failure, similar to the first rack.

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

Cabling for Cisco UCS C240 M5

The Cisco UCS C240 M5 rack server is equipped with 2 x Intel Xeon Processor Scalable Family 6132 (2 x 14 cores, 2.6 GHz), 192 GB of memory, Cisco UCS Virtual Interface Card 1387 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 21 illustrates the port connectivity between the Fabric Interconnect, and Cisco UCS C240 M5 server. Sixteen Cisco UCS C240 M5 servers are used in Master rack configurations.

Figure 21           Cisco UCS C240 M5 and 6300 Series Fabric Interconnect Port Connectivity

For information about physical connectivity and single-wire management, go to:

https://www.cisco.com/c/en/us/td/docs/unified_computing/ucs/c-series_integration/ucsm3-2/b_C-Series-Integration_UCSM3-2/b_C-Series-Integration_UCSM3-2_chapter_010.html?bookSearch=true

For more information about physical connectivity illustrations and cluster setup, go to:

https://www.cisco.com/c/en/us/td/docs/unified_computing/ucs/c-series_integration/ucsm3-2/b_C-Series-Integration_UCSM3-2/b_C-Series-Integration_UCSM3-2_chapter_010.html?bookSearch=true

Software Distributions and Versions

The software distributions required versions are listed below.

Hortonworks Data Platform (HDP 3.1.0)

The Hortonworks Data Platform supported is HDP 3.1.0. For more information, go to: http://www.hortonworks.com.

Red Hat Enterprise Linux (RHEL)

The operating system supported is Red Hat Enterprise Linux 7.6. For more information, go to: http://www.redhat.com.

Software Versions

The software versions tested and validated in this document are shown in Table 6.

Table 6          Software Versions

 

 

Fabric Configuration

This section provides the details to configure a fully redundant, highly available Cisco UCS 6332 fabric configuration. The following is the high-level workflow to setup Cisco UCS:

·         Initial setup of the Fabric Interconnect A and B

·         Connect to Cisco UCS Manager using virtual IP address of using the web browser

·         Launch Cisco UCS Manager

·         Enable server and uplink ports

·         Start discovery process

·         Create pools and polices for service profile template

·         Create Service Profile template

·         Create service profile for each server from service profile template

·         Associate Service Profiles to servers5

Perform Initial Setup of Cisco UCS 6332 Fabric Interconnects

This section describes the initial setup of the Cisco UCS 6332 Fabric Interconnects A and B.

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/3-2/b_UCSM_Getting_Started_Guide_3_2/b_UCSM_Getting_Started_Guide_3_2_chapter_0100.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(4b)

This document assumes the use of UCS 4.0(4b). Refer to the Cisco UCS 4.0 Release  (upgrade Cisco UCS Manager software and Cisco UCS 6332 Fabric Interconnect software to version 4.0(4b). Also, make sure the Cisco UCS C-Series version 4.0(4b) software bundles are installed on the Fabric Interconnects.

Add a Block of IP Addresses for KVM Access

To create a block of KVM IP addresses for server access in the Cisco UCS environment, follow these steps:

1.        Select the LAN tab at the top of the left window.

2.       Select Pools > root > IpPools > Ip Pool ext-mgmt.

3.        Right-click IP Pool ext-mgmt.

4.       Select Create Block of IPv4 Addresses.

Figure 22          Adding a Block of IPv4 Addresses for KVM Access Part 1

5.        Enter the starting IP address of the block and number of IPs needed, as well as the subnet and gateway information.

Figure 23           Adding Block of IPv4 Addresses for KVM Access Part 2

6.       Click OK to create the IP block.

7.        Click OK in the message box.

Enable Uplink Ports

To enable uplinks ports, follow these steps:

1.        Select the Equipment tab on the top left of the window.

2.       Select Equipment > Fabric Interconnects > Fabric Interconnect A (primary) > Fixed Module.

3.        Expand the Unconfigured Ethernet Ports section.

4.       Select port 29-32 that is connected to the uplink switch, right-click, then select Reconfigure > Configure as Uplink Port.

5.        Select Show Interface and select 40GB for Uplink Connection.

6.       A pop-up window appears to confirm your selection. Click Yes then OK to continue.

7.        Select Equipment > Fabric Interconnects > Fabric Interconnect B (subordinate) > Fixed Module.

8.       Expand the Unconfigured Ethernet Ports section.

9.       Select port number 29-32, which is connected to the uplink switch, right-click, then select Reconfigure > Configure as Uplink Port.

10.    Select Show Interface and select 40GB for Uplink Connection.

11.     A pop-up window appears to confirm your selection. Click Yes then OK to continue.

Figure 24          Enabling Uplink Ports Part1

Figure 25           Enabling Uplink Ports Part2

Figure 26          Enabling Uplink Ports Part 3

Configure VLANs

VLANs are configured as in shown in Table 7.

Table 7           VLAN Configurations

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.

To configure VLANs in the Cisco UCS Manager GUI, follow these steps:

1.        Select the LAN tab in the left pane in the UCSM GUI.

2.       Select LAN > LAN Cloud > VLANs.

3.        Right-click the VLANs under the root organization.

4.       Select Create VLANs to create the VLAN.

Figure 27           Creating a VLAN

5.        Enter vlan13 for the VLAN Name.

6.       Keep multicast policy as <not set>.

7.        Select Common/Global for vlan16.

8.       Enter 13 in the VLAN IDs field for the Create VLAN IDs.

9.       Click OK and then, click Finish.

10.    Click OK in the success message box.

Figure 28          Creating VLAN for Data

11.     Click OK and then click Finish.

Enable Server Ports

To enable server ports, follow these steps:

1.        Select the Equipment tab on the top left of the window.

2.       Select Equipment > Fabric Interconnects > Fabric Interconnect A (primary) > Fixed Module.

3.        Expand the Unconfigured Ethernet Ports section.

4.       Select all the ports that are connected to the Servers right-click them and select Reconfigure > Configure as a Server Port.

5.        A pop-up window appears to confirm your selection. Click Yes then OK to continue.

6.       Select Equipment > Fabric Interconnects > Fabric Interconnect B (subordinate) > Fixed Module.

7.        Expand the Unconfigured Ethernet Ports section.

8.       Select all the ports that are connected to the Servers right-click them and select Reconfigure > Configure as a Server Port.

9.       A pop-up window appears to confirm your selection. Click Yes, then OK to continue.

Figure 29          Enabling Server Ports

After the Server Discovery, Port 29-32 will be a Network Port and 1-28 will be Server Ports.

Figure 30           Ports Status after the Server Discover

Create Pools for Service Profile Templates

Create an Organization            

Organizations are used as a means to arrange and restrict access to various groups within the IT organization, thereby enabling multi-tenancy of the compute resources. This document does not assume the use of Organizations; however, the necessary steps are provided for future reference.

To configure an organization within the Cisco UCS Manager GUI, follow these steps:

1.        Click Quick Action icon on the top right corner in the right pane in the Cisco UCS Manager GUI.

2.       Select Create Organization from the options

3.        Enter a name for the organization.

4.       (Optional) Enter a description for the organization.

