Understanding Ethernet and Infiniband on OpenShift

I recently was involved in a conversation around using only infiniband on an OpenShift cluster installation.  That is the customer wanted to only have infiniband connectivity for both the cluster apis and the high speed storage access requirements for the application.   This interaction made me realize we probably need a refresher on the difference between infiniband and ethernet because they are not the same nor can they be swapped interchangeably.

The difference between infiniband and ethernet is very disparate from a design point of view.  Infiniband was designed with the idea of having a high reliability, high bandwidth and low latency to interconnect nodes in a supercomputer cluster.  Whereas ethernet was designed with the idea of how can I move data between multiple systems easily.  This difference becomes more apparent in how each technology is designed to move data.

The design differences show up for example in how latency is handled between the two types of interconnects.  For ethernet interconnects typically a store-and-forward along with MAC address network transport model is used for communication between hosts.  This method increases the process flow because it has to take into account complex services like IP, MPLS and 802.1Q.   Whereas with infiniband layer 2 processing uses a 16 bit LID(local ID) number which is the only one that can be used to search for the forwarding path information.  Further the switching technology in infiniband uses a cut-through approach which reduces the forwarding delay making it significantly faster than ethernet.

Another difference also shows up in network reliability.  The infiniband protocol is a complete network protocol with its own defined layers from layer 1 to layer 4.  This end to end flow control provides the basis for infiniband's network packeting sending and receiving which can provide a lossless network.  Ethernet on the other hand does not have a scheduling based flow control mechanism which results in the lack of a guarantee whether a node on the other end will end up being congested upon packet receipt.  This is why ethernet switches are built with a cache to absorb these sudden bursts of traffic.

Networking mode or methods is another distinction between these two technologies.  An software defined network is built into infiniband by design.  There is a subnet manager present on each layer 2 infiniband network to configure the LID of the nodes.  The subnet manager also calculates the forwarding path through the control plane and issues it to the infiniband exchange.  Conversely ethernet uses a networking mode that generates MAC addresses and the IP protocol must cooperate with the ARP protocol.  Nodes in the ethernet network are required to send packets on a regular basis to guarantee entries, in a ARP table for example, are updated in real time.  All of this leads to more overhead in a ethernet network compared to infiniband.

We can see from the above there are significant differences between the two technologies which makes it impossible to swap them out like for like as in the case of our OpenShift installation request from the customer.  For example take the OpenShift installation which will leverage OVN/OVS for networking.  During the installation there is an expectation that the MAC address will exist on the interface marked for the cluster api.   However in an infiniband network there is no MAC address concept.  Once might see an error similar to the below:

Error: failed to modify 802-3-ethernet.cloned-mac-address:
'00:00:01:49:fe:80:00:00:00:00:00:00:00:11:22:33:01:32:02:00' is not a valid Ethernet MAC.

Further, drivers also become an issue for networking devices like a Mellonox CX-7 or BlueField-3.  This is because the default mlx upstream drivers that ship with Red Hat CoreOS in OpenShift do not contain the RDMA component which is required for infiniband.  To obtain the the RDMA component one needs to leverage the NVIDIA DOCA driver which is part of the NVIDIA network operator.  However this operator cannot be leveraged in an OpenShift day 0 installation.  Even if it could though again the expectation of OVS/OVN networking is to have a MAC address to work with from an ethernet network.

Given all these differences we had to explore how could we meet the customers needs but still apply the correct technology to the systems.   If the customers goal was to ensure a high speed interconnect between the nodes in the cluster we can still do this with OpenShift.  However we need to approach it differently and also break out the cluster apis so they are still working with an ethernet network.   A suitable approach might look like this example below.

In the diagram we have a six node OpenShift cluster each with two single port Mellanox CX-7 cards.  For each node we have one card plugged into an ethernet switch and another plugged into an infiniband switch.   With this design we can now install OpenShift using the one CX-7 card operating in ethernet mode.   Once OpenShift is installed we can then layer on the NVIDIA network operator to provide the RDMA infiniband driver and leverage the second CX-7 card operating in infiniband mode.   This design enables us to not only get OpenShift installed but still provide a secondary network to our workloads with access to the high speed infiniband network.   This same design would also work if we had just one dual ported CX-7 card as we can use the Mellanox tools to configure one port for ethernet and one for infiniband.

