Cisco Network Function Virtualization Infrastructure (NFVI) provides the virtual layer and hardware environment in which virtual network functions (VNFs) can operate. VNFs provide well-defined network functions such as routing, intrusion detection, domain name service (DNS), caching, network address translation (NAT), and other network functions. While these network functions require a tight integration between network software and hardware, the use of VNF enables to decouple the software from the underlying hardware.

The following figure shows the high level architecture of Cisco NFVI.

Cisco NFVI includes a virtual infrastructure layer (Cisco VIM) that embeds the Red Hat OpenStack Platform (OSP 13). Cisco VIM includes the Queens release of OpenStack, which is an open source cloud operating system that controls large pools of compute, storage, and networking resources. Cisco VIM manages the OpenStack compute, network, and storage services, and all NFVI management and control functions. Key Cisco NFVI roles include:

• Control (including Networking)

• Compute

• Storage

• Management (including logging, and monitoring)

Hardware that is used to create the Cisco NFVI pods include a specific combination of the following based on pre-defined BOMs. For more details, contact Cisco VIM Product Management.

• Cisco UCS® C240 M4/M5: Performs management and storage functions and services. Includes dedicated Ceph (UCS 240-M4 or UCS 240-M5) distributed object store and file system. (Only Red Hat Ceph is supported).

• Cisco UCS C220/240 M4/M5: Performs control and compute services.

• HP DL360 Gen9: It is a third-party compute where the control plane is Cisco UCS servers.

• Cisco UCS 220/240 M4/M5 (SFF): In a Micropod environment, expandable to maximum of 16 computes.

• Cisco UCS B200 M4 blades: It can be used instead of the UCS C220 for compute and control services. The B200 blades and C240 Ceph server are connected with redundant Cisco Fabric Interconnects managed by UCS Manager.

• Combination of M5 series servers are supported in M5-based Micropod and VIC/NIC (pure 40G) based Hyper-Converged and Micropod offering.

• Quanta servers as an alternate to Cisco UCS servers: Use of specific Quanta servers for the installation of the cloud both at the core and edge. An automated install of Central Ceph cluster to the edge pods is supported for Glance image services.

The UCS C240 and C220 servers are of type M4 or M5 Small Form Factor (SFF) models where the nodes can boot off a pair of HDDs or SSD as specified in BOM. Each UCS C240, UCS C220, and UCS B200 have two 10 GE Cisco UCS Virtual Interface Cards.

The B-Series pod consists of Cisco UCS B200 M4 blades for the Cisco NFVI compute and controller nodes with dedicated Ceph on a UCS C240 M4. The blades and Ceph server are connected via redundant fabric interconnects (FIs) managed by Cisco UCS Manager. The Cisco VIM installer performs bare metal installation and deploys OpenStack services using Docker™ containers to allow for OpenStack services and pod management software updates.

The following table shows the functions, hardware, and services managed by Cisco NFVI nodes.

Note

Internal SSD is the boot device for the storage node You can use any ToR that supports virtual port channel. Cisco recommends you to use Cisco Nexus 9000 SKUs as ToR, which is released as part of Cisco VIM. When Cisco NCS 5500 acts as a ToR, auto-ToR config is mandatory. You must use the automated ToR configuration feature for Cisco NCS 5500. .

Software applications that manage Cisco NFVI hosts and services include:

• Red Hat Enterprise Linux 7.6 with OpenStack Platform 13.0—Provides the core operating system with OpenStack capability. RHEL 7.6 and OPS 13.0 are installed on all target Cisco NFVI nodes.

• Cisco Virtual Infrastructure Manager (VIM)—An OpenStack orchestration system that helps to deploy and manage an OpenStack cloud offering from bare metal installation to OpenStack services, taking into account hardware and software redundancy, security and monitoring. Cisco VIM includes the OpenStack Queens release with more features and usability enhancements that are tested for functionality, scale, and performance.

• Cisco Unified Management—Deploys, provisions, and manages Cisco VIM on Cisco UCS servers.

• Cisco UCS Manager—Used to perform certain management functions when UCS B200 blades are installed. Supported UCS Manager firmware versions are 2.2(5a) and above.

• Cisco Integrated Management Controller (IMC)— Cisco IMC 2.0(13i) or later is supported, when installing Cisco VIM 2.4.

  For the Cisco IMC lineup, the recommended version is as follows:

  UCS-M4 servers

  Recommended: Cisco IMC 2.0(13n) or later. It is also recommended to switch to 3.0(3a) or later for pure intel NIC based pods.

  For Cisco IMC 3.x and 4.y lineup, the recommended version is as follows:

  UCS-M4 servers	Cisco IMC versions are 3.0(3a) or later, except for 3.0(4a). Recommended: Cisco IMC 3.0(4d). Expanded support of CIMC 4.0(1a), 4.0(1b), 4.0(1c). You can move to 4.0(2f) only if your servers are based on Cisco VIC.
  UCS-M5 servers	We recommend that you use CIMC 3.1(2b) or higher versions. Do not use 3.1(3c) to 3.1(3h), 3.0(4a), 4.0(2c), or 4.0(2d). Support of CIMC 4.0(4e) or later. We recommend that you use CIMC 4.0(4e). The bundle version of a minimum of CIMC 4.0 (4d) is needed for Cascade Lake support. For GPU support, you must ensure that the server is running with CIMC 4.0(2f).

  Enables embedded server management for Cisco UCS C-Series Rack Servers. Supports Cisco IMC firmware versions of 2.0(13i) or greater for the fresh install of Cisco VIM. Because of recent security fixes, we recommend you to upgrade Cisco IMC to 2.0(13n) or higher. Similarly, Cisco IMC version of 3.0 lineup is supported. For this, you must install Cisco IMC 3.0 (3a) or above.

  The Quanta servers need to run with a minimum version of BMC and BIOS as listed below:

  SKU Type	BMC Version	BIOS Version
  D52BQ-2U 3UPI (CDC SKU)	4.68.22	3A11.BT17
  D52BE-2U (GC SKU)	4.68.22	3A11.BT17

• Cisco Virtual Topology System (VTS)—It is an open, overlay management and provisioning system for data center networks. VTS automates DC overlay fabric provisioning for physical and virtual workloads. This is an optional service that is available through Cisco VIM.

• Cisco Virtual Topology Forwarder (VTF)—Included with VTS. VTF leverages Vector Packet Processing (VPP) to provide high performance Layer 2 and Layer 3 VXLAN packet forwarding.

Two Cisco VNF orchestration and management applications that are used with Cisco NFVI include:

• Cisco Network Services Orchestrator, enabled by Tail-f—Provides end-to-end orchestration spanning multiple network domains to address NFV management and orchestration (MANO) and software-defined networking (SDN). For information about Cisco NSO, see Network Services Orchestrator Solutions.

• Cisco Elastic Services Controller—Provides a single point of control to manage all aspects of the NFV lifecycle for VNFs. ESC allows you to automatically instantiate, monitor, and elastically scale VNFs end-to-end. For information about Cisco ESC, see the Cisco Elastic Services Controller Data Sheet.

At a high level, the NFVI architecture includes a VNF Manager and NFV Infrastructure.

For cloud networking, Cisco NFVI supports Open vSwitch over VLAN as the cloud network solution for both UCS B-series and UCS C-Series pods. Both B-Series and C-Series deployments support provider networks over VLAN.

In addition, with a C-series pod, you can choose:

• To run with augmented performance mechanism by replacing OVS/LB with VPP/VLAN (for Intel NIC).

• To have cloud that is integrated with VTC which is an SDN controller option.

The Cisco NFVI uses OpenStack services running inside containers with HAProxy load balancing and providing high availability to API and management network messaging. Transport Layer Security (TLS) protects the API network from external users to the HAProxy. Cisco VIM installation also includes service assurance, OpenStack CloudPulse, built-in control, and data plane validation. Day two pod management allows you to add and remove both compute and Ceph nodes, and replace the controller nodes. The Cisco VIM installation embeds all necessary RHEL licenses as long as you use the Cisco VIM supported BOM and the corresponding release artifacts.

The following illustration shows a detailed view of the Cisco NFVI architecture and the Cisco NFVI installation flow.

Cisco Virtualized Infrastructure Manager (VIM) is a fully automated cloud lifecycle management system. Cisco VIM helps to bring up a fully functional cloud in hours, with integrated end-to-end control and data plane verification in place. Cisco VIM offers fully automated day 1 to day n cloud lifecycle management. These include capabilities such as pod scaling (expansion), software update, upgrade, or reconfigure parameters, consolidated logging with rotation and export, software update and upgrade. These have been implemented in line with the operational and security best practices of service providers and enterprises.

The following figure provides the high-level overview of all day-0 and day-n items of Cisco VIM.

Cisco VIM supports installation on two different type of pods. The blade B-series and rack C-series based offering supports NICs that are from Cisco (called as Cisco VIC). You can choose the C-series pod to run in a pure Intel NIC environment, and thereby obtain SRIOV support on the C-series pod. This section calls out the differences in networking between the Intel NIC and Cisco VIC installations.