5.        Click OK.

6.       Click OK in the success message box.

Create MAC Address Pools

To create MAC address pools, follow these steps:

1.        Select the LAN tab on the left of the window.

2.       Select Pools > root > MAC Pools

3.        Right-click MAC Pools under the root organization.

4.       Select Create MAC Pool to create the MAC address pool. Enter ucs for the name of the MAC pool.

5.        (Optional) Enter a description of the MAC pool.

6.       Select Assignment Order Sequential.

7.        Click Next.

8.       Click Add.

9.       Specify a starting MAC address.

10.    Specify a size of the MAC address pool, which is sufficient to support the available server resources.

11.     Click OK.

Figure 31           Specifying first MAC Address and Size

12.    Click Finish.

13.     When the message box displays, click OK.

Create a Server Pool

A server pool contains a set of servers. These servers typically share the same characteristics. Those characteristics can be their location in the chassis, or an attribute such as server type, amount of memory, local storage, type of CPU, or local drive configuration. You can manually assign a server to a server pool or use server pool policies and server pool policy qualifications to automate the assignment.

To configure the server pool 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 Pools > root.

3.        Right-click the Server Pools.

4.       Select Create Server Pool.

5.        Enter your required name (ucs) for the Server Pool in the name text box.

6.       (Optional) enter a description for the organization.

7.        Click Next > to add the servers.

8.       Select all the Cisco UCS C240M5 servers to be added to the server pool that was previously created (ucs), then Click >> to add them to the pool.

9.       Click Finish.

10.    Click OK and then click Finish.

Create Policies for Service Profile Templates

Create Host Firmware Package Policy

Firmware management policies allow the administrator to select the corresponding packages for a given server configuration. These include adapters, BIOS, board controllers, FC adapters, HBA options, and storage controller properties as applicable.

To create a firmware management policy for a given server configuration using 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 Host Firmware Packages.

4.       Select Create Host Firmware Package.

5.        Enter the required Host Firmware package name (ucs).

6.       Select Simple radio button to configure the Host Firmware package.

7.        Select the appropriate Rack package that has been installed.

8.       Click OK to complete creating the management firmware package.

9.       Click OK.

Create QoS Policies

To create the QoS policy for a given server configuration using the Cisco UCS Manager GUI, follow these steps:

Platinum Policy

To create the Platinum policy, follow these steps:

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

2.       Select Policies > root.

3.        Right-click QoS Policies.

4.       Select Create QoS Policy.

 

5.        Enter Platinum as the name of the policy.

6.       Select Platinum from the drop-down list.

7.        Keep the Burst(Bytes) field set to default (10240).

8.       Keep the Rate(Kbps) field set to default (line-rate).

9.       Keep Host Control radio button set to default (none).

10.    When the pop-up window appears, click OK to complete the creation of the Policy.

Set Jumbo Frames

To set Jumbo frames and enable QoS, follow these steps:

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

2.       Select LAN Cloud > QoS System Class.

3.        In the right pane, select the General tab

4.       In the Platinum row, enter 9216 for MTU.

5.        Check the Enabled Check box next to Platinum.

6.       In the Best Effort row, select none for weight.

7.        In the Fiber Channel row, select none for weight.

8.       Click Save Changes.

9.       Click OK.

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.       Go to Policies > root.

3.        Right-click Local Disk Config Policies.

4.       Select Create Local Disk Configuration Policy.

5.        Enter ucs as the local disk configuration policy name.

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

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

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

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

10.    Click OK.

 

Create the Server BIOS Policy

The BIOS policy feature in Cisco UCS automates the BIOS configuration process. The traditional method of setting the BIOS is manually and is often error-prone. By creating a BIOS policy and assigning the policy to a server or group of servers, can enable transparency within the BIOS settings configuration.

To create a server BIOS policy using the Cisco UCS Manager GUI, follow these steps:

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

2.       Select Policies > root.

3.        Right-click BIOS Policies.

4.       Select Create BIOS Policy.

5.        Enter your preferred BIOS policy name (ucs).

6.       Change the BIOS settings as shown in the following figures.

7.        Only changes that need to be made are in the Processor and RAS Memory settings.

Create the 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 as 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.

Create the Power Control Policy

To create Power Control 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 Power Control Policies.

4.       Select Create Power Control Policy.

 

5.        Enter ucs as the Power Control policy name.

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

7.        Select Performance for Fan Speed Policy.

8.       Select No cap for Power Capping selection.

9.       Click OK to create the Power Control Policy.

10.    Click OK.

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 C240M5-BIOS as the BIOS policy name.

Figure 32           BIOS Configuration

Create the Service Profile Template

To create the Service Profile Template, follow these steps:

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

2.       Right-click Service Profile Templates.

3.        Select Create Service Profile Template.

The Create Service Profile Template window appears.

To identify the service profile template, follow these steps:

1.        Name the service profile template as ucs. Select the Updating Template radio button.

2.       In the UUID section, select Hardware Default as the UUID pool.

3.        Click Next to continue to the next section.

 

Configure the Storage Provisioning for the Template

To configure storage policies, follow these steps:

1.        Go to the Local Disk Configuration Policy tab and select ucs for the Local Storage.

2.       Click Next to continue to the next section.

 

3.        Click Next once the Networking window appears to go to the next section.

Configure Network Settings for the Template

To configure the network settings for the templates, follow these steps:

1.        Keep the Dynamic vNIC Connection Policy field at the default.

2.       Select Expert radio button for the option how would you like to configure LAN connectivity?

3.        Click Add to add a vNIC to the template.

4.       The Create vNIC window displays. Name the vNIC as eth0.

5.        Select ucs in the Mac Address Assignment pool.

6.       Select the Fabric A radio button and check the Enable failover check box for the Fabric ID.

7.        Check the VLAN13 check box for VLANs and select the Native VLAN radio button.

8.       Select MTU size as 9000.

9.       Select adapter policy as Linux.

10.    Select QoS Policy as Platinum.

11.     Keep the Network Control Policy as Default.

12.    Click OK.

13.     Click Next to continue with SAN Connectivity.

14.    Select no vHBAs for How would you like to configure SAN Connectivity?

15.     Click Next to continue with Zoning.

16.    Click Next to continue with vNIC/vHBA placement.

17.     Click Next to configure vMedia Policy.

Configure the vMedia Policy for the Template

To configure the vMedia policy for the template, follow these steps:

1.        Click Next once the vMedia Policy window appears to go to the next section.

Configure the Server Boot Order for the Template

To set the boot order for the servers, follow these steps:

1.        Select ucs in the Boot Policy name field.

2.       Review to make sure that all of the boot devices were created and identified.

3.        Verify that the boot devices are in the correct boot sequence.

4.       Click OK.

5.        Click Next to continue to the next section.

6.       In the Maintenance Policy window, apply the maintenance policy.

7.        Keep the Maintenance policy at no policy used by default. Click Next to continue to the next section.

Configure the Server Assignment for the Template

To assign the servers to the pool, In the Server Assignment window, follow these steps:

1.        Select ucs for the Pool Assignment field.

2.       Select the power state to be Up.

3.        Keep the Server Pool Qualification field set to <not set>.

4.       Check the Restrict Migration check box.

5.        Select ucs in Host Firmware Package.

 

Configure the Operational Policies for the Template

To configure the operational policies for the template, in the Operational Policies Window, follow these steps:

1.        Select ucs in the BIOS Policy field.

2.       Select ucs in the Power Control Policy field.

3.        Click Finish to create the Service Profile template.

4.       Click OK in the pop-up window to proceed.

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

6.       Go to Service Profile Templates > root.

7.        Right-click Service Profile Templates ucs.

8.       Select Create Service Profiles From Template.

 

The Create Service Profiles from Template window appears.