Hopefully this blog provided some insight into the difference between infiniband and ethernet and why one simply cannot swap out ethernet for infiniband on an OpenShift installation.    

Build RDMA GPU-Tools Container

 

The purpose of this blog is to build a container that automates building the testing tooling for validating RDMA connectivity and performance when used in conjunction with NVIDIA Network Operator and NVIDIA GPU Operator.  Specifically I want to be able to use the ib_write_bw command with the --use_cuda switch to demonstrate RDMA from one GPU in a node to another GPU in another node in an OpenShift cluster. The ib_write_bw command is part of the perftest suite which is a collection of tests written over uverbs intended for use as a performance micro-benchmark. The tests may be used for HW or SW tuning as well as for functional testing.

The collection contains a set of bandwidth and latency benchmark such as:

• Send - ib_send_bw and ib_send_lat
• RDMA Read - ib_read_bw and ib_read_lat
• RDMA Write - ib_write_bw and ib_write_lat
• RDMA Atomic - ib_atomic_bw and ib_atomic_lat
• Native Ethernet (when working with MOFED2) - raw_ethernet_bw, raw_ethernet_lat

In previous blogs, here and here,  I used a Fedora 35 container and manually added the components I wanted but here we will provide the tooling to build a container that will instantiate itself upon deployment. The workflow is as follows:

• Dockerfile.tools - which provides the content for the base image and the entrypoint.sh script.
• Entrypoint.sh - which provides the start up script for the container to pull in both the NVIDIA cuda libraries and also build and deploy the perftest suite with the cuda option available.
• Additional RPMs - there are some packages that were not part of the UBI image repo but are dependencies for CUDA toolkit.

The first thing we need to do is create a working directory for our files and an rpms directory for the rpms we will need for our base image. I am using root here but it could be a regular user as well.

$ mkdir -p /root/gpu-tools/rpms
$ cd /root/gpu-tools

Next we need to download the following rpms from Red Hat Package Downloads and place them into the rpms directory.

• infiniband-diags-51.0-1.el9.x86_64.rpm
• libglvnd-opengl-1.3.4-1.el9.x86_64.rpm
• libibumad-51.0-1.el9.x86_64.rpm
• librdmacm-51.0-1.el9.x86_64.rpm
• libxcb-1.13.1-9.el9.x86_64.rpm
• libxcb-devel-1.13.1-9.el9.x86_64.rpm
• libxkbcommon-1.0.3-4.el9.x86_64.rpm
• libxkbcommon-x11-1.0.3-4.el9.x86_64.rpm
• pciutils-devel-3.7.0-5.el9.x86_64.rpm
• rdma-core-devel-51.0-1.el9.x86_64.rpm
• xcb-util-0.4.0-19.el9.x86_64.rpm
• xcb-util-image-0.4.0-19.el9.x86_64.rpm
• xcb-util-keysyms-0.4.0-17.el9.x86_64.rpm
• xcb-util-renderutil-0.3.9-20.el9.x86_64.rpm
• xcb-util-wm-0.4.1-22.el9.x86_64.rpm

Once we have all our rpms for the base image we can move onto creating the dockerfile.tools file which we will use to build our image.

$ cat <<EOF >dockerfile.tools
# Start from UBI9 image
FROM registry.access.redhat.com/ubi9/ubi:latest

# Set work directory
WORKDIR /root
RUN mkdir /root/rpms
COPY ./rpms/*.rpm /root/rpms/

# DNF install packages either from repo or locally
RUN dnf install `ls -1 /root/rpms/*.rpm` -y
RUN dnf install wget procps-ng pciutils jq iputils ethtool net-tools git autoconf automake libtool -y

# Cleanup 
WORKDIR /root
RUN dnf clean all

# Run container entrypoint
COPY entrypoint.sh /root/entrypoint.sh
RUN chmod +x /root/entrypoint.sh

ENTRYPOINT ["/root/entrypoint.sh"]
EOF

We also need to create the entrypoint.sh script which is referenced in the dockerfile and does the heavy lifting of pulling in the cuda toolkit and the perftest suite.

