Created on Aug 20, 2019
Introduction
This Reference Deployment Guide (RDG) describes the process of building a scalable Ethernet fabric with NVIDIA end-2-end solutions for deploying a Kubernetes (k8s) or an OpenShift cluster with RoCE acceleration enabled. RDMA/RoCE network technologies are essential to unlocking superior performance in Kubernetes or OpenShift clusters for machine learning (ML) and artificial intelligence (AI) workloads. Today, RDMA and RoCE are integrated into the mainstream code of popular AI/ML frameworks. Kubernetes pod's use of several additional high-performance network interfaces affects suitable network designs. One of way to scale-out a high-performance L2 network is to use a solution based on the Border Gateway Protocol (BGP) with EVPN and VXLAN overlays, which are used for network virtualization. VXLAN tunnels can satisfy this L2 adjacency requirement, and EVPN serves as a standard for scale-out L2 Ethernet fabrics.
VXLAN can virtualize the data center network, enabling layer 2 segments to be extended over an IP core (the underlay).
EVPN is the control plane for modern VXLAN deployments, allowing VTEPs to discover each other via EVPN and exchange reachability information such as MAC and IPs across racks.
This RDG is based on RDG: Scale-Out fabric deployment with VXLAN-BGP-EVPN over NVIDIA Onyx and describes the deployment process of an additional Ethernet fabric for high-performance POD network interfaces that are suitable for:
- Red Hat OpenShift Container Platform 4.1 and above
- Kubernetes cluster with version 1.10 and above
References
- RDG: RoCE accelerated K8s cluster deployment for ML and HPC workloads
- Multiple networks | Networking | OpenShift Container Platform 4.1
- NVIDIA Spectrum® Open Ethernet Switches
- LinkX® Ethernet Cables and Transceivers
- EVPN VXLAN: The Next Generation of Network Virtualization
- Ethernet VPN User manual for NVIDIA Onyx
- Border Gateway Protocol
Components Overview
NVIDIA Spectrum® Open Ethernet Switches
The NVIDIA Spectrum® switch family provides the most efficient network solution for the ever-increasing performance demands of Data Center applications. The Spectrum product family includes a broad portfolio of Top-of-Rack (TOR) and aggregation switches that range from 16 to 128 physical ports, with Ethernet data rates of 1GbE, 10GbE, 25GbE, 40GbE, 50GbE, 100GbE, 200GbE and 400GbE per port. Spectrum Ethernet switches are ideal to build cost-effective and scalable data center network fabrics that can scale from a few nodes to tens-of-thousands of nodes.
LinkX® Ethernet Cables and Transceivers
NVIDIA LinkX cables and transceivers make 100Gb/s deployments as easy and as universal as 10Gb/s links. Because NVIDIA offers one of industry’s broadest portfolio of 10, 25, 40, 50,100, 200 and 400Gb/s Direct Attach Copper cables (DACs), Copper Splitter cables, Active Optical Cables (AOCs) and Transceivers, every data center reach from 0.5m to 10km is supported. To maximize system performance. NVIDIA tests every product in an end-to-end environment assuring a Bit Error Rate of less than 1E-15. A BER of 1E-15 is 1000x better than many competitors.
EVPN VXLAN: The Next Generation of Network Virtualization
VXLAN (Virtual Extensible LAN) has become mainstream for network virtualization. EVPN (Ethernet Virtual Private Networks) VXLAN – using BGP as the control plane – is becoming the preferred way to virtualize data center networks without dedicated controllers. BGP EVPN with VXLAN provides greater network efficiency and scalability through enhanced support for both L2 and L3 services. It provides administrators with fine-grained policy control to help boost efficient utilization of network resources.
Compute Platform
This solution is designed for multi-homing servers with 4 x 100 GbE ports like DGX-1 or HPE Apollo 6500.
Solution Overview
The solution is comprised of NVIDIA Onyx and NVIDIA Open Ethernet Spectrum switches.
Only the Ethernet fabric configuration steps are detailed in this user guide. For Kubernetes cluster installation steps please refer to RDG: RoCE accelerated K8s cluster deployment for ML and HPC workloads.
Logical Design
The following figure shows a high-speed Ethernet network diagram with the L2 network topology that is required to deploy a Kubernetes cluster with RoCE acceleration enabled.