Install Red Hat Enterprise Linux 7.6

This section provides detailed procedures to install Red Hat Enterprise Linux 7.6 using Software RAID (OS based Mirroring) on Cisco UCS C240 M5 servers. There are multiple ways to install the Red Hat Linux operating system. The installation procedure described in this deployment guide uses the KVM console and virtual media from Cisco UCS Manager.

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

1.        Log into the Cisco UCS 6332 Fabric Interconnect and launch the Cisco UCS Manager application.

2.       Select the Equipment tab.

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

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

5.        Click OK on the KVM Console – Select IP address pop-up window.

6.       Click the link to launch the KVM console.

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

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

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

10.    Select Choose File in the Virtual Disk Management windows.

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

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

13.     Click Map Drive after selecting the .iso file.

14.    In the KVM window, select the KVM tab to monitor during boot.

15.     In the KVM window, select the Macros > Static Macros > Ctrl-Alt-Del button in the upper left corner.

16.    Click OK.

17.     Click OK to reboot the system.

18.    Press F6 key on the keyboard to select install media.

19.    On reboot, the machine detects the presence of the Red Hat Enterprise Linux Server 7.6 install media.

20.    Select the Install Red Hat Enterprise Linux 7.6.

21.    Skip the Media test and start the installation. Select language of installation and click Continue.

22.    Select Date and time, which pops up another window as shown below:

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

24.    Click INSTALLATION DESTINATION.

25.    This opens a new window with the boot disks. Make the selection and choose I will configure partitioning. Click Done.

26.    This opens the new window for creating the partitions. Click the + sign to add a new partition as shown below, boot partition of size 2048 MB.

27.    Click Add Mount Point to add the partition.

28.    Change the Device type to RAID and make sure the RAID Level is RAID1 (Redundancy) and click Update Settings to save the changes.

29.    Click the + sign to create the swap partition of size 2048 MB as shown below.

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

31.     Click + to add the / partition. The size can be left empty, so it uses the remaining capacity and click Add Mount Point.

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

33.     Click Done to return to the main screen and continue the Installation.

34.    Click SOFTWARE SELECTION.

35.     Select Infrastructure Server and select the Add-Ons as noted below. Click Done.

36.    Click NETWORK & HOSTNAME and configure Hostname and Networking for the Host.

37.     Type in the hostname as shown below.

38.    Click Configure to open the Network Connectivity window. Click IPV4Settings.

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

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

41.    Click Begin Installation on the Main menu.

42.    Select Root Password in the User Settings.

43.    Enter the Root Password and click Done.

44.    When the installation is complete reboot the system.

45.    Repeat the steps to install Red Hat Enterprise Linux 7.6 on the remaining servers.

The hostnames and their corresponding IP addresses are shown in Table 8.

Table 8          Hostnames and IP Addresses

 

Post OS Install Configuration

Choose one of the nodes of the cluster or a separate node as the Admin Node for management such as HDP installation, Ansible, creating a local Red Hat repo and so on. In this document, rhel1 has been used for this purpose.

Configure /etc/hosts

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.        Log into the Admin Node (rhel1).

2.       Populate the host file with IP addresses and corresponding hostnames on the Admin node (rhel1) and other nodes as follows:

3.        On Admin Node (rhel1):

Set Up the Passwordless Login

To manage all of the cluster nodes 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 a passwordless 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 (rhel1).

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

Figure 33           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-host’s  .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 the Red Hat Enterprise Linux (RHEL) 7.6 Local Repository

To create the repository using RHEL DVD or ISO on the admin node (in this deployment rhel1 is used for this purpose), create a directory with all the required RPMs, run the createrepo command and then publish the resulting repository.

1.        Log into rhel1. Create a directory that would contain the repository.

2.       Alternatively, if you have access to a Red Hat ISO Image, Copy the ISO file to rhel1.

3.        On rhel1 create a .repo file to enable the use of the yum command:

4.       Copy rheliso.repo file from /var/www/html/rhelrepo  to /etc/yum.repos.d on rhel1:

Create the Red Hat Repository Database

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

1.        Install the createrepo package on admin node (rhel1). 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 34           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-2.4.6.0-1.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 34 shows the outcome of the ping command:

Figure 35           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:

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/4.0.2

1.        In the ISO image, the required driver kmod-enic-…..rpm can be located at  Network\Cisco\VIC\RHEL\RHEL7.6.

2.       From a node connected to the Internet, download, extract and transfer kmod-enic-.rpm to rhel1 (admin node).

3.        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 rhel1:

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

5.        It is recommended to download the kmod-megaraid driver for higher performance. The RPM can be found in the same package at: \Linux\Storage\LSI\Cisco_Storage_12G_SAS_RAID_controller\RHEL\RHEL7.6

Install xfsprogs

From the admin node rhel1 run the command shown below to Install  xfsprogs on all the nodes for xfs filesystem:

Set Up JAVA

To setup JAVA, follow these steps:

1.        Download jdk-7u76-linux-x64.rpm and src the rpm to admin node (rhel1) from the link  (http://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 (rhel1):

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:

 

Configure NTP

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 (rhel1). 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 (rhel1):

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:

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.

One of the following commands should suffice to confirm that the service is properly configured:

Set ulimit

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 rhel1 and add the following lines:

2.       Copy the /etc/security/limits.conf file from admin node (rhel1) 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 must be disabled during the install procedure and cluster setup. SELinux can be enabled after installation and while the cluster is running.

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:

The command (above) may fail if SELinux is already disabled. This require reboot to take effect.

2.       q 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 rhel1 and add the following lines:

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

Disable the Linux Firewall

The default Linux firewall settings are far too restrictive for any Hadoop deployment. Since the UCS Big Data deployment will be in its own isolated network there is no need for that additional firewall.

Disable Swapping

To disable swapping, follow these steps:

1.        In order to reduce Swapping, 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 Transparent Huge Pages

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

To disable Transparent Huge Pages, follow these steps:

1.        The following commands must be run for every reboot, so 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:

Disable IPv6 Defaults

To disable IPv6 defaults, follow these steps:

1.        Disable IPv6 as the addresses used are IPv4:

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

Configure Data Drives on Name Node and Other Management Nodes

This section describes the steps 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 drives on Name node and other nodes, follow these steps:

1.        From the website download storcli Error! Hyperlink reference not valid.https://www.broadcom.com/support/download-search/?pg=&pf=&pn=&po=&pa=&dk=storcli.