$ cat <<EOF > entrypoint.sh 
#!/bin/bash
# Set working dir
cd /root

# Configure and install cuda-toolkit
dnf config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel9/x86_64/cuda-rhel9.repo
dnf clean all
dnf -y install cuda-toolkit-12-6

# Export CUDA library paths
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
export LIBRARY_PATH=/usr/local/cuda/lib64:$LIBRARY_PATH

# Git clone perftest repository
git clone https://github.com/linux-rdma/perftest.git

# Change into perftest directory
cd /root/perftest

# Build perftest with the cuda libraries included
./autogen.sh
./configure CUDA_H_PATH=/usr/local/cuda/include/cuda.h
make -j
make install

# Sleep container indefinitly
sleep infinity & wait
EOF

Next we can use the dockerfile we just created to build the base image.

$ podman build -f dockerfile.tools -t quay.io/redhat_emp1/ecosys-nvidia/gpu-tools:0.0.2
STEP 1/10: FROM registry.access.redhat.com/ubi9/ubi:latest
STEP 2/10: WORKDIR /root
--> Using cache 75f163f12503272b83e1137f7c1903520f84493ffe5aec0ef32ece722bd0d815
--> 75f163f12503
STEP 3/10: RUN mkdir /root/rpms
--> Using cache ade32aa6605847a8b3f5c8b68cfcb64854dc01eece34868faab46137a60f946c
--> ade32aa66058
STEP 4/10: COPY ./rpms/*.rpm /root/rpms/
--> Using cache 59dcef81d6675f44d22900f13a3e5441f5073555d7d2faa0b2f261f32e4ba6cd
--> 59dcef81d667
STEP 5/10: RUN dnf install `ls -1 /root/rpms/*.rpm` -y
--> Using cache ebb8b3150056240378ac36f7aa41d7f13b13308e9353513f26a8d3d70e618e3b
--> ebb8b3150056
STEP 6/10: RUN dnf install wget procps-ng pciutils jq iputils ethtool net-tools git autoconf automake libtool -y
--> Using cache 5ca85080c103ba559994906ada0417102f54f22c182bbc3a06913109855278cc
--> 5ca85080c103
STEP 7/10: WORKDIR /root
--> Using cache 68c8cd47a41bc364a0da5790c90f9aee5f8a8c7807732f3a5138bff795834fc1
--> 68c8cd47a41b
STEP 8/10: RUN dnf clean all
Updating Subscription Management repositories.
Unable to read consumer identity

This system is not registered with an entitlement server. You can use subscription-manager to register.

26 files removed
--> a219fec5df49
STEP 9/10: COPY entrypoint.sh /root/entrypoint.sh
--> aeb03bf74673
STEP 10/10: ENTRYPOINT ["/bin/bash", "/root/entrypoint.sh"]
COMMIT quay.io/redhat_emp1/ecosys-nvidia/gpu-tools:0.0.2
--> 45c2113e5082
Successfully tagged quay.io/redhat_emp1/ecosys-nvidia/gpu-tools:0.0.2
45c2113e5082fb2f548b9e1b16c17524184c4079e2db77399519cf29829af1e7

Once the image is created we can push it to our favorite registry.

$ podman push quay.io/redhat_emp1/ecosys-nvidia/gpu-tools:0.0.2
Getting image source signatures
Copying blob 62ee1c6c02d5 done   | 
Copying blob 6027214db22e done   | 
Copying blob 4822ebd5a418 done   | 
Copying blob 422a0e40f90b done   | 
Copying blob 5916e2b21ab2 done   | 
Copying blob 10bf375a4d78 done   | 
Copying blob ca1c18e183d5 done   | 
Copying config 3bbb6e1f9b done   | 
Writing manifest to image destination

Now that we have an image let's test it out on the cluster where we have compatible RDMA hardware configured. I am using the same setup as I used in a previous blog so I am going to skip the details about setting up a service account and providing the privileges to it. We will however create our workload pod yaml files which we will use to deploy the image.