The diagram includes:
- CLOS based network topology with BGP underlay routing protocol
- Spine switch layer
- Workload rack with one or more Leaf switches with independent connectivity to Spine switch layer
- Service and storage rack with IPL connection between two Leaf switches that are connected to the Spine switch layer
The diagram does not contain the Kubernetes Management network and does not affect it.
Below is the logical design of the Leaf switch with four wired SRIOV enabled physical interfaces from the Worker node:
Note that the Worker Nodes (servers like DGX-1 or HPE Apollo 6500) must be connected to the leaf switches without any bond, LAG or MLAG interface connectivity options.
Below is the logical design of the Leaf switches configuration with Service node:
Setup
Below is the ethernet fabric deployment and in the service node setup specifications.
NVIDIA SN2000 and SN3000 series ethernet switches can be deployed in the above network design.
Physical Design and Wiring
Below is an example of the physical design for high-speed back-end setup. The design can be customized to your use case and requirements.
Using different switch models may require different types of cables.
Deployment Guide
Network / Fabric Deployment and Configuration
Prerequisites
This solution is based on NVIDIA Onyx switch operation system and supported only from NVIDIA Onyx version 3.8.1208 and above.
Below is a list of recommendations and prerequisites that are important for the configuration process:
- Refer to the NVIDIA Onyx User Manual to become familiar with switch software (located at http://www.mellanox.com/page/mlnx_onyx?mtag=onyx_software)
- Upgrade the switch to the latest NVIDIA Onyx version
Configuration
Before we start the configuration process, we recommend reading the Ethernet VPN User manual for NVIDIA Onyx to become familiar with EVPN deployment using NVIDIA Onyx.
Deployment configuration steps:
- Change the Leaf switch ports mode that are connected to the Worker nodes to switchport mode hybrid.
- Create an MLAG between the Leaf switches in the Service rack.
- Create VLANs on each Leaf switch and allow them on the hybrid ports.
- Create Loopback interface on each Leaf switch for HW VTEP.
- Choose and assign ASN for Spine switches( same ASN for ALL Spine switches).
- Choose and assign ASN for each Leaf switch(different for each Leaf switch).
- Configure BGP IPv4 underlay on Spine and Leaf switches.
- Advertise all Loopback network addresses on Leaf switches and MLAG tunnel IP address on both MLAG domain switches (in IPv4 address-family).
- Configure EVPN overlay on all switches.
- Configure NVE interface on each Leaf switch.
- Map VLANs to VNI's on all Leaf switches.
Example of our configuration files for the above physical solution design with the configuration steps implemented can be found here: https://github.com/Mellanox/roce_backend_at_scale.
for a Detailed .
Below are the switch CLI commands to configure the deployment of the spine switches, workload-rack leaf switches, and service-rack leaf switches (valid for NVIDIA Onyx version 3.8.1208).
The example shows mapping of four VLANs (111-114) to four VNIs (100111-100114):
Deployment Verification
The below commands can verify the deployment in the Spine switch, Workload rack Leaf switch and Service rack Leaf switch.