2.       Extract the zip file and copy storcli-1.14.12-1.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:

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

9.       Get the enclosure id as follows:

10.    Convert to unconfigured good:

11.      Verify status by running the following command:

12.    Run this script as root user on rhel1 to rhel3 to create the virtual drives for the management nodes:

13.     Run the following command to get enclosure id:

14.    Run MegaCli script:

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

16.    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 these steps. 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:

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.        From the Admin node, 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:

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

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

Delete Partitions

To delete a partition, follow these steps:

Cluster Verification

To verify a cluster, follow these steps:

1.        Create the script cluster_verification.sh as shown, on the Admin node (rhel1):

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 HDP 3.1.0

HDP is an enterprise grade, hardened Hadoop distribution. HDP combines Apache Hadoop and its related projects into a single tested and certified package. HPD 3.1.0 components are depicted in below. This section details how to install HDP 3.1.0 on the cluster.

Prerequisites for HDP Installation

This section details the prerequisites for the HDP installation, such as setting up HDP repositories.

Hortonworks Repository

1.        From a host connected to the Internet, create a Hortonworks folder and download the Hortonworks repositories as shown below, then transfer it to the admin node:

2.       Download Hortonworks HDP repo:

3.        Download Hortonworks HDP-Utils repo:

4.       Download HDP-GPL repo:

5.        Download the Ambari repo:

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

7.        Extract the tar ball:

8.       Create HDP repo with the following contents:

9.       Create the Ambari repo:

Downgrade Snappy on All Nodes

Downgrade snappy on all data nodes by running this command from admin node:

HDP Installation

To install HDP, complete the following the steps:

Install and Setup Ambari Server on rhel1

1.        Run the following command in rhel1 to install ambary-server:

2.       Log into rhel2.

3.        Install Red Hat Package Manager (RPM) according to the requirements of your operating system:

4.       Install PostgreSQL version 9.5 or later:

5.        Initialize the database as shown in the below figure by running the following command:

6.       Start PostgreSQL:

7.        Open /var/lib/pgsql/9.6/data/postgresql.conf and update to the following:

8.       Update these files on rhel2 in the location chosen to install the databases for Hive, Oozie and Ambari, using the host ip addresses:

9.       Restart PostgreSQL:

10.    Run the following:

Create Database for Ambari

To create the database for Ambari, follow these steps:

1.        Run the following commands mentioned below in bold to create and prepare database for Ambari:

2.       Restart PostgreSQL:

3.        Verify the ambari user by logging into psql:

4.       Load the Ambari Server database schema:

5.        Find the Ambari-DDL-Postgres-CREATE.sql file in the /var/lib/ambari-server/resources/ directory of the Ambari Server host after you have installed Ambari Server.

6.       Run the following command to launch the Ambari-DDL-Postgres-CREATE.sql script:

7.        Check the table is created by running \dt command:

8.       Restart PostgreSQL:

Create Database for Hive

Run the following command as shown in bold to create and prepare database for Hive:

Create Database for Oozie

Run the following command to create and prepare database for Oozie:

Setup Ambari Server On Admin Node(Rhel1)

To setup the Ambari server, follow these steps:

1.        Install the PostgreSQL JDBC driver:

2.       Configure Ambari server to use the JDBC driver for connectivity to Ambari database in PostgreSQL:

3.        Setup Ambari Server by running the following command:

4.       Start the Ambari Server:

 

5.        To check status of Ambari Server, run the following command:

 

Launch the Ambari Server

When the Ambari service starts, access the Ambari Install Wizard through the browser. To launch the Ambari server, follow these steps:

1.        Point the browser to http://<ip address for rhel1>:8080 or http://rhel1.hdp3.cisco.com:8080

The Ambari Login screen opens.

2.       Log into the Ambari Server using the default username/password: admin/admin. This can be changed at a later period of time.

3.        When logged in the “Welcome to Apache Ambari” window opens.

Create the Cluster

To create the cluster, follow these steps:

1.        To create a cluster, click “LAUNCH INSTALL WIZARD.”

2.       From the Get started page type “Cisco_HDP” for the name for the cluster.

3.        Click Next.

Select Version

To select the version, follow these steps:

1.        In the Select Version section, choose the HDP 3.1.0. version.

2.       Under Repositories, select “Use Local Repository.”

3.        Update the Redhat 7 HDP-3.0 URL to http://rhel1.hdp3.cisco.com/Hortonworks/HDP/centos7/3.1.0.0-187/

4.       Update the Redhat 7 HDP-3.0-GPL URL to http://rhel1.hdp3.cisco.com/Hortonworks/HDP-GPL/centos7/3.1.0.0-187/

5.        Update the Redhat 7 HDP-UTILS-1.1.0.22 to http://rhel1.hdp3.cisco.com/Hortonworks/HDP-UTILS/centos7/1.1.0.22/

Select Hosts

To build up the cluster, you need to provide the general information about how you want to set up the cluster. This requires providing the Fully Qualified Domain Name (FQDN) of each of the hosts. You also need to provide access to the private key file that was created in Set Up Password-less SSH; this is used to locate all the hosts in the system and to access and interact with them securely.

To select hosts, follow these steps:

1.        Use the Target Hosts text box to enter the list of host names, one per line. Ranges inside brackets can also be used to indicate larger sets of hosts.

2.       Select the option Provide your SSH Private Key in the Ambari cluster install wizard.

3.        Copy the contents of the file /root/.ssh/id_rsa on rhel1 and paste it in the text area provided by the Ambari cluster install wizard.

4.       Click the Register and Confirm to continue.

Figure 36           Install Options

Hostname Pattern Expressions

1.        Click OK on the Host Name Pattern Expressions popup.

Confirm Hosts

Confirm Hosts helps ensure that Ambari has located the correct hosts for the cluster and checks those hosts to make sure they have the correct directories, packages, and processes to continue the install.

To confirm host, follow these steps:

1.        If any host was selected in error, remove it by selecting the appropriate checkboxes and clicking the grey Remove Selected button.

2.       To remove a single host, click the small white Remove button in the Action column.

3.        When the list of hosts is confirmed, click Next.

Choose Services

HDP is made up of a number of components. Go to Hortonworks Understand the Basics for more information.

To choose services, follow these steps:

1.        Select all to preselect all items.

2.       When you have made your selections, click Next.

Assign Masters

The Ambari installation wizard attempts to assign the master nodes for various services that have been selected to appropriate hosts in the cluster, as listed in Table 9. The right column shows the current service assignments by host, with the hostname and its number of CPU cores and amount of RAM indicated.