To achieve network level security and isolation of tenant traffic, Cisco VIM segments the various OpenStack networks. The Cisco NFVI network includes six different segments in the physical infrastructure (underlay). These segments are presented as VLANs on the Top-of-Rack (ToR) Nexus switches (except for the provider network) and as vNIC VLANs on Cisco UCS servers. You must allocate subnets and IP addresses to each segment. Cisco NFVI network segments include: API, external, management and provisioning, storage, tenant and provider.

API Segment

The API segment needs one VLAN and two IPv4 addresses (four if you are installing Cisco VTS) in an externally accessible subnet different from the subnets assigned to other Cisco NFVI segments. These IP addresses are used for:

• OpenStack API end points. These are configured within the control node HAProxy load balancer.

• Management node external connectivity.

• Cisco Virtual Topology Services (VTS) if available in your Cisco NFVI package.

• Virtual Topology Controller (VTC). It is optional for VTS.

External Segment

The external segment needs one VLAN to configure the OpenStack external network. You can provide the VLAN during installation in the Cisco NFVI setup_data.yaml file, but you must configure the actual subnet using the OpenStack API after the installation. Use the external network to assign OpenStack floating IP addresses to VMs running on Cisco NFVI.

Management and Provisioning Segment

The management and provisioning segment needs one VLAN and one subnet with an address pool large enough to accommodate all the current and future servers planned for the pod for initial provisioning (PXE boot Linux) and, thereafter, for all OpenStack internal communication. This VLAN and subnet can be local to Cisco NFVI for C-Series deployments. For B-Series pods, the UCS Manager IP and management network must be routable. You must statically configure Management IP addresses of Nexus switches and Cisco UCS server Cisco IMC IP addresses, and not through DHCP. They must be through the API segment. The management/provisioning subnet can be either internal to Cisco NFVI (that is, in a lab it can be a non-routable subnet limited to Cisco NFVI only for C-Series pods), or it can be an externally accessible and routable subnet. All Cisco NFVI nodes (including the Cisco VTC node) need an IP address from this subnet.

Storage Segment

Cisco VIM has a dedicated storage network used for Ceph monitoring between controllers, data replication between storage nodes, and data transfer between compute and storage nodes. The storage segment needs one VLAN and /29 or larger subnet internal to Cisco NFVI to carry all Ceph replication traffic. All the participating nodes in the pod will have IP addresses on this subnet.

Tenant Segment

The tenant segment needs one VLAN and a subnet large enough to manage pod tenant capacity internal to Cisco NFVI to carry all tenant virtual network traffic. Only Cisco NFVI control and compute nodes have IP addresses on this subnet. The VLAN/subnet can be local to Cisco NFVI.

Provider Segment

Provider networks are optional for Cisco NFVI operations but are often used for real VNF traffic. You can allocate one or more VLANs for provider networks after installation is completed from OpenStack.

Cisco NFVI renames interfaces based on the network type it serves. The segment Virtual IP (VIP) name is the first letter of the segment name. Combined segments use the first character from each segment for the VIP, with the exception of provisioning whose interface VIP name is "mx" instead of "mp" to avoid ambiguity with the provider network. The following table shows Cisco NFVI network segments, usage, and network and VIP names.

For each C-series pod node, two vNICs are created using different ports and bonded for redundancy for each network. Each network is defined in setup_data.yaml using the naming conventions listed in the preceding table. The VIP Name column provides the bonded interface name (for example, mx or a) while each vNIC name has a 0 or 1 appended to the bonded interface name (for example, mx0, mx1, a0, a1).

The Cisco NFVI installer creates the required vNICs, host interfaces, bonds, and bridges with mappings created between all elements. The number and type of created vNICs, interfaces, bonds, and bridges depend on the Cisco NFVI role assigned to the UCS server. For example, the controller node has more interfaces than the compute or storage nodes. The following table shows the networks that are associated with each Cisco NFVI server role.

Note	*ACIINFRA is only applicable when using ACI as a mechanism driver. ** Provider network is extended to controller nodes when VMs are on provider network with virtio.

The network arrangement on third-party HP compute is slightly different from that of Cisco compute running with Intel NIC, because the HP computes have 2 less NIC ports than that are available in the Cisco Intel NIC BOM.

Following table lists the differences in the network arrangement between the Cisco compute and third-party HP compute.

In the initial Cisco NFVI deployment, two bridges are created on the controller nodes, and interfaces and bonds are attached to these bridges. The br_api bridge connects the API (a) interface to the HAProxy. The HAProxy and Keepalive container has VIPs running for each OpenStack API endpoint. The br_mgmt bridge connects the Management and Provisioning (mx) interface to the HAProxy container as well.

The following diagram shows the connectivity between Cisco NFVI nodes and networks.

Supported Layer 2 networking protocols include:

• VLAN over Open vswitch(SRIOV with Intel 710NIC).

• VLAN over VPP/VLAN for C-series Only.

• Single Root Input/Output Virtualization (SRIOV) for UCS B-Series pods. SRIOV allows a single physical PCI Express to be shared on a different virtual environment. The SRIOV offers different virtual functions to different virtual components, for example, network adapters, on a physical server.

The footprint of the cloud offering supported by Cisco VIM has continued to evolve over multiple releases to support customer needs that can vary across multiple dimensions such as cloud capacity, power, physical space, and affordability. The following table shows the available Cisco NFVI hardware and data path deployment combinations.

In the above table:

• fullon indicates the dedicated control, compute and ceph nodes.

• micro indicates converged control, compute and ceph nodes with expandable computes.

• Hyperconverged (HC) indicates dedicated control and compute nodes, but all ceph nodes are compute nodes.

• edge indicates converged control and compute nodes with expandable computes. It communicates with Central ceph cluster for Glance Image service. Persistent storage is not supported.

• ceph indicates converged cephcontrol & cephosd nodes, with an option to add cephosd nodes for glance image services.

Note	The SRIOV support is applicable only for Intel NIC-based pods.

Note	VTF with VTC is only supported on C-series Cisco VIC.

Pod with Intel NICs— In case of the pod having Intel NICs (X710), the networking is slightly different. You need to have atleast two NICs (4x10G) on a single server to support NIC level redundancy. Each NIC is connected to each ToR (connections explained later in the chapter). Since vNICs are not supported in the Intel card, bond the physical interfaces at the host and then create sub-interfaces based on the segment VLAN. Lets call the two NIC cards as NIC_1 and NIC_2 and call their four ports as A, B, C, D. Unlike Cisco VIC based pod, the traffic here is classified as follows:

1. Control plane

2. Data plane (external, tenant and non-SRIOV provider network).

3. SRIOV (optional for provider network). If SRIOV is used, the data plane network only carries external and tenant network traffic.

Control Plane

The control plane is responsible for carrying all the control and management traffic of the cloud. The traffic that flows through control plane are:

1. Management/Provision

2. Storage

3. API

The control plane interface is created by bonding the NIC_1 A port with NIC_2 A port. The bonded interface name is called as samx, indicating that it is carrying Storage, API, Management/Provision traffic (naming convention is similar to Cisco VIC pod). The slave interfaces (physical interfaces) of the bonded interface are renamed as samx0 and samx1. samx0 belongs to NIC_1 and samx1 belongs to NIC_2. Sub interfaces are then carved out of this samx interface based on the Storage, API VLANs. The management/provision traffic will be untagged/native VLAN in order to support pxe booting.

Data Plane

The data plane is responsible for carrying all the VM data traffic. The traffic that flows through the data plane are

• Tenant

• Provider

• External

The data plane is created by bonding the NIC_1 B port with NIC_2 B port. The bonded interface name here would be pet, indicating that it is carrying Provider, External and Tenant traffic. The slave interfaces of this bonded interface would be visible as pet0 and pet1. pet0 belongs to the NIC_1 and pet1 belongs to NIC_2.

In case of OVS/VLAN, the "pet" interface is used as it is (trunked to carry all the data VLANs) to the Openstack cloud, as all the tagging and untagging happens at the Openstack level. In case of Linux Bridge/VXLAN, there will be sub-interface for tenant VLAN to act as the VXLAN tunnel endpoint.

SRIOV

In case of Intel NIC pod, the third (and optionally the fourth) port from each NIC can be used for SRIOV traffic. This is optional and is set or unset through a setup_data.yaml parameter. Unlike the control and data plane interfaces, these interfaces are not bonded and hence there is no redundancy. Each SRIOV port can have maximum of 32 Virtual Functions and the number of virtual function to be created are configurable through the setup_data.yaml. The interface names of the SRIOV will show up as sriov0 and sriov1 on each host, indicating that sriov0 belongs to NIC_1 C port and sriov1 belongs to NIC_2 C port.

In the case of Intel NIC pod, the following table summarizes the above discussion

The following table shows the interfaces that are present on each type of server (role based).

Note	On an Intel pod, all kind of OpenStack networks are created using the physnet1 as the physnet name.

Cisco NFVI UCS servers connect to the ToR switches using Cisco UCS dual-port Virtual Interface Cards (VICs). The VIC is an Enhanced Small Form-Factor Pluggable (SFP+) 10 Gigabit Ethernet and Fiber Channel over Ethernet (FCoE)-capable PCI Express (PCIe) card designed for Cisco UCS C-Series Rack Servers. Each port connects to a different ToR using a Virtual Port Channel (VPC). Each VIC is configured with multiple vNICs that correspond to specific Cisco VIM networks. The UCS Cisco IMC port is connected to an out-of-band (OOB) Cisco management switch.