cat >>EOF >rdma-32-workload.yaml
apiVersion: v1
kind: Pod
metadata:
  name: rdma-eth-32-workload
  namespace: default
  annotations:
    k8s.v1.cni.cncf.io/networks: rdmashared-net
spec:
  nodeSelector: 
    kubernetes.io/hostname: nvd-srv-32.nvidia.eng.rdu2.dc.redhat.com
  serviceAccountName: rdma
  containers:
  - image: quay.io/redhat_emp1/ecosys-nvidia/gpu-tools:0.0.2
    name: rdma-32-workload
    securityContext:
      privileged: true
      capabilities:
        add: [ "IPC_LOCK" ]
    resources:
      limits:
        nvidia.com/gpu: 1
        rdma/rdma_shared_device_eth: 1
      requests:
        nvidia.com/gpu: 1
        rdma/rdma_shared_device_eth: 1
EOF

$ cat >>EOF >rdma-33-workload.yaml
apiVersion: v1
kind: Pod
metadata:
  name: rdma-eth-33-workload
  namespace: default
  annotations:
    k8s.v1.cni.cncf.io/networks: rdmashared-net
spec:
  nodeSelector: 
    kubernetes.io/hostname: nvd-srv-33.nvidia.eng.rdu2.dc.redhat.com
  serviceAccountName: rdma
  containers:
  - image: quay.io/redhat_emp1/ecosys-nvidia/gpu-tools:0.0.2
    name: rdma-33-workload
    securityContext:
      privileged: true
      capabilities:
        add: [ "IPC_LOCK" ]
    resources:
      limits:
        nvidia.com/gpu: 1
        rdma/rdma_shared_device_eth: 1
      requests:
        nvidia.com/gpu: 1
        rdma/rdma_shared_device_eth: 1
EOF

Next we can deploy the containers.

$ oc create -f rdma-32-workload.yaml 
pod/rdma-eth-32-workload created

$ oc create -f rdma-33-workload.yaml 
pod/rdma-eth-33-workload created

Validate the pods are up and running.

$ oc get pods
NAME                   READY   STATUS    RESTARTS   AGE
rdma-eth-32-workload   1/1     Running   0          51s
rdma-eth-33-workload   1/1     Running   0          47s

Now open two terminals and rsh into each pod in one of the terminals and validate that the perftest commands are present. We can also get the ipaddress of our pod inside the containers.

$ oc rsh rdma-eth-32-workload
sh-5.1# ib
ib_atomic_bw         ib_read_lat          ib_write_bw          ibcacheedit          ibfindnodesusing.pl  iblinkinfo           ibping               ibroute              ibstatus             ibtracert            
ib_atomic_lat        ib_send_bw           ib_write_lat         ibccconfig           ibhosts              ibnetdiscover        ibportstate          ibrouters            ibswitches           
ib_read_bw           ib_send_lat          ibaddr               ibccquery            ibidsverify.pl       ibnodes              ibqueryerrors        ibstat               ibsysstat            
sh-5.1# ip a
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN group default qlen 1000
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
    inet 127.0.0.1/8 scope host lo
       valid_lft forever preferred_lft forever
    inet6 ::1/128 scope host 
       valid_lft forever preferred_lft forever
2: eth0@if96: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1400 qdisc noqueue state UP group default 
    link/ether 0a:58:0a:83:00:34 brd ff:ff:ff:ff:ff:ff link-netnsid 0
    inet 10.131.0.52/23 brd 10.131.1.255 scope global eth0
       valid_lft forever preferred_lft forever
    inet6 fe80::858:aff:fe83:34/64 scope link 
       valid_lft forever preferred_lft forever
3: net1@if78: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default 
    link/ether 32:1a:83:4a:e2:39 brd ff:ff:ff:ff:ff:ff link-netnsid 0
    inet 192.168.2.1/24 brd 192.168.2.255 scope global net1
       valid_lft forever preferred_lft forever
    inet6 fe80::301a:83ff:fe4a:e239/64 scope link 
       valid_lft forever preferred_lft forever