1.        Reconfigure the service assignments to match Table 9.

Table 9          Reconfigure the Service Assignments

2.       Click Next.

Assign Slaves and Clients

The Ambari install wizard attempts to assign the slave components (DataNodes, NFSGateway, NodeManager, RegionServers, Supervisor, and Client) to appropriate hosts in the cluster.

To assign slaves and clients, follow these steps:

1.        Reconfigure the service assignment to match the values shown in Table 10.

2.       Assign DataNode, NodeManager, RegionServer, and Supervisor on nodes rhel3- rhel28.

3.        Assign Client to all nodes.

4.       Click Next.

Table 10       Services and Hostnames

 

Customize Services

This section shows the tabs that manage configuration settings for Hadoop components. The wizard attempts to set reasonable defaults for each of the options here, but this can be modified to meet specific requirements. The following sections provide configuration guidance that should be refined to meet specific use case requirements.

The following changes need to be made:

·         Memory and service level settings for each component and service level tuning.

·         Customize the log locations of all the components to make sure growing logs do not cause the SSDs to run out of space.

Credentials

Specify the credentials as per your organizational policies for services in CREDENTIALS tab as shown below:

Databases

In the DATABASES tab, to configure the database for DRUID, HIVE, OOZIE, and RANGER, follow these steps:

1.        Configure DRUID as shown below:

2.       Change the default log location by finding the Log Dir property and modifying the /var prefix to /data/disk1.

3.        Configure HIVE:

a.        Select Existing PostgreSQL database in the Hive Database drop-down list.

b.       Enter Database Name as hive

c.        Enter Database Username as hive

d.       Enter Database URL as jdbc:postgresql://rhel2.hdp3.cisco.com/hive

e.       Enter Database password (use the password created during hive database setup in earlier steps; for example, bigdata)

f.         Click TEST CONNECTION to verify the connectivity

g.       In Advanced tab, Change the default log location by filtering the Log Dir property and modifying the /var prefix to /data/disk1.

h.       Change the WebHCat log directory by filtering the Log Dir property and modifying the /var prefix to /data/disk1.

4.       Configure OOZIE:

a.        Select Existing PostgreSQL database in the Hive Database drop-down list.

b.       Enter Database Name as oozie.

c.        Enter Database Username as oozie.

d.       Enter Database URL as jdbc:postgresql://rhel2.hdp3.cisco.com/oozie.

e.       Enter Database password (use the password created during hive database setup in earlier steps; for example, bigdata).

f.         Click TEST CONNECTION to verify the connectivity.

g.       In Advanced tab, Change the default log location by filtering the Log Dir property and modifying the /var prefix to /data/disk1.

5.        Configure RANGER:

a.        Select POSTGRES from DB FLAVOR drop-down list.

b.       Provide a Ranger DB name. For example, ranger.

c.        Provide Ranger DB Username such as rangeradmin.

d.       Enter JDBC connect string for a Ranger Database as jdbc:postgresql//rhel2.dhp3.cisco.com:54323/ranger.

e.       Ranger DB Host as rhel2.hdp3.cisco.com.

f.         Enter Ranger DB password. (Ranger database is not previously created. provide password string that would be configured in the DB. For example, bigdata).

g.       Select “Yes” for Setup Database and Database User.

h.       Enter “postgres” in Database Administrator (DBA) username.

i.         Enter Database Administrator (DBA) password.

j.         Enter jdbc:postgresql://rhel2.hdp3.cisco.com:5432/postgres in JDBC connect string for root user.

k.        Update Ranger Admin Log Dir from /var to /data/disk1.

HDFS

1.        In Ambari, select the HDFS Service tab and use the “Search” box to filter for the properties mentioned in Table 11and update their values.

2.       Update the following HDFS configurations in Ambari.

Table 11        HDFS Configurations in Ambari

3.        Change the default log location by filtering the Log Dir property and modifying the /var prefix to /data/disk1.

MapReduce2

1.        In Ambari, choose the MapReduce Service tab and update the values as shown below.

2.       Under the MapReduce2 tab, change the default log location by finding the Log Dir property and modifying the /var prefix to /data/disk1.

YARN

1.        In Ambari, select the YARN Service, and update the following as shown in Table 12.

Table 12       YARN Configuration

2.       Under YARN tab, change the default log location by filtering the Log Dir property and modifying the /var prefix to /data/disk1.

The Ambari web UI provides a wizard-driven user experience that allows to configure high availability of the components in many Hortonworks Data Platform (HDP) stack services. The high availability of the components are achieved by setting up primary and secondary components. In the event that the primary component fails or becomes unresponsive, services failover to secondary component. After configuring the high availability for a service, Ambari enables you to manage and disable (roll back) the high availability of components within that service.

Configure the HDFS NameNode High Availability

The HDFS NameNode high availability feature enables you to run redundant NameNodes in the same cluster in an Active/Passive configuration with a hot standby. This eliminates the NameNode as a potential single point of failure (SPOF) in an HDFS cluster. With the release of Hadoop 3.0, you can configure more than one backup NameNode.

Prior to the release of Hadoop 2.0, the NameNode represented a single point of failure (SPOF) in an HDFS cluster. Each cluster had a single NameNode, and if that machine or process became unavailable, the cluster as a whole would be unavailable until the NameNode was either restarted or brought up on a separate machine. This situation impacted the total availability of the HDFS cluster in two major ways:

·         In the case of an unplanned event such as a machine crash, the cluster would be unavailable until an operator restarted the NameNode.

·         Planned maintenance events such as software or hardware upgrades on the NameNode machine would result in periods of cluster downtime.

HDFS NameNode High Availability  avoids this by facilitating either a fast failover to one or more backup NameNodes during machine crash, or a graceful administrator-initiated failover during planned maintenance.

HBase

Under the HBase tab, change the default log location by finding the Log Dir property and modifying the /var prefix to /data/disk1.

Zookeeper

Under the Zookeeper tab, change the default log location by filtering the Log Dir property and modifying the /var prefix to /data/disk1.

Storm

Under the Storm tab, change the default log location by finding the Log Dir property and modifying the /var prefix to /data/disk1.

Ambari Metrics

1.        Choose the Ambari Metrics Service and expand the General tab and make the changes shown below.

2.       Enter the Grafana Admin password as per organizational policy.

3.        Change the default log location for Metrics Collector, Metrics Monitor and Metrics Grafana by finding the Log Dir property and modifying the /var prefix to /data/disk1.

4.       Change the default data dir location for Metrics Grafana by finding the data Dir property and modifying the /var prefix to /data/disk1.

Accumulo

Select Accumulo Service and change the default log location by finding the Log Dir property and modifying the /var prefix to /data/disk1.

Atlas

Under the Atlas tab, change the default log location by finding the Log Dir property and modifying the /var prefix to /data/disk1.

Kafka

Under the Kafka tab, change the default log location by finding the Log Dir property and modifying the /var prefix to /data/disk1.