The following figure shows the UCS C-Series pod Cisco NFVI host to ToR topology.

In the case of Intel NIC, a single two port Cisco VIC in the preceding diagram, is replaced with two 4-port 710 Intel NIC. An extra Intel NIC is added to provide card level redundancy.

Of the four ports that are available in each NIC card, port A is used for management traffic (provision, API, storage, etc), whereas the port B is used for data plane (tenant and provider network) traffic. Port C (and optionally Port D) is dedicated for SRIOV (configured optionally based on setup_data.yaml). Sub-interfaces are carved out of the data and control plane interfaces to provide separate traffic based on specific roles. While the ports A and B from each NIC help in forming bonded interface, the ports C and D over which SRIOV traffic for provider network flows is not bonded. Extreme care should be taken during pod setup, so that ports A, B and C for the Intel NIC is connected to the ToRs. Port D can be optionally used as a second pair of SRIOV ports by appropriate intent defined in the setup_data.yaml file. From Cisco VIM release 2.4.2 onwards, this port option is available for both M4 and M5 based systems or pods.

The following table provides the default link aggregation member pairing support for the pods based on server type:

Note	In M5, a NIC_LEVEL_REDUNDANCY option is introduced to support the M4 default option for link aggregation settings.

From Cisco VIM 2.4.2 onwards, support of M5 full on pods with two port XL-710 across control, compute and dedicated Ceph Nodes, and with NIC_LEVEL_REDUNDANCY is available. This deployment can be achieved with Cisco Nexus 9000 series or Cisco NCS 5500 as ToR. SRIOV is not supported in computes with XL-710. However, the pod can also support computes with four port X-710, where SRIOV is over port C and D.

In Cisco VIM, computes (M4 based testbed) running a Cisco 1227 VIC, and 2 2-port Intel 520 NIC are supported. In this combination, SRIOV is running on the Intel NIC, whereas the control and data plane are carried by virtual interfaces over Cisco VIC.

Cisco VIM 2.4 introduces the support of C220/C240 M5 servers in a micropod configuration with an option to augment the pod with additional computes (upto a max of 16). The M5 micropod environment is based on X710 for control and data plane and an additional XL710 or 2xX710 for SRIOV. The SRIOV card is optional. Once the SRIOV card is chosen, all the computes must have same number of SRIOV ports across the pod.

The following diagram depicts the server network card diagram for the M5 setup.

Cisco VIM 2.4 introduces the first third-party compute. The first SKU chosen is HPE ProLiant DL360 Gen9. In Cisco VIM 2.4, the supported deployment is a full-on pod, with OVS as the mechanism driver, where the management, control, and storage nodes are based on existing Cisco UCS c220/240M4 BOM, and the compute nodes are on HPE ProLiant DL360 Gen9 hardware:

ProLiant DL360 Gen9 with HP Ethernet 1Gb 4-port 331i Adapter - NIC (755258-B21) 2 x E5-2695 v4 @ 2.10GHz CPU
8 x 32GB DDR4 memory (Total 256GB)
1	x Smart Array P440ar hardware RAID card with battery
2	x 1.2 TB - SAS 12GB/S 10k RPM HDD
1	x FlexLOM HP Ethernet 10Gb 2-port 560FLR-SFP+ Adapter
2	x PCIe HP Ethernet 10Gb 2-port 560SFP+ Adapter
System ROM: P89 v2.40 (02/17/2017)
iLO Firmware Version: 2.54 Jun 15 2017

In the case of HP Computes, the FlexLOM HP Ethernet 10Gb interface is used for management and tenant network, and the two additional HP Ethernet 10Gb 2-port 560SFP+ Adapters are used for SRIOV for the provider network. Listed below is network schematic of the HP Compute node.

The Cisco NFVI controller node has four bonds: mx, a, t, and e. Each has a slave interface that is named with the network name association and a mapped number. For example, the management and provisioning network, mx, maps to mx0 and mx1, the API network, a, to a0 and a1, and so on. The bonds map directly to the vNICs that are automatically created on the controller node when it is deployed.

Cisco VIM 3.0 manages a third-party infrastructure based on Quanta servers, thereby bringing in true software abstraction. In the implementation, the supported deployment is a full-on or edge pod, with OVS as the mechanism driver. With the power limitation and rack restrictions on the edge pod, it cannot support hard-drives for the Ceph service. As the Edge pod does not need persistent storage, it is designed to communicate with a central ceph cluster for providing glance image services only.

The installation and management of the Central Ceph cluster is fully automated and it is assumed that the management network of the edge cloud is routable to that of the central Ceph cluster.

In the case of Quanta servers, the networking is similar to that of the HP computes except for the two port 25G (xxv710) Intel NICs. The 2x25GE OCP card is used for control and data plane network over virtio, and the two additional 25GE 2-port xxv710 based Intel NIC Adapters are used for SRIOV via the provider network.

The following figure shows the controller node network-to-bond-to-vNIC interface mapping.

The Cisco NFVI compute node has three bonds: mx, t, and p. Each has a slave interface that is named with the network name association and a mapped number. For example, the provider network, p, maps to p0 and p1. The bonds map directly to the vNICs that are automatically created on the compute node when it is deployed. The following figure shows the compute node network-to-bond-to-vNIC interfaces mapping.

The Cisco NFVI storage node has two bonds: mx and s. Each has a slave interface that is named with the network name association and a mapped number. For example, the storage network, s, maps to s0 and s1. Storage nodes communicate with other storage nodes over the mx network. The storage network is only used for Ceph backend traffic. The bonds map directly to the vNICs that are automatically created on the storage node when it is deployed. The following figure shows the network-to-bond-to-vNIC interfaces mapping for a Cisco NFVI storage node.

Cisco NFVI installation creates two bridges on the controller nodes and interfaces and bonds are attached to the bridges. The br_api bridge connects the API (a) interface to the HAProxy container. The HAProxy and Keepalive container has VIPs running for each OpenStack API endpoint. The br_mgmt bridge connects the Management and Provisioning (mx) interface to the HAProxy container as well.

The following figure shows the connectivity between the mx interface and the br_mgmt bridge. It also shows the connectivity between the br_mgmt and the HAProxy container/namespace using mgmt_out and mgmt interfaces. The figure shows the connectivity between the api interface and the br_api bridge as well as the link between the br_mgmt bridge and the HAProxy container using api_out and mgmt_out interfaces.

A sample routing table is shown below. br_api is the default route and br_mgmt is local to the pod.

[root@c43-bot-mgmt ~]# ip route
default via 172.26.233.193 dev br_api  proto static  metric 425 
172.26.233.0/25 dev br_mgmt  proto kernel  scope link  src 172.26.233.104  metric 425 
172.26.233.192/26 dev br_api  proto kernel  scope link  src 172.26.233.230  metric 425 

[root@c43-bot-mgmt ~]# ip addr show br_api
6: br_api: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP 
    link/ether 58:ac:78:5c:91:e0 brd ff:ff:ff:ff:ff:ff
    inet 172.26.233.230/26 brd 172.26.233.255 scope global br_api
       valid_lft forever preferred_lft forever
    inet6 fe80::2c1a:f6ff:feb4:656a/64 scope link 
       valid_lft forever preferred_lft forever

[root@c43-bot-mgmt ~]# ip addr show br_mgmt
7: br_mgmt: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP 
    link/ether 58:ac:78:5c:e4:95 brd ff:ff:ff:ff:ff:ff
    inet 172.26.233.104/25 brd 172.26.233.127 scope global br_mgmt
       valid_lft forever preferred_lft forever
    inet6 fe80::403:14ff:fef4:10c5/64 scope link 
       valid_lft forever preferred_lft forever

In Cisco VIM, the management node has an interface for API and another interface for provisioning. This is primarily done for security reasons so that internal pod management or control plane messages (RabbitMQ, Maria DB, and so on) do not leak out, and hence reduce the attack vector to the pod. As the name indicates, the API interface is to access the VIM installer API and is also used to SSH to the management node. All external services (installer API, Insight, ELK, and so on) are password that is protected and hang off the API interface. Default route of the management node points to the API interface.

The second interface, also called the provisioning interface is used to PXE boot the various nodes that constitute the OpenStack pod. Typically, provisioning interface is a non-routable interface that is reserved for OpenStack management traffic.

In B-series pod, the networks between provisioning and the UCSM IP need to be routable. Proper ACL has to be applied in the upstream router so that other networks do not interfere with the provisioning network. Depending on the overall deployment, the management node acts as a jump-server to the OpenStack nodes.

Cisco NFVI UCS C-series management node physically connects to the network. Unlike other nodes, the management node does not use multiple vNICs corresponding to specific Cisco NFVI networks. Instead, it connects to the management and API networks using two different physical connections. The management node connects to the management network using a Cisco two-port VIC or first two ports of intel X710-DA4, with each port connecting to a different ToR switch in a VPC configuration. The Cisco VIC card utilizes the default vNICs, but requires the vNICs to be in trunk mode and the default VLAN set to the management network VLAN. The management node connects to the API network using both one Gbps LAN On Motherboard (LOM) ports connected in a port channel configuration. These ports can either connect to the Nexus 9000 Series switch in a VPC configuration, or to an operator-managed switch(es), depending on how the operator wants to segment their network. The Cisco IMC port can optionally be connected to an out-of-band management Catalyst switch.