Spine switch deployment verification:
show ip bgp evpn summary - displays the BGP peers participating in the layer 2 EVPN address-family and their state:
VRF name : default BGP router identifier : 172.16.2.2 local AS number : 65005 BGP table version : 24702 Main routing table version: 24702 IPV4 Prefixes : 6 IPV6 Prefixes : 0 L2VPN EVPN Prefixes : 21 ------------------------------------------------------------------------------------------------------------------ Neighbor V AS MsgRcvd MsgSent TblVer InQ OutQ Up/Down State/PfxRcd ------------------------------------------------------------------------------------------------------------------ 172.16.2.1 4 65001 80877 80910 24702 0 0 39:03:43:05 ESTABLISHED/4 172.16.4.1 4 65002 80826 80867 24702 0 0 39:03:43:02 ESTABLISHED/7 172.16.5.1 4 65003 78009 80864 24702 0 0 39:03:42:58 ESTABLISHED/5 172.16.7.1 4 65004 78073 80851 24702 0 0 39:03:42:55 ESTABLISHED/5
Leaf switch deployment verification:
Workload rack Leaf switch:
show interface nve 1 - displays the configured VTEP on a network device participating in BGP EVPN:
Interface NVE 1 status: Admin state: up Source interface: loopback 1 (ip 172.16.100.100) Controller mode: BGP Mlag tunnel IP: (not configured) Effective tunnel IP: 172.16.100.100 Global Neigh-Suppression: Enable NVE member interfaces: (not configured) Counters 12266 encapsulated (Tx) NVE packets 29046 decapsulated (Rx) NVE packets 0 dropped NVE-encapsulated packets 0 NVE-encapsulated packets with errors
show interface nve 1 detail - displays the configured VNIs on a network device participating in BGP EVPN:
---------------------------------------------- Vlan VNI Neigh Suppression ---------------------------------------------- 111 100111 Enable 112 100112 Enable 113 100113 Enable 114 100114 Enable
show ip bgp evpn summary:
VRF name : default BGP router identifier : 172.16.1.1 local AS number : 65001 BGP table version : 35287 Main routing table version: 35287 IPV4 Prefixes : 8 IPV6 Prefixes : 0 L2VPN EVPN Prefixes : 24 ------------------------------------------------------------------------------------------------------------------ Neighbor V AS MsgRcvd MsgSent TblVer InQ OutQ Up/Down State/PfxRcd ------------------------------------------------------------------------------------------------------------------ 172.16.1.2 4 65005 80848 80822 35287 0 0 39:03:42:25 ESTABLISHED/12 172.16.2.2 4 65005 80874 80834 35287 0 0 39:03:43:33 ESTABLISHED/12
Service rack Leaf switch
In the following command output "Mlag tunnel IP" must contain a value/parameter. The rest of the commands have the same output.
show interface nve 1 :
Interface NVE 1 status: Admin state: up Source interface: loopback 1 (ip 172.16.102.102) Controller mode: BGP Mlag tunnel IP: 172.16.104.104 Effective tunnel IP: 172.16.104.104 Global Neigh-Suppression: Enable NVE member interfaces: (not configured) Counters 41283 encapsulated (Tx) NVE packets 0 decapsulated (Rx) NVE packets 0 dropped NVE-encapsulated packets 0 NVE-encapsulated packets with errors
Service Host Deployment and Configuration
We selected the Service Node to provide IP addresses for high-performance network adapters via DHCP service that was deployed in a docker container. This server has been deployed with Ubuntu 18.04 OS. For DHCP service we are using the docker image storytel/dnsmasq with the dnsmasq application.
Service Node deployment steps:
- Install the OS
- Install the required packages
- Configure the network interfaces
- Install Docker container
- Create a dnsmasq configuration file - dnsmasq.conf
- Run the docker image with dnsmasq application
Required packages:
# apt-get -y install ifenslave vlan
Interfaces configuration file:
Install Docker:
# apt-get install docker.io
Configuration file for dnsmasq application - /etc/dnsmasq/dnsmasq.conf:
keep-in-foreground log-queries log-dhcp interface=bond0.111,bond0.112,bond0.113,bond0.114 bind-interfaces port=0 bogus-priv dhcp-range=bond0.111,10.10.111.1,10.10.111.253,255.255.255.0,60m dhcp-range=bond0.112,10.10.112.1,10.10.112.253,255.255.255.0,60m dhcp-range=bond0.113,10.10.113.1,10.10.113.253,255.255.255.0,60m dhcp-range=bond0.114,10.10.114.1,10.10.114.253,255.255.255.0,60m dhcp-option=bond0.111,3 dhcp-option=bond0.112,3 dhcp-option=bond0.113,3 dhcp-option=bond0.114,3
Run docker container with customized dnsmasq application configuration file.
docker run --name dnsmasq --cap-add=NET_ADMIN --net=host -v /etc/dnsmasq:/etc/dnsmasq storytel/dnsmasq
Done!
About the Authors
| Vitaliy Razinkov Over the past few years, Vitaliy Razinkov has been working as a Solutions Architect on the NVIDIA Networking team, responsible for complex Kubernetes/OpenShift and Microsoft's leading solutions, research and design. He previously spent more than 25 years in senior positions at several companies. Vitaliy has written several reference designs guides on Microsoft technologies, RoCE/RDMA accelerated machine learning in Kubernetes/OpenShift, and container solutions, all of which are available on the NVIDIA Networking Documentation website. |
Related Documents