$ oc rsh rdma-eth-33-workload
sh-5.1# ib
ib_atomic_bw         ib_read_lat          ib_write_bw          ibcacheedit          ibfindnodesusing.pl  iblinkinfo           ibping               ibroute              ibstatus             ibtracert            
ib_atomic_lat        ib_send_bw           ib_write_lat         ibccconfig           ibhosts              ibnetdiscover        ibportstate          ibrouters            ibswitches           
ib_read_bw           ib_send_lat          ibaddr               ibccquery            ibidsverify.pl       ibnodes              ibqueryerrors        ibstat               ibsysstat            
sh-5.1# ip a
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN group default qlen 1000
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
    inet 127.0.0.1/8 scope host lo
       valid_lft forever preferred_lft forever
    inet6 ::1/128 scope host 
       valid_lft forever preferred_lft forever
2: eth0@if105: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1400 qdisc noqueue state UP group default 
    link/ether 0a:58:0a:80:02:3d brd ff:ff:ff:ff:ff:ff link-netnsid 0
    inet 10.128.2.61/23 brd 10.128.3.255 scope global eth0
       valid_lft forever preferred_lft forever
    inet6 fe80::858:aff:fe80:23d/64 scope link 
       valid_lft forever preferred_lft forever
3: net1@if82: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default 
    link/ether 22:3e:02:c9:d0:87 brd ff:ff:ff:ff:ff:ff link-netnsid 0
    inet 192.168.2.2/24 brd 192.168.2.255 scope global net1
       valid_lft forever preferred_lft forever
    inet6 fe80::203e:2ff:fec9:d087/64 scope link 
       valid_lft forever preferred_lft forever

Now let's run the RDMA perftest with the --use_cuda switch. Again we will need to have two rsh sessions one on each pod. In the first terminal we can run the following.

sh-5.1# ib_write_bw -R -T 41 -s 65536 -F -x 3 -m 4096 --report_gbits -q 16 -D 60  -d mlx5_1 -p 10000 --source_ip 192.168.2.1 --use_cuda=0
 WARNING: BW peak won't be measured in this run.
Perftest doesn't supports CUDA tests with inline messages: inline size set to 0

************************************
* Waiting for client to connect... *
************************************
~

In the second terminal we will run the following command which will dump the output.

sh-5.1# ib_write_bw -R -T 41 -s 65536 -F -x 3 -m 4096 --report_gbits -q 16 -D 60  -d mlx5_1 -p 10000 --source_ip 192.168.2.2 --use_cuda=0 192.168.2.1
 WARNING: BW peak won't be measured in this run.
Perftest doesn't supports CUDA tests with inline messages: inline size set to 0
Requested mtu is higher than active mtu 
Changing to active mtu - 3
initializing CUDA
Listing all CUDA devices in system:
CUDA device 0: PCIe address is E1:00

Picking device No. 0
[pid = 4101, dev = 0] device name = [NVIDIA A40]
creating CUDA Ctx
making it the current CUDA Ctx
CUDA device integrated: 0
cuMemAlloc() of a 2097152 bytes GPU buffer
allocated GPU buffer address at 00007f3dfa600000 pointer=0x7f3dfa600000
---------------------------------------------------------------------------------------
                    RDMA_Write BW Test
 Dual-port       : OFF        Device         : mlx5_1
 Number of qps   : 16        Transport type : IB
 Connection type : RC        Using SRQ      : OFF
 PCIe relax order: ON        Lock-free      : OFF
 ibv_wr* API     : ON        Using DDP      : OFF
 TX depth        : 128
 CQ Moderation   : 1
 Mtu             : 1024[B]
 Link type       : Ethernet
 GID index       : 3
 Max inline data : 0[B]
 rdma_cm QPs     : ON
 Data ex. method : rdma_cm     TOS    : 41
---------------------------------------------------------------------------------------
 local address: LID 0000 QPN 0x00c6 PSN 0x2986aa
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00c7 PSN 0xa0ef83
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00c8 PSN 0x74badb
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00c9 PSN 0x287d57
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00ca PSN 0xf5b155
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00cb PSN 0x6cc15d
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00cc PSN 0x3730c2
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00cd PSN 0x74d75d
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00ce PSN 0x51a707
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00cf PSN 0x987246
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00d0 PSN 0xa334a8
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00d1 PSN 0x5d8f52
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00d2 PSN 0xc42ca0
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00d3 PSN 0xf43696
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00d4 PSN 0x43f9d2
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 local address: LID 0000 QPN 0x00d5 PSN 0xbc4d64
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00c6 PSN 0xb1023e
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00c7 PSN 0xc78587
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00c8 PSN 0x5a328f
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00c9 PSN 0x582cfb
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00cb PSN 0x40d229
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00cc PSN 0x5833a1
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00cd PSN 0xcfefb6
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00ce PSN 0xfd5d41
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00cf PSN 0xed811b
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00d0 PSN 0x5244ca
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00d1 PSN 0x946edc
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00d2 PSN 0x4e0f76
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00d3 PSN 0x7b13f4
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00d4 PSN 0x1a2d5a
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00d5 PSN 0xd22346
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00d6 PSN 0x722bc8
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
---------------------------------------------------------------------------------------
 #bytes     #iterations    BW peak[Gb/sec]    BW average[Gb/sec]   MsgRate[Mpps]
 65536      10384867         0.00               181.46              0.346100
---------------------------------------------------------------------------------------
deallocating GPU buffer 00007f3dfa600000
destroying current CUDA Ctx