Management node services, which are required to start the other topology nodes, listen on the management network and the traffic flowing over the vNICs or NICs on that network. These services and the other management network services are unsecured. Secure management node services listen on the management node API network, and their traffic flows over the LOM ports. This service division allows tenants to utilize tighter network access control to the management network than the management node API network. The following figure shows the Cisco NFVI management node (UCS C-Series) API network connections.

Note	Connecting Cisco IMC port to a Cisco OOB management switch is optional.

For the day-0 server automation in Cisco VIM, ensure that the reachability to:

CIMC/ILO/BMC of the individual servers from the management node is available through the br_api network.

Cloud API, external network (for ssh to floating IPs) and provider network from the management node is available, as the VMTP and NFVbench are typically run from the management node.

Note	From the Cisco VIM release 2.4.3 onwards, you can enable or disable the default behavior of the management node reachability from cloud API, external network, and provider network as part of their day-0 configuration.

If you disable the reachability to cloud api, external, and provider network for security reasons, then:

• VMTP and NFVbench are not accessible from the management node.

• Cloud api, external network and provider network must be properly routed as the Cisco VIM cannot automatically valid the same.

You can switch from IPv4 to IPv6 as the number of available routable IPv4 networks is limited. In Cisco VIM, the management network uses the default IPv4 route to reach external service like NTP, DNS, AD/LDAP, SwiftStack, and so on, if it is not locally hosted.

Due to the limited availability of IPv4 address space, if you cannot provide a routable IPv4 network or local or dual-home of the external services that require routing, for example, AD or LDAP, deployment hindrance can occur.

IPv4 is obligatory in Cisco VIM, as the provision network colocates with the management network (mx/samx interface) for baremetal PXE install and Ansible orchestration.

As CEPH and OpenStack control plane communication are on the same management network, you cannot completely remove IPv4 from the management network. However, you can run IPv4+IPv6 dual stack in which IPv4 network can exist in a non-routable private network and IPv6 network can exist in a routable semi private network. This ensures to satisfy the requirements of the CiscoVIM and accessibility to the external services.

In Cisco VIM, the management network supports IPv6 addresses for servers, while the management node is statically allocated from a given pool. The external services that support both IPv4 and IPv6 addresses are DNS, NTP, and AD or LDAP. You can run IPv4+IPv6 (optionally) as the cloud API endpoint. CIMC/BMC can have IPv6 addresses.

You can deploy Cisco NFVI using a combination of Cisco C-Series and B-Series servers. The C-Series management node is connected to the Cisco Nexus 9000 Series ToRs through the Cisco VIC in a VPC configuration. The UCS Fabric Interconnects (FIs) are connected to the ToRs and the UCS B-Series blade chassis is connected to the FIs. The C-Series storage nodes are connected to the ToRs as well. For C-series implementation, see Cisco NFVI Networking Overview. For the combination of the C-Series and B-Series implementation, two exceptions are listed below:

• For UCS B-Series, the Cisco UCS Manager IP address must be available to the Cisco NFVI management network. For UCS C-Series, this requirement is optional.

• The UCS Manager cluster and VIP connections are not attached to one of the Cisco NFVI network segments.

Following figure shows a high-level view of Cisco UCS C-Series and B-Series servers that are used in a Cisco NFVI deployment.

For C-Series pods, each host has a 2x10-GE Cisco network card 1227 from which the installer creates two vNICs for each network to ensure that the network topology has built-in redundancy. The provider network, if needed, is also created from the same network card. Each link of a given network type terminates to a unique Cisco Nexus 9000 switch, which acts as the ToR. The Cisco Nexus 9000s are configured in VPC mode to ensure that the network redundancy. The networking redundancy is extended to the management node, which has a redundant vNIC for the installer API and management or provisioning networks. The following figure shows the C-Series topology.

Note	While the figure depicts UCS 220 M4s as the controller and compute, it also supports UCS 240 M4s as control and compute nodes.

Cisco NFVI uses multiple networks and VLANs to isolate network segments. For UCS C-series management and storage nodes, VLANs are trunked between the ToR switches and the Cisco VICs on the C-Series nodes. For UCS B-series controllers and compute nodes, VLANs are trunked between the ToR switches, the UCS Fabric interconnects, and the B-Series blades. The figure shows the network segments and how each node is attaches to them. The network segments are VLANs that are trunked between the respective upstream switch/FI and the C-Series or B-Series node.

Cisco NFVI high availability (HA) is provided by HAProxy, a single-threaded, event-driven, non-blocking engine combining a fast I/O layer with a priority-based scheduler. HAProxy architecture is layered with bypass mechanisms at each level to ensure that the data does not reach higher levels than needed. Most processing is performed in the kernel.

The following figure shows a detailed view of Cisco NFVI controllers connecting to the API and Management and Provisioning network. It also shows how the bridges are configured and the roles of the HAProxy container and network namespace. The dedicated HAProxy container network namespace was created to avoid split default gateway problems. The namespace allows API segment ingress and egress traffic to have a different default gateway than the one configured on each controller host for non-API traffic. In the illustration, two of the three Cisco NFVI controllers have HAProxy containers and a dedicated Linux network namespace. (Cisco NFVI supports three HAProxy containers).

In the figure, Control Node 1 is attached to the API network segment through the br_api bridge. The br_api bridge connects to the Linux network namespace where the HAProxy container has an interface that is mapped through the api < > api_out interface mapping. The HAProxy container has a default gateway configured that points to the upstream API Layer 3 First Hop Redundancy Protocol (FHRP) VIP. This gateway is used for the HAProxy container incoming and outgoing API traffic.

Outside traffic coming in through the API interface is routed into the API network. The traffic traverses the br_api bridge, goes into the Linux network namespace and then the API VIP (based on the IP address or port) that is listening on the HAProxy container. The HAProxy container establishes a connection with the backend API endpoint (for example, the OpenStack Horizon dashboard) and the return traffic passes through the container and back out the API network following the default gateway for the container on the API network. All other non-API traffic such as the management access over SSH to the Cisco VIM controller comes into the management or provisioning network and access the node directly. Return traffic uses the host-level default gateway that is configured on the Linux (RHEL) operating system.

If an HA event occurs in a Cisco NFVI pod, Cisco VIM automatically shuts down machines by failing over services. Examples include:

• For API servers, HAProxy automatically ensures that the other redundant control services handle requests, avoiding the shutdown/terminated/non-responding one.

• For quorum services, such as Galera, the remaining members of the quorum continue to provide service and HAProxy ensures that new requests go to the remaining processes.

• For an active/standby process such as HAProxy, the system moves the endpoint IP to a standby copy and continues to operate.

All these behaviors are automatic and do not require manual intervention. When the server is restarted, the services automatically come into service and are added to the load balancing pool, joining their quorums or are added as backup services, depending on the service type.

While manual intervention is not needed, some specific failure scenarios (for example, Mariadb, rabbit) can cause problems that require manual intervention. For example, if a complete network failure occurs, the Galera and RabbitMQ clusters can go into three-way partition. While the Cisco NFVI cluster is resilient to single-point failures, two switches failing simultaneously—something highly unlikely in long-running systems—can sometimes happen due to administrative error, in which case, manual intervention is needed. To repair the pod, the management node must be up and running and all the nodes accessible through password-less SSH from the management node. From the installer<tagid> dir, execute:

# ciscovim cluster-recovery

Control nodes recovers after the network partitions are resolved. After executing this command, control nodes services come back to working state. To make sure that the Nova services are good across the compute nodes, execute the following command after sourcing /root/openstack-configs/openrc:

# nova service-list

To check for the overall cloud status, execute the following:

# ciscovim cloud-sanity create test all

#ciscovim cloud-sanity show result all –id <uid of the test >

Block Storage

Cisco NFVI storage nodes utilize Ceph, an open source software for creating redundant, scalable data storage using clusters of standardized servers to store petabytes of accessible data. OpenStack Object Storage is a long-term storage system for large amounts of static data that can be retrieved, leveraged, and updated. It uses a distributed architecture with no central point of control, providing greater scalability, redundancy, and permanence. Objects are written to multiple hardware devices, with the OpenStack software responsible for ensuring data replication and integrity across the cluster. Storage clusters scale horizontally by adding new nodes. if a node fail, OpenStack replicates its content across other active storage nodes. Because Ceph uses software logic to ensure data replication and distribution across different devices, inexpensive commodity hard drives and servers can be used in lieu of more expensive equipment.

Cisco NFVI storage nodes include object storage devices (OSDs), hard disk drives (HDDs), and solid state drives (SSDs). OSDs organize data into containers called objects that a user or application determines are related. The objects reside in a flat address space where they all exist at the same level and cannot be placed inside one another. Each OSD has a unique object identifier (OID) that allows the Cisco NFVI control node to retrieve it without knowing the physical location of the data it contains.