Knox

1.        Select the Knox Service tab and expand the Knox gateway tab and make the changes shown below.

2.       Enter the Knox Master Secret password as per organizational policy.

3.        For Knox, change the gateway port to 8444 to ensure no conflicts with local HTTP server.

SmartSense

The SmartSense account requires the Hortonworks support subscription. Subscribers can populate the properties as shown below:

Spark

Select the Spark tab, change the default log location by finding the Log Dir property and modifying the /var prefix to /data/disk1.

Review

The assignments that have been made are displayed. Check to make sure all is correct before clicking the Deploy button. If any changes are necessary, use the left navigation bar to return to the appropriate screen.

Deploy

1.        When the review is complete, click DEPLOY.

2.       Follow the onscreen installation process. Watch for warnings and failures by clicking the link as shown below:

Summary of the Installation Process

1.        On the Summary page click “COMPLETE.”

High Availability for HDFS NameNode and YARN ResourceManager

High availability for NameNode and YARN Resource Manager can be configured using Ambari or also on non-Ambari clusters. This deployment guide covers the configuration of high availability using Ambari – Use the Ambari wizard interface to configure HA of the components.

The Ambari web UI provides a wizard-driven user experience that allows to configure high availability of the components in many Hortonworks Data Platform (HDP) stack services. The high availability of the components are achieved by setting up primary and secondary components. In the event that the primary component fails or becomes unresponsive, services failover to secondary component. After configuring the high availability for a service, Ambari enables you to manage and disable (roll back) the high availability of components within that service.

Configure the HDFS NameNode High Availability

The HDFS NameNode high availability feature enables you to run redundant NameNodes in the same cluster in an Active/Passive configuration with a hot standby. This eliminates the NameNode as a potential single point of failure (SPOF) in an HDFS cluster. With the release of Hadoop 3.0, you can configure more than one backup NameNode.

Prior to the release of Hadoop 2.0, the NameNode represented a single point of failure (SPOF) in an HDFS cluster. Each cluster had a single NameNode, and if that machine or process became unavailable, the cluster as a whole would be unavailable until the NameNode was either restarted or brought up on a separate machine. This situation impacted the total availability of the HDFS cluster in two major ways:

·         In the case of an unplanned event such as a machine crash, the cluster would be unavailable until an operator restarted the NameNode.

·         Planned maintenance events such as software or hardware upgrades on the NameNode machine would result in periods of cluster downtime.

HDFS NameNode HA avoids this by facilitating either a fast failover to one or more backup NameNodes during machine crash, or a graceful administrator-initiated failover during planned maintenance.

Active NameNode honors all the client requests and the Standby NameNode acts as a backup. The Standby NameNode keeps its state synchronized with Active NameNode through a group of JournalNodes(JNs). When the Active node performs any namespace modification, the Active node durably logs a modification record to a majority of these JNs. The Standby node reads the edits from the JNs and continuously watches the JNs for changes to the edit log.

Prerequisites for NameNode High Availability

The following are the prerequisites for NameNode high availability:

·         NameNode Machine: Hardware for Active and Standby node should be exactly identical.

·         JournalNodes Machine: JournalNode daemon is relatively lightweight, therefore it can be co-located on machines with other Hadoop daemons; it is typically located on the management nodes.

·         There MUST be at least three JournalNodes, because the edit log modifications must be written to a majority of JNs. This allows the system tolerate failure of a single machine. You may also run more than three JournalNodes, but in order to increase the number of failures that the system can tolerate, you must run an odd number of JNs (3, 5, 7, and so on).

·         ZooKeeper Machines: For automatic failover capability, an existing Zookeeper cluster must exist. The Zookeeper service can also co-exist with other Hadoop daemons.

·         In HA Cluster, the Standby NameNode also performs the checkpoints of the namespace state, therefore do not deploy a Secondary NameNode, CheckpointNode, or BackupNode in an high availability cluster.

Deploy the NameNode High Availability Cluster

To deploy the NameNode high availability on a Ambari managed cluster, follow these steps:

1.        Log into Ambari. Click HDFS > CONFIGS. Click the ACTIONS drop-down list and click Enable NameNode HA to launch the wizard.

Figure 37           Enable NameNode HA

2.       Step 1 launches the Enable NameNode HA wizard. On the Get Started page, specify the Nameservice ID as shown below. Click Next.

Figure 38           Enable NameNode HA Wizard – Get Started

3.        On the Select Hosts page, select the Additional NameNode and JournalNode. Click Next.

Figure 39           Enable NameNode HA Wizard – Select Hosts

4.       On the Review page, confirm the selection. To change any values, click Back, or to continue click Next.

Figure 40          Enable NameNode HA Wizard – Review

5.        Create a checkpoint on the NameNode on the linux server (rhel1.hdp3.cisco.com) as shown below:

Figure 41           Enable NameNode HA Wizard – Create Checkpoint

6.       SSH to current NameNode, rhel1.hdp3.cisco.com and run the following commands:

Figure 42          Current NameNode – Safe Mode and Create Checkpoint Command

7.        Return to the Ambari web UI, verify that the Checkpoint was created. Click Next.

8.       See the progress bar on the Configure Components page. When the configuration steps are completed, click Next.

Figure 43           Enable NameNode HA Wizard – Configure Components

9.       Initialize the JournalNodes as shown below:

Figure 44          Enable NameNode HA Wizard – Initialize JournalNodes

10.    SSH to the current NameNode, for example rhel1.hdp3.cisco.com.

11.     Run the following command:

Figure 45           Initialize JournalNodes

12.    Return to the Ambari UI, when Ambari detects success, click Next.

13.     On the Start Components page, when completed, click Next.

Figure 46          Enable NameNode HA Wizard – Start Components

14.    On the Initialize Metadata page, add the information as shown below:

Figure 47           Enable NameNode HA Wizard – Initialize Metadata

15.     SSH to rhel1.hdp3.cisco.com and run the following command:

Figure 48          Initialize the Metadata for NameNode

16.    SSH to an additional NameNode, for example, rhel3.hdp3.cisco.com and run the following command:

Figure 49          Initialize the Metadata for Additional NameNode

17.     Return to the Ambari web UI, click NEXT. Click OK on the confirmation pop-up window. Make sure the initialization of metadata was performed in NameNode and an additional NameNode as mentioned in step 15 and 16.

Figure 50           Enable NameNode HA Wizard – Confirmation for Initialize the Metadata

18.    On the Finalize HA Setup page, you can see the progress bar as the high availability completes.

Figure 51           Enable NameNode HA Wizard – Finalize High Availability Setup

19.    Click COMPLETE when done.

20.    Click HDFS > SUMMARY tab, verify the Active and Standby NameNode. The Quick Links pane also shows that rhel1.hdp3.cisco.com is running the Active NameNode and rhel3.hdp3.cisco.com is running in Standby NameNode.

Figure 52           Ambari – HDFS – Summary Information

Configure the YARN ResourceManger HA

This section provides instructions on setting up the ResourceManager (RM) HA feature in a HDFS cluster. The Active and Standby ResourceManagers embed the ZooKeeper based ActiveStandbyElector to determine which RM should be active.