And if we return to the first terminal we should see it updated with the same output.

sh-5.1# ib_write_bw -R -T 41 -s 65536 -F -x 3 -m 4096 --report_gbits -q 16 -D 60  -d mlx5_1 -p 10000 --source_ip 192.168.2.1 --use_cuda=0
 WARNING: BW peak won't be measured in this run.
Perftest doesn't supports CUDA tests with inline messages: inline size set to 0

************************************
* Waiting for client to connect... *
************************************
Requested mtu is higher than active mtu 
Changing to active mtu - 3
initializing CUDA
Listing all CUDA devices in system:
CUDA device 0: PCIe address is 61:00

Picking device No. 0
[pid = 4109, dev = 0] device name = [NVIDIA A40]
creating CUDA Ctx
making it the current CUDA Ctx
CUDA device integrated: 0
cuMemAlloc() of a 2097152 bytes GPU buffer
allocated GPU buffer address at 00007f8bca600000 pointer=0x7f8bca600000
---------------------------------------------------------------------------------------
                    RDMA_Write BW Test
 Dual-port       : OFF        Device         : mlx5_1
 Number of qps   : 16        Transport type : IB
 Connection type : RC        Using SRQ      : OFF
 PCIe relax order: ON        Lock-free      : OFF
 ibv_wr* API     : ON        Using DDP      : OFF
 CQ Moderation   : 1
 Mtu             : 1024[B]
 Link type       : Ethernet
 GID index       : 3
 Max inline data : 0[B]
 rdma_cm QPs     : ON
 Data ex. method : rdma_cm     TOS    : 41
---------------------------------------------------------------------------------------
 Waiting for client rdma_cm QP to connect
 Please run the same command with the IB/RoCE interface IP
---------------------------------------------------------------------------------------
 local address: LID 0000 QPN 0x00c6 PSN 0xb1023e
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00c7 PSN 0xc78587
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00c8 PSN 0x5a328f
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00c9 PSN 0x582cfb
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00cb PSN 0x40d229
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00cc PSN 0x5833a1
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00cd PSN 0xcfefb6
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00ce PSN 0xfd5d41
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00cf PSN 0xed811b
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00d0 PSN 0x5244ca
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00d1 PSN 0x946edc
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00d2 PSN 0x4e0f76
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00d3 PSN 0x7b13f4
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00d4 PSN 0x1a2d5a
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00d5 PSN 0xd22346
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 local address: LID 0000 QPN 0x00d6 PSN 0x722bc8
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:32
 remote address: LID 0000 QPN 0x00c6 PSN 0x2986aa
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00c7 PSN 0xa0ef83
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00c8 PSN 0x74badb
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00c9 PSN 0x287d57
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00ca PSN 0xf5b155
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00cb PSN 0x6cc15d
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00cc PSN 0x3730c2
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00cd PSN 0x74d75d
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00ce PSN 0x51a707
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00cf PSN 0x987246
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00d0 PSN 0xa334a8
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00d1 PSN 0x5d8f52
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00d2 PSN 0xc42ca0
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00d3 PSN 0xf43696
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00d4 PSN 0x43f9d2
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
 remote address: LID 0000 QPN 0x00d5 PSN 0xbc4d64
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:10:06:145:33
---------------------------------------------------------------------------------------
 #bytes     #iterations    BW peak[Gb/sec]    BW average[Gb/sec]   MsgRate[Mpps]
 65536      10384867         0.00               181.46              0.346100
---------------------------------------------------------------------------------------
deallocating GPU buffer 00007f8bca600000
destroying current CUDA Ctx