HDDs store and retrieve digital information using one or more rigid rapidly rotating disks coated with magnetic material. The disks are paired with magnetic heads arranged on a moving actuator arm, which read and write data to the disk surfaces. Data is accessed in a random-access manner; individual data blocks can be stored or retrieved in any order and not only sequentially. HDDs are a type of non-volatile memory, retaining stored data even when powered off.

SSDs are solid-state storage devices that use integrated circuit assemblies as memory to store data persistently. SSDs primarily use electronic interfaces compatible with traditional block input/output (I/O) hard disk drives, which permit simple replacements in common applications.

Cisco NFVI storage nodes are managed by the control node applications including Ceph monitoring dashboard, Glance, and Cinder. The Ceph monitoring dashboard provides a view into the overall storage node health. Glance virtualizes pools of block storage devices and provides a self-storage API to request and consume those resources. Cinder is an OpenStack block storage service designed to present storage resources to the OpenStack compute node.

In Cisco VIM, depending on the needs of the user, the number of OSDs a pod can have is between 3 and 20. From release Cisco VIM 3.0.0 onwards, you can choose to have multi-backend Ceph in the same pod, to support different I/O requirements. Currently, this is a day-0 decision. You must decide whether to start with single or multi back-end ceph, with a minimum of three nodes for each backend type. Only 2 backends (one of type HDD and another of type SSD) for each pod is supported. For details on how to use HDD or SSD based ceph, see Cisco Virtualized Infrastructure Administrator Guide.

Cisco VIM supports NetApp devices running ONTAP 9.X or higher. NetApp devices are added as an alternate to Ceph for block storage. Cisco VIM has been integrated and tested with FAS2650 SKU of NetApp, however it does not preclude Cisco VIM from working with SKUs of NetApp that are compatible FAS2650. Now, you have to choose the blockstorage and the hardware from Day 0.

Object Storage

Cisco VIM provides an integration with SwiftStack, an object storage solution. In this case, the SwiftStack is installed and managed outside the Cisco VIM ahead of time, and the VIM orchestrator adds the relevant Keystone configuration to access the SwiftStack endpoint. In addition to Keystone integration, the Cinder service is also configured to support backup of the volumes to SwiftStack object store. In the current integration, the SwiftStack endpoint has to be in a network routable to/from the CiscoVIM API network (as the VIM API is the same as the Keystone public endpoint network). In the current release, because of limitations in SwiftStack, Cisco VIM is integrated only with KeystoneV2.

In Cisco VIM, you can choose to use Solidfire as an option for block storage along with Ceph. In this scenario, the backend for Glance is Ceph, and the customers have a choice for the Cinder backend to be Ceph or Solidfire. The Cinder block storage service manages the creation, attachment, and detachment of these volumes between a storage system, such as, SolidFire, and different host servers. Also, in Cisco VIM, the date in Solidfire will be backed by Ceph. The Solidfire cluster is pre-deployed and has 2 networks: management and storage. It is recommended that:

• The storage network for Cisco VIM is same as that for Solidfire.

• The management network for Solidfire is reachable from Cisco VIM control nodes.

The Cisco Virtual Topology System (VTS) is a standards-based, open, overlay management and provisioning system for data center networks. It automates the data center overlay fabric provisioning for both physical and virtual workloads.

Cisco VTS provides a network virtualization architecture and software-defined networking (SDN) framework that meets multitenant data center cloud service requirements. It enables a policy-based approach for overlay provisioning.

Cisco VTS automates network overlay provisioning and management tasks, integrates with OpenStack and simplifies the management of heterogeneous network environments. Cisco VTS provides an embedded Cisco VTS GUI and a set of northbound Representational State Transfer (REST) APIs that is consumed by orchestration and cloud management systems.

Cisco VTS architecture has two main components: the Policy Plane and the Control Plane. These perform core functions such as SDN control, resource allocation, and core management function.

• Policy Plane—Enables Cisco VTS to implement a declarative policy model that captures user intent and converts it into specific device-level constructs. Cisco VTS includes a set of modular policy constructs that can be organized into user-defined services for use cases across service provider and cloud environments. The policy constructs are exposed through REST APIs that is consumed by orchestrators and applications to express user intent, or instantiated through the Cisco VTS GUI. Policy models are exposed as system policies or service policies.

• Control Plane—Serves as the SDN control subsystem that programs the various data planes including the VTFs residing on the x86 servers, hardware leafs, DCI gateways. The control plane hosts the Cisco IOS XRv Software instance that provides route peering capabilities between the DCI gateways or to a BGP route reflector. (Cisco IOS XRv is the virtualized version of Cisco IOS XR Software.) The control plane enables an MP-BGP EVPN-based control plane for VXLAN overlays originating from leafs or software VXLAN tunnel endpoints (VTEPs)

The Cisco NFVI implementation of Cisco VTS includes the VTS Virtual Topology Forwarder (VTF). VTF provides a Layer 2/Layer 3 (L2/L3) software switch that can act as a software VXLAN terminal endpoint (VTEP). VTF is a lightweight, multitenant software data plane designed for high performance packet processing on x86 servers. VTF uses Vector Packet Processing (VPP). VPP is a full-featured networking stack with a software forwarding engine. VTF leverages VPP and the Intel Data Path Development Kit (DPDK) for high performance L2, L3, and VXLAN packet forwarding.

VTF allows Cisco VTS to terminate VXLAN tunnels on host servers by using the VTF as a software VXLAN Tunnel Endpoint (VTEP). Cisco VTS also supports hybrid overlays by stitching together physical and virtual endpoints into a single VXLAN segment.

The figure below shows the Cisco VTS architecture and high-level flow when installed in Cisco NFVI. Cisco VTS is installed on separate UCS servers, the Virtual Topology Controller plugin is installed on the control node, and the VTF is installed on the compute node.

1. The OpenStack user invokes the OpenStack Neutron API.

2. Neutron uses the VTS plugin and driver to make calls to the VTC REST API.

3. VTS control components interact with the VTF agent to carry out the corresponding dataplane setup.

4. During Cisco NFVI installation, the Cisco NFVI Installer installs the OpenStack Neutron VTC plugin and driver on the Cisco NFVI controller node, and installs the VTF component (including VPP) on the Cisco NFVI compute node.

The following illustration shows that the Cisco NFVI networking after the Cisco VTS is installed. The SDN controller nodes are an addition to the existing Cisco NFVI pod.

Cisco VIM solution uses Cisco NFVI Monitor (NFVIMON) to monitor the health and performance of the NFVI. This includes monitoring both the physical and logical components of one or multiple NFVI pods. NFVIMON feature is enabled by the Zenoss which provides for extensive monitoring and collection of performance data for various components of the cloud infrastructure including Cisco UCS blade and rack servers, service profiles, Nexus top of rack switches, fabric interconnects, and also the OpenStack instances. The monitoring system is designed such that it can monitor single or multiple pods from a single management system. NFVIMON is integrated into Cisco VIM as an optional component. NFVIMON is enabled by extending the setup_data.yaml file with relevant information. To enable the NFVIMON, refer to Enabling NFVIMON on Cisco VIM. Also, NFVIMON can be enabled on an existing pod, through the reconfigure option. To reconfigure through Insight UI, refer to Reconfiguring Optional Services. Then, the pod is added as a new VIM resource to be monitored in the Monitoring UI.

The NFVIMON architecture supports monitoring of one or more Cisco VIM pods. There is no limit on the number of pods, but note that the setup supports up to ~ 25000 managed resources across pods, where a managed resource is a physical device, network device or virtual machine tracked from a monitoring perspective.

NFVIMON consists of four components: ceilometer services (for data collection), collector, resource manager (RM), and control-center (CC) with Cisco Zenpacks. As NVIFMON is a third-party software, its integration with the VIM is loosely coupled and the VIM automation only deals with installing the ceilometer service software required to monitor the pod.

The installation of other NFVIMON components (collector, RM, and CC with Cisco NFVI Zenpacks) are Cisco Advance Services led activity and those steps are outside the scope of the current install guide. Make sure that you have engaged with Cisco Advance Services on the planning, image information (of collector with Cisco NFVI Zenpacks and RM), and installation of the NFVIMON accessories along with its network requirements.

Start with one Cisco VIM pod (Pod A in the picture) and two external nodes (one to host 2 Collector VMs and one for remote management to host 1 control-center with Cisco Zenpacks and 2 RM VMs) of multiple pods.

Monitor the Cisco VIM pods at the time of installation if NFVIMON is enabled, or by adding NFVIMON after installation. Install the collectors manually in the external collector node, so that the pod is added for monitoring in the control center.

Note	From Cisco VIM 3.2.0, you can use non-root admin keys for monitoring purposes.

Note	From Cisco VIM 3.4.1, you can run NFVIMON along with CVIM MON. This capability eases the transition from NFVIMON to CVIM MON and allows operators to evaluate CVIM MON while using NFVIMON.

NFVIMON supports the functionality of high availability (HA). HA is achieved through dual polling of the redundant collectors over an active-active deployment. VM is deployed between the two physical collector servers with two sets of collectors. Two separate Zenoss CC-RMs are connected to one set of collectors each, to aid in simultaneous monitoring of the pod. Ceilometer is deployed in Cisco VIM pod such that it sends data to two separate collector VIPs simultaneously. To enable the NFVIMON, refer to Enabling NFVIMON on Cisco VIM. The NFVIMON HA architecture is depicted in the below figure.