Prerequisites for ResourceManager HA

The following are the prerequisites for ResourceManager HA:

·         The servers where Active and Standby RMs are run should have identical hardware.

·         For automated failover configurations, there must be an existing Zookeeper cluster available. The ZooKeeper service nodes can be co-located with the other Hadoop daemons.

Deploy the ResourceManager HA

ResourceManager HA can be configured manually or through Ambari. These instructions are based on configuring ResourceManager HA using Ambari.

To setup ResourceManager HA, follow these steps:

1.        From the Ambari web UI, click Services > YARN. Click the ACTIONS drop-down list and select Enable ResourceManager HA.

Figure 53           Services/YARN - Enable ResourceManger HA

2.       This launches the ResourceManager HA wizard as shown below. Click NEXT.

Figure 54           Enable ResourceManager HA Wizard – Get Started

3.        From the Select Host page, select the host for Additional Resource Manager. Click NEXT.

Figure 55           Enable ResourceManger HA Wizard – Select Host

4.       Proceed to the Review page.

Figure 56           Enable ResourceManager HA Wizard - Review

5.        The Configure Components page shows the progress bar as the Additional ResourceManager is being deployed.

Figure 57           Enable ResourceManager HA Wizard – Configure Components

6.       Click COMPLETE when done.

7.        Verify the ResourceManger HA setup by clicking Services > YARN > SUMMARY tab. The Quick Links pane identifies Active and Standby ResourceManager.

Figure 58           Services/YARN/Summary – Verify ResourceManger HA

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 13 for BOM for the Hadoop Base rack, Table 14 for BOM for Data Science(AI/ML) Expansion Rack, Table 15 for BOM for Hadoop Tiered Storage. Table 16, Table 17, and Table 18 for software components. Table 19 lists Cloudera SKUs available from Cisco.

Table 13        Bill of Materials for C240M5SX Hadoop Nodes Base Rack

Table 14       Bill of Materials for Hadoop Nodes Expansion Rack

Table 15        Bill of Materials for Hadoop Nodes Tiered Storage Rack

Table 16       Bill of Materials for CDSW Nodes Rack

Table 17        Red Hat Enterprise Linux License

Table 18       Cloudera Data Science Workbench Software

Table 19       Cloudera SKU’s available at Cisco

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 drives on Name node and other nodes, If the drive state shows up as JBOD, creating RAID in the subsequent steps will fail with the error “The specified physical disk does not have the appropriate attributes to complete the requested command.”

To configure data drives on Name Node and others, 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. 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 these steps. 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)

Cloudera Data Science Workbench runs on one or more dedicated gateway / edge hosts on HDP clusters. A gateway host is one that does not have any cluster services running on them. They only run the clients for cluster services (such as the HDFS Client, YARN Client, Spark2 Client, and so on). These clients ensure that Cloudera Data Science Workbench has all the libraries and configuration files necessary to securely access the HDP cluster and their respective services.

Cloudera Data Science Workbench does not support running any other services on these gateway hosts. Each gateway host must be dedicated solely to Cloudera Data Science Workbench. This is because user workloads require dedicated CPU and memory, which might conflict with other services running on these hosts.

From the gateway hosts assigned to Cloudera Data Science Workbench, one will serve as the master host, which also runs the CDSW web application, while others will serve as worker hosts. You should note that worker hosts are not required for a fully-functional Cloudera Data Science Workbench deployment. For proof-of-concept deployments you can deploy a 1-host cluster with just a Master host. The Master host can run user workloads just as a worker can.

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

·         CUDA pre-requisites

·         CUDA

·         CDSW pre-requisites

·         CDSW

 

Figure 59           Cisco UCS C240 M5 resource inventory as seen through Cisco UCS Manager

 

Install the Prerequisites for CUDA

To install the prerequisites for CUDA, follow these steps:

 

1.        List GPUs and CPUs installed:

 

 

 

 

Install GCC

To install GCC, follow these steps:

1.        Make sure gcc is installed in the system:

 

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.

 

 

 

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.

To install DKMS, follow these steps:

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

1.        Copy dkms rpm to all the GPU servers:

2.       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 https://www.nvidia.com/Download/index.aspx?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 these steps:

1.        Download CUDA 10.1.

The latest version of CUDA is available here:

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

 

2.       Copy .rpm file to all GPU nodes as follows:

3.        Install CUDA in all GPU nodes by running the following set of commands:

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

Download cuDNN 10.1

To download and set up NVIDIA cuDNN, follow these steps:

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

2.       Copy cuDNN into all the GPU servers.

https://developer.nvidia.com/rdp/cudnn-archive

3.        Run the following Ansible commands in all GPU nodes to setup cuDNN:

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 need to include /usr/local/<cuda>/bin:

3.        Reboot the GPU nodes:

 

Verify Drivers

To verify the drivers have been installed, run the following commands in all GPU nodes as shown below:

 

 

 

 

Installation Prerequisites for CDSW

To install the prerequisites for CDSW, complete the steps in the following sections 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.

With dnsmasq, to add a wildcard DNS subdomain, do the following for just the master host:

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

address=/cdsw/10.15.1.53

#service dnsmasq restart

2.       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 HDP cluster, including Cloudera Data Science Workbench gateway nodes, must use Oracle JDK. OpenJDK is not supported by CDH, HDP Ambari, or Cloudera Data Science Workbench.

Spark 2 which is needed by CDSW to run Spark jobs (on Hadoop nodes) requires JDK 1.8. On CSD-based deployments, the version of Java installed on Cloudera Data Science Workbench gateway hosts and also updating the JAVA home directory in the “cdsw.conf” file located in /etc/cdsw/config directory..

To update the JDK version, go to “Changing the JDK version on an Existing HDP Cluster”

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 HDP 3.1.0.

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 HDP 3.1.0

To download and install CDSW, complete the following steps:

1.        Log into the Ambari Server

2.       Go to the Hosts page and select Actions > + Add New Hosts.

3.        On the Install Options page, enter the fully-qualified domain names for your new hosts.

The wizard also needs the private key file you created when you set up password-less SSH. Using the host names and key file information, the wizard can locate, access, and interact securely with all the hosts in the cluster. Alternatively, you can manually install and start the Ambari agents on all the new hosts.

4.       Click Register and Confirm. For more detailed instructions, refer to “https://docs.hortonworks.com/HDPDocuments/Ambari-2.6.2.2/bk_ambari-installation/content/install_options.html”

5.        The Confirm Hosts page prompts you to confirm that Ambari has located the correct hosts for your cluster and to check those hosts to make sure they have the correct directories, packages, and processes required to continue the install. When you are satisfied with the list of hosts, click Next.

For detailed instructions, go to https://docs.hortonworks.com/HDPDocuments/Ambari-2.6.2.2/bk_ambari-installation/content/confirm_hosts.html.

6.       On the Assign Slaves and Clients page, select the Clients that should be installed on the new hosts. To install clients on all hosts, select the Client checkbox for every host. You can use the all option for each available client to expedite this.