Hopefully this helped demonstrate a much cleaner and automated way to build a perftest container with cuda enabled to perform RDMA testing on OpenShift with NVIDIA Network Operator and NVIDIA GPU Operator.

RDMA+CUDA with NVIDIA on OpenShift

In a previous blog I described how to configure an OpenShift cluster with RDMA when using the NVIDIA Network Operator and NVIDIA GPU Operator.  However in that blog we only did simple RDMA testing across the network interfaces with no involvement of the GPU.   In this blog I will show the testing so it does involve the GPU and CUDA libraries.   Keep in mind though this testing is for validating that the configuration is setup correctly and should not replace real world workload testing of an application.

In this example we are using the same versions of OpenShift and the operators as in the previous blog so I will not go into those details here.  What we will capture below is how to configure the container appropriately to do the RDMA+CUDA testing.

The first thing we need to do is create a ServiceAccount in the default namespace. We can do so by generating the custom resource file below and creating on the cluster.

$ cat <<EOF > default-serviceaccount.yaml
apiVersion: v1
kind: ServiceAccount
metadata:
  name: rdma
  namespace: default
EOF

$ oc create -f default-serviceaccount.yaml
serviceaccount/rdma created

Now that the rdma account is created let's give it privileged access.

oc -n default adm policy add-scc-to-user privileged -z rdma
clusterrole.rbac.authorization.k8s.io/system:openshift:scc:privileged added: "rdma"

Next we will generate two pod custom resource files to run our workload pod image on the two baremetal a100 nodes in our environment.

$ cat <<EOF > rdma-eth-a100-01-workload.yaml
apiVersion: v1
kind: Pod
metadata:
  name: rdma-eth-a100-01-workload
  namespace: default
  annotations:
    k8s.v1.cni.cncf.io/networks: rdmashared-net
spec:
  nodeSelector: 
    kubernetes.io/hostname: a100-1.private.openshiftvcn.schmaustech.com
  serviceAccountName: rdma
  containers:
  - image: quay.io/redhat_emp1/ecosys-nvidia/gpu-operator:tools
    name: rdma-eth-a100-01-workload
    command:
      - sh
      - -c
      - sleep inf
    securityContext:
      privileged: true
      capabilities:
        add: [ "IPC_LOCK" ]
    resources:
      limits:
        nvidia.com/gpu: 1
        rdma/rdma_shared_device_eth: 1
      requests:
        nvidia.com/gpu: 1
        rdma/rdma_shared_device_eth: 1
EOF

$ cat <<EOF > rdma-eth-a100-02-workload.yaml
apiVersion: v1
kind: Pod
metadata:
  name: rdma-eth-a100-02-workload
  namespace: default
  annotations:
    k8s.v1.cni.cncf.io/networks: rdmashared-net
spec:
  nodeSelector: 
    kubernetes.io/hostname: a100-2.private.openshiftvcn.schmaustech.com
  serviceAccountName: rdma
  containers:
  - image: quay.io/redhat_emp1/ecosys-nvidia/gpu-operator:tools
    name: rdma-eth-a100-02-workload
    command:
      - sh
      - -c
      - sleep inf
    securityContext:
      privileged: true
      capabilities:
        add: [ "IPC_LOCK" ]
    resources:
      limits:
        nvidia.com/gpu: 1
        rdma/rdma_shared_device_eth: 1
      requests:
        nvidia.com/gpu: 1
        rdma/rdma_shared_device_eth: 1
EOF

With the pod files generated we can create them on the cluster.