Note	Pods running with NFVIMON in standalone mode, cannot be moved to HA mode through reconfiguration.

You can enable NFVIMON HA on day-0 or day-1, when the pod is not running with NFVIMON in the first place.

You can deploy Cisco VIM with a lightweight pod-level monitoring solution known as CVIM-MON which is based on the open source PTG stack (Prometheus, Telegraf, Grafana). This solution is available as an add-on from both commercial and feature point of view, and provides the following services:

• Infrastructure-level metric collection based on metric collection agents installed on all nodes in the pod and on specialized collectors running on the management node.

• Metric aggregation into a time series database (TSDB) installed on the management node.

• Rule-based alerting engine integrated in the management node.

• TSDB visualization web server installed on the management node with pre-defined dashboards customized for Cisco VIM.

This image is not available in preview/cisco.com

All CVIM-MON components are containerized, except for the Telegraf agents which run on bare metal on all nodes in the pod (including the management node). The two sub-components of CVIM-MON are:

CVIM_MON: Provides the base functionality of monitoring and KPIs.

SNMP: It is enabled only if CVIM_MON is enabled.

Cisco VIM provides telemetry services to collect meters within an OpenStack deployment. Cisco VIM Telemetry service is built on Ceilometer and Gnocchi in OpenStack Queens release. You can retrieve metrics using OpenStack CLI and REST APIs. Pods must have Ceph for persistent storage of the metrics which are collected every five minutes and retained for 48 hours. As Ceph is required for ceilometer, you can install ceilometer as part of fresh installation of the cloud, that is, ceilometer cannot be brought in as a reconfigure option. Also, the ceilometer is supported only on fullon pod. The diagram below illustrates the high-level architecture of the telemetry services.

To view the summary of the metrics, see Telemetry for OpenStack of Cisco Virtualized Infrastructure Admin Guide, Release 3.0.0.

Cisco VIM UM, a light-weight UI, is introduced in Cisco VIM to ease the deployment and management of the NFVI platform. This feature is available as an add-on from both commercial and feature point of view. Also, Cisco VIM Insight offers a single pane of glass service to provide deployment visualization and to manage multiple Cisco VIM pods thereby reducing user-errors.

Cisco VIM UM supports multi-tenancy with local RBAC support and is easily integrated with the CiscoVIM REST layer. The container based UI platform is loosely coupled, and can help manage multiple CiscoVIM pods right from day-0, or later in the lifecycle of the cloud.

The architecture of the CiscoVIM UM is light-weight, hierarchical and scalable. While it introduces an ease of management from the global UI, each local site is autonomous with localized toolsets. The Global Unified Management UI, provides ease of management with multi-site multi-pod capability for distributed NFV deployment at scale. Also, CiscoVIM UM is designed to operate in HA as an option. The platform is a modular, loosely coupled architecture, that will provide the capability to manage multiple pods, with RBAC support as shown in the figure .

NFVbench is a containerized network benchmarking tool that is introduced in Cisco VIM, to bring consistent methodology to measure the network performance of the cloud. NFVbench is released in a container that is preinstalled on the management node if the NFVBENCH option is selected in the Cisco VIM configuration file.

The main goal of NFVbench is to measure the cloud performance that is based on real cloud deployment traffic patterns. During the test, the packet path traverses through every network element that participates in the production environment; that is traffic flows through a switch (ToR) to v-switch on compute node, continues to VM representing any basic VNF in NFV deployment and comes back in similar way on different ports. Network performance or throughput is computed based on sent and received traffic.

NFVbench can bring up one or more chains of test VMs, where each chain consists of one or two VMs. The example below illustrates a case with a two-VM chain.

Reports from NFVbench show data measurements from every hop in the path, which makes it easier to detect configuration errors or potential bottlenecks. NFVbench sends UDP packets that are generated by open-source traffic generator (TRex) already included in the container. Advanced testing using NFVbench allows to conduct multi-chain and multi-flow testing. Multi-chain testing enables running multiple parallel independent packet paths at the same time, while the multi-flow testing performs IP ranging in packet headers within every chain. The below figure illustrates a NFVbench result test execution with two parallel chains with one VM each.

NDR/PDR and Fixed Rate Tests

NDR/PDR Test: NFVbench offers a more advanced test (called the NDR/PDR test), provides information about network throughput using any of the standard defined packet sizes - 64B, IMIX,1518B. NDR (No Drop Rate) value represents throughput at which no packets are dropped (satisfied by less than 0.001% of packets being dropped). Similarly, PDR (Partial Drop Rate) represents throughput at which only small number of packets is dropped ( less than 0.1% of packets sent).

Fixed Rate Test: NFVbench offers a simple test to run traffic at fixed rate, which verifies that every network component of packet path works properly. It is useful for identifying bottlenecks in the test environment. Traffic generator generates packets at fixed rate for the given time by the user. From the statistics that is collected, drop rates and latencies are computed and displayed.

Both the NDR/PDR Test and Fixed Rate Test types of test provide a way of verifying network performance of NFV solution.

Supported Encapsulation and Protocols

NFVbench supports all networking options that can be deployed with Cisco VIM:

• OVS

• VPP with VLAN or VxLAN

• SR-IOV

While the use of ACI plugin brings in the flexibility of dynamic allocation of tenant and provider VLANs on demand, it also ties the OVS version to the ACI plugin. This leads to an extreme tight coupling of Cisco VIM and ACI. Also, with an APIC plugin there are might be gaps to cover certain use-cases, for example, where there is a need to have flexibility of different access type (tagged vs non-tagged) for the same VLAN but for different servers.

To address such use-case or avoid tight coupling of OVS with ACI plugin, an optional solution is available to automate the target VLANs on the right switch port based on server role on day-0 along with corresponding fabric access and tenant policy configurations via the ACI API.

With this option, the setup_data for each Cisco VIM instance is the single source for the server to switch port mappings. This solution can handle switch provisioning with the correct VLANs during addition/removal of server and provider/tenant VLAN range expansion via reconfiguration option. This solution is based on the fact that the PV (port*VLAN) count in a given ACI Fabric domain is under the scale limits 10000 PV/ToR and 450000 PV/Fabric.

Cisco VIM supports NCS-5500 as an alternate to a Nexus ToR. NCS-5500 is an IOS XR-based router, which is similar to Nexus switches. You can use the 48 10/25G ports or the 6 40/100G uplink ports model to implement NCS-5500 (port-numbers depend on NCS version). Also, other SKUs of NCS-5500 are supported as long as the NCS-5500 software supports the EVLAG feature. NCS-5500 uses the technology of bridge domain to connect to the server. Enable the Auto ToR configuration feature to support NCS-5500 as ToR. NCS-5500 supports a micropod with more computes running on Intel 710 NICs with the mechanism driver of VPP over LACP. The support is extended to include 40G/100G based NCS-5500 SKUs with splitter cables (of 4x10) connecting to the servers, which helps in increasing the server port density by four folds.

Cisco VIM uses the disk-maintenance tool that gives you the ability to check the status of all hard disk drives present in the running and operational mode in the following nodes:

• management node

• specific or all controller servers

• specific or all compute servers

Status of the disks such as online, offline, rebuilding helps you to identify which particular disks in which slot has potentially gone bad and require to be physically replaced in the server. It can be run on servers that have either a RAID controller or an SAS passthrough controller.

Once the disk is physically replaced, Disk management tool can be used to add the new disk back into the system as part of the RAID system (recommended one server at a time).

Note

Disk Maintenance tool is useful only when one or at most two (in RAID6) go bad. Failure of more than one disk at a time puts the entire server in an irrecoverable state. Replace the server using remove and add operations through ciscovim. Disk management is not supported on a third party compute due to the licensing issue with the HPE SmartArray Utility tool.

OSD maintenance tool gives you the ability to check the status of all OSDs and their corresponding physical hard disk drives present in the running and operational storage nodes. The status of the OSDs is reported along with the HDD mapping.

OSD Maintenance tool helps you to identify the status of the OSD (Up or Down) and its corresponding hard disk drive slot in the server that requires to be physically replaced. OSD Maintenance tool can run on servers that have either a RAID or an SAS passthrough controller.

Once the HDD to be physically replaced is identified, the same OSD tool can be used to rebalance the ceph tree, remove the OSD from the cluster, and unmount the disk drive, in preparation for the disk removal. After the disk has been physically replaced, the tool can be used to add the new disk back into the system as part of the Ceph cluster and recreate the OSD (only one HDD/OSD at a time). It ensures to replace a bad HDD, it is not required to remove the ceph cluster from operation and then add it back through remove-storage and add-storage options in ciscovim.

Note	OSD tool does not support the replacement of the internal OS drives and the external journal drives, for which you still have to use add or remove of OSD nodes.

Cisco VIM pods has many compute servers, but the actual usage of the compute servers are limited at times. To optimize the overall power consumption of the data center, we have to power down the server through an API/CLI.

To prevent the cloud destabilization, you cannot power off all the compute nodes. For example, one cannot power off all the compute nodes, at least one pod has to be Active.

Pod management operation(s) applies to the entire pod during updating and reconfigure, the server.