7.        On the Configurations page, select the “configuration groups” for the new hosts.

8.       The Review page displays the host assignments you have made. Check to make sure everything is correct. If you need to make changes, use the left navigation bar to return to the appropriate screen.

9.       Click Deploy.

10.    The Install, Start and Test page displays progress as the clients are installed and deployed on each host. When the process is complete, click Next.

11.     The Summary page provides you a list of the accomplished tasks. Click Complete and you will be directed back to the Hosts page.

Create HDFS User Directories

To run workloads that leverage HDP cluster services, make sure that HDFS directories (/user/<username>) are created for each user so that they can seamlessly connect to HDP from Cloudera Data Science Workbench.

Follow these steps for each user directory that must be created:

1.        SSH to a host in the cluster that includes the HDFS client.

2.       Switch to the hdfs system account user:

3.        Create an HDFS directory for the user. For example, you would create the following directory for the default user admin:

4.       Assign ownership of the new directory to the user. For example, for the new /user/admin directory, make the admin user the owner of the directory:

Install Cloudera Data Science Workbench on the Master Host

To install Cloudera Data Science Workbench on the master host, follow these steps: 

1.        Download Cloudera Data Science Workbench:

For non-airgapped installations, download this file and save it to /etc/yum.repos.d/: cloudera-cdsw.repo

2.       The Cloudera Public GPG repository key verifies that you are downloading genuine packages. Add the repository key:

3.        Install the latest RPM with the following command:

4.       Initialize and start Cloudera Data Science Workbench:

The application will take a few minutes to bootstrap. You can watch the status of the application installation and startup with watch cdsw status.

Install Cloudera Data Science Workbench on Worker Hosts

To install Cloudera Data Science Workbench on worker hosts, follow these steps:

1.        Download Cloudera Data Science Workbench:

For non-airgapped installations, download this file and save it to /etc/yum.repos.d/: cloudera-cdsw.repo

2.       The Cloudera Public GPG repository key verifies that you are downloading genuine packages. Add the repository key:

3.        Install the latest RPM with the following command:

4.       Copy cdsw.conf file from the master host:

After initialization, the cdsw.conf file includes a generated bootstrap token that allows worker hosts to securely join the cluster. You can get this token by copying the configuration file from master and ensuring it has 600 permissions.

If your hosts have heterogeneous block device configurations, modify the Docker block device settings in the worker host configuration file after you copy it. Worker hosts do not need application block devices, which store the project files and database state, and this configuration option is ignored.

5.        Create /var/lib/cdsw on the worker host. This directory must exist on all worker hosts. Without it, the next step that registers the worker host with the master will fail.

6.       Unlike the master host, the /var/lib/cdsw directory on worker hosts does not need to be mounted to an Application Block Device. It is only used to store client configuration for HDP services on workers.

7.        Initialize and start Cloudera Data Science Workbench:

This causes the worker hosts to register themselves with the Cloudera Data Science Workbench master host and increase the available pool of resources for workloads.

8.       Return to the master host and verify the host is registered with this command:

Create the Administrator Account

After your installation is complete, set up the initial administrator account.

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.        To disable Kerberos, delete the file /etc/krb5.conf on the CDSW nodes.

2.       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).

3.        Create User admin in hdfs.

 

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

 

5.        Restart CDSW from command-line using “cdsw stop” followed by “cdsw start”.

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 removed and since by default CDSW version 1.5 does not come with nvidia-docker 1.0 plugins which is a thin wrapper around the Docker CLI and a Docker plugin, shall be installed separately.

1.        Set the following parameter in /etc/cdsw/config/cdsw.conf on all Cloudera Data Science Workbench hosts. You must make sure that cdsw.conf is consistent across all hosts, irrespective of whether they have GPU hardware installed on them.

The following sample steps demonstrate how to use nvidia-docker to set up the directory structure for the drivers so that they can be easily consumed by the Docker containers that will leverage the GPU. Perform these steps on all nodes with GPU hardware installed.

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

 

3.        In CDSW, local mount point (/usr/local/nvidia) for the docker containers will be established manually while creating the directory and copying the files from the CUDA library locations:

 

Test Cloudera Data Science Workbench to Detect 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.

Run a Local Registry

To start the registry container, run the following command:

 

 

 

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.

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

2.       Build a custom engine image out of cuda.Dockerfile using the following sample command:

 

3.        Push this new engine image to a public Docker registry so that it can be made available for Cloudera Data Science Workbench workloads. For example:

 

Image Management

Images can be preloaded on all NodeManager hosts or they can be implicitly pulled at runtime if they are available in a public Docker registry, such as Docker hub. If the image does not exist on the NodeManager and cannot be pulled, the container will fail.

 

Configure CDSW to Run Docker Containers

Set Up Docker Registry

Private trusted registry is required to provision YARN container. This topic provides basic information about deploying and configuring a registry.

Designate a Server for Docker and Start the Registry

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

1.        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.

2.       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.

3.        Create /var/lib/registry folder:

4.       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:

5.        Restart Docker on all nodes.

6.       Provision the registry container by running the following command:

7.        Verify the registry container is provisioned by running docker ps command:

 

Create a Custom CUDA-capable Engine Image

Push Image to the Local Registry

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:

6.       Sign into Cloudera Data Science Workbench as a site administrator.

7.        Click Admin.

8.       Go to the Engines tab.

9.       From the Maximum GPUs per Session/Job drop-down list, select the maximum number of GPUs that can be used by an engine.

10.    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 particular project by completing the following steps:

1.        Go to Projects Overview.

2.       Click on 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.

 

 

Yogesh Ramesh, Big Data Solutions Architect, Computing Systems Product Group, Cisco Systems, Inc.

Yogesh Ramesh is a Big Data Solutions Architect in the Computing Systems Product Group. He is part of the solution engineering team focusing on big data infrastructure, solutions, and performance.

Muhammad Afzal, Engineering Architect, Computing Systems Product Group, Cisco Systems, Inc.

Muhammad Afzal is an Engineering Architect and Technical Marketing Engineer in Cisco UCS Product Management and Datacenter Solutions Engineering. He is currently responsible for designing, developing, and producing validated architectures for Big Data and analytics while working collaboratively with product partners. Previously, Afzal had been a lead architect for various cloud and data center solutions in Solution Development Unit at Cisco. Prior to this, Afzal has been a Solutions Architect in Cisco’s Advanced Services group, where he worked closely with Cisco's large enterprise and service provider customers delivering data center and cloud solutions. Afzal holds an MBA in Finance and a BS in Computer Engineering.

Acknowledgements

For their support and contribution to the design, validation, and creation of this Cisco Validated Design, the authors would like to thank:

·         Karthik Kulkarni, Architect, Computing Systems Product Group, Cisco Systems, Inc.

·         Silesh Bijjahalli, Product Management, Computing Systems Product Group, Cisco Systems, Inc.

·         Ali Bajwa, Hortonworks

·         Wangda Tan, Hortonworks

·         Harsh Shah, Hortonworks

·         Sicong Ji, NVIDIA

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).