Updating and reconfiguration are not possible under the following circumstances:

• If one or more compute nodes are powered off.

• Computes on which VMs are running cannot be powered-off.

• Computes with. All-in-one (AIO) nodes in a micro-pod) cannot be powered-off through this API.

When there is a power-off, internally cloud-sanity is run and if the cloud sanity fails, then the power-off action is aborted.

Cisco VIM has been tuned to deliver performance from an infrastructure and VNF point of view. The following are the details of the physical cores (regardless of hyper-thread enabled or not) that the infrastructure needs. Number of cores that are reserved for the system (host system + OpenStack services) is 2 in all cases and is included in the count that is shown in the following table.

For OVS deployments, use V=0 and Vr=0

Some VPP deployments with high throughput requirements may require more than 2 VPP cores.

Cisco VIM is supported in an air-gapped (disconnected mode) environment. You can use a USB or Cisco VIM Software Hub for an air-gapped install. When the number of pods is more, shipping USBs for an air-gapped install and update is not scalable. In such scenarios, we recommend that you use Cisco VIM Software Hub.

Cisco VIM Software Hub contains the Cisco VIM release artifacts such as buildnode ISO, Cisco VIM code, docker registry, and docker images. Using the management node, you can access the release artifacts from the Cisco VIM Software Hub.

You can install the artifacts available on the Cisco VIM Software Hub server through a connected or a disconnected install procedure. For a connected install, one end of the Cisco VIM Software Hub server is connected to the internet, and the other end is connected to the datacenter.

The following figure shows the architecture of a connected install.

For a disconnected install, both interfaces are private and the artifacts are installed on the Cisco VIM Software Hub using the USB procedure. You must ensure that the ssh interface (br_api) of the management node for each Cisco VIM pod can connect to the enterprise facing interface of the Cisco VIM Software Hub server through Layer 2 or Layer 3 networking. From release Cisco VIM 3.0.0, the Cisco VIM Software Hub is supported over dual-stack network.

From release Cisco VIM 2.4.3 onwards, seamless connectivity from VNFs of the private cloud to the customer premise private cloud is enabled. The architecture of the Cisco VIM Tenant L2 Connectivity is depicted below:

To set up Cisco VIM tenant L2 connectivity architecture, the following assumptions are made:

• OpenStack can manage VLAN allocation.

• You must manage VXLAN network and subnet for overlays, and enable OpenStack to use the EVI/VNID by creating appropriate networks/subnets in OpenStack. Cisco VIM supports VNI ranging from 1 to 65535.

• BGP configuration (peer, ASes) will be provided at the time of Cisco VIM cloud deployment through setup_data.yaml.

VXLAN tunnel is used for traffic between the VNF and customer Private cloud, while the VLAN is used for the traffic within the pod or across VNFs. EVPN is used to share L2 reachability information to the remote end, and Cisco NCS 5500 in EVLAG mode acts as a conduit for the traffic. For the VXLAN/EPVN solution to work, Cisco VIM and VXLAN tunnel peers with an external BGP route reflector to exchange IP address to Mac Binding information as shown in the below figure.

From a control plane point of view, three instances of GoBGP (in Active-Active-Active mode) run on the controller nodes to establish L3 peering with the external BGP RR for importing or exporting VxLAN routes into or from Cisco VIM respectively. The imported information is then pushed into etcd, to maintain a single source of the information within Cisco VIM.

VPP agents create and program VTEP on VPP, and also create a VXLAN tunnel interface for the VM based on the VNI information from Neutron. VPP updates VNF IP/MAC mapping in etcd, which gets exported out through EVPN to the BGP RR.

Multi-VXLAN EVPN Design

From release Cisco VIM 2.4.6 onwards, multiple-AS VXLAN EVPN overlay networks are supported. The following image depicts the schematic view of the multiple-AS VXLAN EVPN overlay network.

One set of VXLAN overlays manage the Cloud exchange traffic, while the other set of VXLAN overlays manage the Cloud management traffic. The multi-VXLAN (multi refers to 2) is used to conserve the number of bridge domains (BD) consumed on the Cisco NCS 5500 ToR.

From the control plane point of view, it is similar to that of a single VXLAN architecture.

The multi-VXLAN EVPN based design optionally supports a static implementation of VXLAN technology through head-end replication (HER). HER helps leverage the VXLAN technology, regardless of the hardware/software limitation in the VXLAN feature set at the remote end of the VTEP tunnel.

With the static information defined in the setup_data, VPP performs the HER to all defined remote VTEPs and updates L2FIB (MAC-IP) table based on flood and learn. If EVPN co-exists with HER, Cisco VIM treats it as if two different sets of BGP speakers exist and provides information from each speaker in the same etcd FIB table.

Only drawback of this implementation is that VPP may peform unnecessary flooding. Cisco VIM uses EVPN as the primary mechanism and HER as the fallback methodology. You can add or remove HER to or from an existing EVPN pod through Cisco VIM reconfigure option.

From release CVIM 2.4.3 onwards, all the network traffic between the VM and VPP is over a vhost interface which is in memory and does not use a traditional kernel side interface, when VPP is used as the vSwitch in OpenStack. The network interface is no longer on the host and available within VM, to trace packets or capture them for debugging or other administrative purposes.

Underlying Architecture of the Port Mirroring Tool

Port mirroring works by setting up the following:

1. A span port on vpp to mirror the VirtualEthernet interface corresponding to the VMs vhost interface. This is a tap interface in VPP

   .

2. A tap device (tap0b20e7b0) on the compute host side is set as a kernel interface. A veth pair is created between the tap device on the VPP side (tapcli-0) and kernel side tap device (tap0b20e7b0) as shown in the below figure.

Note	This feature is released as a tech preview in Cisco VIM 3.4.1.

An important aspect of any Telco cloud is how the cloud is connected to the rest of the SP network. From release 3.4.1, Cisco VIM can attach to an existing Segment Routing (SR) Ethernet VPN (EVPN). This implementation is an evolution of the existing VPP-based standard VLAN and VXLAN EVPN designs. Cisco VIM attaches to the SR EVPN without an additional SDN controller, by peering with the route reflectors of the EVPN and the ToR with BGP Labeled Unicast (BGP-LU). There is no contention with any controller that manages the EVPN. MPLS is used as the data plane for the SR-labelled traffic.

Overview of Cisco VIM VPP Architecture

To connect a Cisco VIM pod to an SR EVPN, the existing VPP forwarding architecture has been enhanced. Networking-vpp is the Vector Packet Processing (VPP) based software accelerated virtual switch that is part of Cisco VIM. The architecture of networking-vpp is similar to a distributed SDN controller. However, it is not a separate controller and is integrated with Cisco VIM as a core component. Networking-vpp makes installation, updates and operational day-2 tasks a seamless experience.

Networking-vpp uses a publish and subscribe model to configure the virtual switches running on each compute node. The controllers translate any requests to the OpenStack API into the desired forwarding behavior of the VPP forwarders and publish that information into the distributed state database. The VPP agents that run on each compute node monitor the distributed state database and program according to the desired state that is published in the distributed state database.

The controllers do not have to:

• Push configuration directly to the compute nodes

• Validate the configuration

• Continuously monitor the configuration of the compute nodes.

This behavior makes networking-vpp extremely efficient, resilient, and scalable. When a compute node restarts, it only needs to look at the distributed state database to check how it forwards traffic to and from the virtual machines running on this compute node and configure itself. All this is transparent to the controllers.

The following steps explain the figure above:

1. The agent running on the controllers sees incoming requests and processes the API calls that relate to the state of the network data plane forwarding.

2. The agent translates the request into the desired state of the network data plane and publishes this state into a distributed state database.

3. The vpp-agents running on the compute nodes watch the distributed state database. When they see a change relevant to the compute node, they translate the desired state and update the VPP forwarding on the compute node.

Overview of Cisco VIM SR EVPN Architecture

To extend the networking-vpp architecture to support Segment Routing over MPLS, Cisco VIM 3.4.1 adds several additional components.

This image is not available in preview/cisco.com

For the control plane, a BGP process runs on each of the three controller nodes. The controllers peer to the EVPN route reflectors. This peering allows the Cisco VIM pod to exchange reachability information with the EVPN. Cisco VIM does an L2 stretch with only single-homed L2 routes. The peering with the route reflectors allows Cisco VIM to exchange Type-2 routes that have MAC to IP bindings. Type-3 route updates are exchanged to handle broadcast, unknown unicast and multicast (BUM) traffic. There are unidirectional MPLS tunnels in both directions.

In addition to the BGP peering with the EVPN route reflectors, the controllers have a BGP Labeled Unicast (BGP-LU) peering with the NCS ToRs to exchange label information.

This image is not available in preview/cisco.com

Cisco VIM does not process any SR policies coming from BGP or an SDN controller to program any explicit static or dynamic paths into VPP. BGP-LU and EVPN address family routes are processed by the BGP process running on each of the controllers. The BGP process publishes the SR policy into the distributed state database so that the VPP agents can configure VPP forwarding on the compute nodes to apply the MPLS label to the outbound traffic. The label between the compute node and the NCS ToR represents the path to node SID that identifies the VPP node. Bridge domain is identified by the Openstack Segmentation ID that is used for tenant isolation.

The data plane ports in the Cisco VIM pod are connected to the NCS with an L3 link and EVLAG provides link redundancy to the nodes. The VPP agents encapsulate the traffic leaving the compute node with an MPLS label representing the Segment Routing segment ID. The virtual machines are unaware that traffic will be sent over an SR-enabled network. For traffic arriving on the compute node from the NCS, the MPLS label represents the bridge domain, and the resulting L2 lookup determines the virtual machine to which to forward the traffic after stripping the MPLS header.

This image is not available in preview/cisco.com

The control plane ports on the Cisco VIM nodes have two uplinks, one to each NCS. From the node, the two ports are bundled in a port-channel using the linux teaming driver. ISIS runs on the NCS and provides link redundancy between the two ToRs using EVLAG. The control plane ports carry any or all of the following networks: storage network to access the CEPH cluster, API network hosting the OpenStack API, Management and provision networks used at install time to provision the nodes and during day-2 operations for management tasks.

This image is not available in preview/cisco.com

The VPP backed data plane ports are configured differently. Each NIC has a port connected to one NCS. This link is a routed link with an IP address both on the NIC port in the Cisco VIM node and an IP address on the NCS port. These addresses come from a /30 subnet. ISIS runs between the NCS and BGP-LU. On the Cisco VIM nodes, ECMP load balancing occurs over the two links to the NCS. On VPP, there is a loopback interface configured that is used as the SR node SID.

Ports C and D are SR-IOV ports that give the virtual machines direct access to the NIC through the VF interfaces. These are out of scope for the SR EVPN implementation on Cisco VIM.

Overview of Cisco VIM Installation for SR EVPN

Before You Begin

• A pair of NCS 5500 acts as the ToR. As Cisco VIM currently supports only a single pair of NCS, you must properly scale your NCS and choose the right NCS-5500 SKU so that all ports of all the nodes can be connected to this single pair of NCS. You must ensure that you include future growth in the port count when selecting the NCS model.

• You cannot enable both SR EVPN and VXLAN EVPN on the same Cisco VIM pod. Only one forwarding mechanism can exist within a single pod.

• Cisco VIM administrators must provide the SR network and subnet information for overlays, and the EVI or BD label. Cisco VIM does not manage the SR EVPN, it only attaches to an existing SR EVPN.

• You must provide the BGP configuration to the Cisco VIM installer during deployment.

• Cisco VIM 3.4.1 does not support traffic engineering policies for dynamic path selection through BGP SR extensions, or any explicit path selection through CLI or any interaction with an SDN controller.

• Only single homing L2 routes are supported. Only EVPN Type 2 and Type 3 are exchanged with the EVPN Route Reflectors (RR) in this release.

• You can enable this feature in Cisco VIM as a day-0 option using either setup_data.yaml file or Unified Management when you install the pod.

• You cannot convert an existing pod to this SR-enabled forwarding mechanism, even if it is set up to use VPP for VLAN or VXLAN based forwarding.

Note	Cisco VIM 3.4.1 does not support VMTP and NFVbench with SR EVPN.

You must update the setup_data.yaml file as shown below for the SR EVPN deployment:

SERVERS:
  control-server-1:
    bgp_speaker_addresses: {sr-mpls-tenant: <ip address from sr-mpls-tenant network but not part of the pool>}
    bgp_mgmt_addresses: {sr-mpls-tenant: <ip address from management network but not part of the pool>}  # <== optional, only when NETWORK_OPTIONS is sr-mpls network, for controller node only, needed when BGP peer is over L3
    vtep_ips: sr-mpls-tenant: <ip address from sr-mpls-tenant network and part of the pool>}  # <== needed for compute and control nodes
    sr_global_block: {prefix_sid_index: <int between 0 and 8000>, base: <int between 16000 and 1048576>}  # for sr-mpls, unique and must exist for all controllers/computes, sum of prefix_sid_index and base has to be unique across all servers
    dp_tor_info: {switch_a_hostname: ethx/y, switch_b_hostname: ethx/y} #For sr-mpls dp_tor_info is mandatory and has no po info

sr_global_block will exist for all computes and controllers. These are the segment identifiers used to associate with the Loopback IPs of the VPP nodes
tor_info will exist only for the control plane links. 
dp_tor_info will not be port channel. SR being a Layer 3 feature will rely on ECMP links for data plane redundancy.
vtep_ips is the loopback IP of the VPP which will be distributed by BGP LU for VPP reachability.
prefix sid is computed from sr_global_block section where base and index is defined.
NETWORK OPTIONS:
A new NETWORK OPTIONS section called sr-mpls has been introduced to support SR MPLS.
#  ######sr-mpls options #####
#  physnet_name: <unique_name> # Optional. Default is "physnet1"
#  enable_ecmp: <true or false>. # Optional (For SR, it is true. default is false), and only for vpp
#  ecmp_private_pool: <cidr> # Optional, Only if enable_ecmp is true.
#  sr-mpls:     # mutually exclusive to vxlan; will work with VPP and NCS-5500 only for now
#    sr-mpls-tenant:
#       physnet_name: <unique_name>
#       bgp_as_num: <int value between 1 and 4294967295> # unique across all AS
#       bgp_peers: ['ip1'] ---> list of length 1, Peer Route Reflector IPs
#       bgp_router_id: 'ip3' ---> The router ID to use for local GoBGP cluster

ecmp_private_pool is the range used to burn private IPs for the VPP uplinks to the aggregation layer.
enable_ecmp is used to enable ecmp links
sr-mpls-tenant is the key word to enable sr provider networking. 
bgp_as_num is the bgp as number of the remote PE device
bgp_peer is the bgp speaker loopback address
bgp_router_id will be the router id address used by the gobgp speakers


NETWORKING: 
A new networking section called sr-mpls-tenant has been introduced to support SR MPLS. 
NETWORKING:
  Networks: 
 -                 
    vlan_id: <vlan_id>
    subnet: <subnet in cidr>
    gateway: <gateway address>
    pool:
      - <a.b.c.d,e.f.g.h>
      - <i.j.k.l to m.n.o.p>
      - <q.r.s.t>

    segments:
      - sr-mpls-tenant

gateway is the BVI ip address to be used for the L3 gateway for CVIM control plane
pool is the address pool of the public addresses
segments will contain sr-mpls-tenant as a provider network
vlan_id is the id used to carve a sub interface for SR control plane from bonded samx CP links

Cisco VIM supports VM, baremetal, or container-based workloads. To support the container-based workloads, Cisco VIM hosts Cisco Container Platform as an application. The orchestrator creates a common OpenStack tenant and deploys the Cisco Container Platform control plane on it. The orchestrator can also create a tenant cluster if needed.

The Kubernetes clusters deployed are multi-master clusters with three master nodes and N worker nodes. The Cisco Container Platform control plane consists of three masters and three workers. The master and worker nodes run as VMs on OpenStack.

For more information on enabling Cisco Container Platform over Cisco VIM, see Container Support in Cisco VIM.

From Cisco VIM 3.4.1, GPU is supported in a passthrough mode. In this mode, the entire physical GPU card is available to the VM. Currently, Cisco VIM supports only Tesla T4 GPU cards.

For more information on enabling P-GPU, see Enabling P-GPU.

Cisco VIM supports Tap As a Server (TAAS) with Open vSwitch (OVS) as the mechanism driver. Tap-as-a-Service (TaaS) is an extension to the OpenStack network service (Neutron). It provides remote port mirroring for tenant virtual networks. Port mirroring involves sending a copy of all packets of one port to another port.

The TAAS implementation has two parts:

1. Tap-service: Represents the neutron port that receives the mirrored traffic (mirror destination). Tap service creation mirrors the VLAN allocation from an available pool and installs flow rules into the OVS tables.

2. Tap-flow: Represents the neutron port from which the traffic needs to be mirrored (mirror source). It determines the location from where traffic should be copied and the direction to which it should be forwarded.

Tap-flow creation initiates flow rules to copy the mirror source traffic to a VLAN associated with a tap-service. All mirrored traffic is redirected from br-int to br-tap with a specific mirror VLAN number (VLAN pool is preallocated in a configuration). Depending upon the mirror destination location, traffic is redirected back to the br-int (if the destination is on the same compute) or to br-tun (if the destination is on a different compute). If traffic is redirected to br-tun it sends traffic towards the compute node where the mirror destination is located using GRE encapsulation. br-tap is an intermediate point where specific mirrored traffic is filtered towards the mirror destination.

The figure below is a schematic of the layout for the TAAS networking diagram with OVS as the mechanism driver.

Following are the assumptions for this feature in CVIM:

1. Filtering of specific traffic is not supported.

2. TaaS is used for debugging and not for mirroring production traffic as it affects OVS performance.  

3. If ovs_vswitch container restarts or a compute node reloads, all tap services and tap flows have to be recreated.

4. Tap-flow and Tap-service can be created only for existing neutron ports, that is the corresponding VMs must be in running state. If the neutron ports point to a remote traffic destination, the remote ports will stay in down state.

5. Tap-flow with the same source mirror port cannot be used with different tap-services.

For information on setting up remote and local TaaS OVS, see TAAS Support.