$ oc create -f oci-agent-pod-a100-01.yaml
pod/oci-agent-pod-a100-01 created

$ oc create -f oci-agent-pod-a100-02.yaml
pod/oci-agent-pod-a100-02 created

Validate that the pods are running.

$ oc get pods
NAME                        READY   STATUS    RESTARTS   AGE
rdma-eth-a100-01-workload   1/1     Running   0          1m
rdma-eth-a100-02-workload   1/1     Running   0          1m

Next we can rsh into each of them in separate terminal windows.

$ oc rsh rdma-eth-a100-01-workload
sh-5.1# cd /root
sh-5.1#

$ oc rsh rdma-eth-a100-02-workload
sh-5.1# cd /root
sh-5.1# 

Building RDMA Validation Tests

The next steps are required on both running pods and enable perftest to have CUDA capable binaries.

First we need to download the CUDA repo and since our image is Fedora 35 based we will pulling down a Fedora 35 based package with wget. Note one might have to install wget.

sh-5.1# wget https://developer.download.nvidia.com/compute/cuda/11.7.0/local_installers/cuda-repo-fedora35-11-7-local-11.7.0_515.43.04-1.x86_64.rpm
--2024-11-20 16:06:08--  https://developer.download.nvidia.com/compute/cuda/11.7.0/local_installers/cuda-repo-fedora35-11-7-local-11.7.0_515.43.04-1.x86_64.rpm
Resolving developer.download.nvidia.com (developer.download.nvidia.com)... 152.199.20.126
Connecting to developer.download.nvidia.com (developer.download.nvidia.com)|152.199.20.126|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 3795608809 (3.5G) [application/x-rpm]
Saving to: 'cuda-repo-fedora35-11-7-local-11.7.0_515.43.04-1.x86_64.rpm'

cuda-repo-fedora35-11-7-local-11.7.0_515.43.04-1.x86_64.rpm 100%[=========================================================================================================================================>]   3.53G  28.0MB/s    in 2m 10s  

2024-11-20 16:08:18 (27.9 MB/s) - 'cuda-repo-fedora35-11-7-local-11.7.0_515.43.04-1.x86_64.rpm' saved [3795608809/3795608809]

Once the package is downloaded install it with rpm command.

sh-5.1# rpm -i cuda-repo-fedora35-11-7-local-11.7.0_515.43.04-1.x86_64.rpm
warning: cuda-repo-fedora35-11-7-local-11.7.0_515.43.04-1.x86_64.rpm: Header V4 RSA/SHA512 Signature, key ID d42d0685: NOKEY

Then clean all local repos with dnf clean all.

sh-5.1# dnf clean all
42 files removed

And finally install the CUDA toolkit.

sh-5.1# dnf -y install cuda
Fedora 35 - x86_64 - Updates                                                                                                                                                                                   32 MB/s |  34 MB     00:01    
Fedora Modular 35 - x86_64 - Updates                                                                                                                                                                          7.2 MB/s | 3.9 MB     00:00    
Dependencies resolved.
==============================================================================================================================================================================================================================================
 Package                                                            Architecture                              Version                                                       Repository                                                   Size
==============================================================================================================================================================================================================================================
Installing:
 cuda                                                               x86_64                                    11.7.0-1                                                      cuda-fedora35-11-7-local                                    2.7 k
Upgrading:
 systemd-libs                                                       x86_64                                    249.13-6.fc35                                                 updates                                                     599 k
Installing dependencies:
 NetworkManager-libnm                                               x86_64                                    1:1.32.12-2.fc35                                              updates                                                     1.7 M
 acl                                                                x86_64                                    2.3.1-2.fc35                                                  fedora                                                       71 k
(...)
  tracker-3.2.1-1.fc35.x86_64                              tracker-miners-3.2.2-1.fc35.x86_64                            ttmkfdir-3.0.9-64.fc35.x86_64                           tzdata-java-2022g-1.fc35.noarch                           
  uchardet-0.0.6-14.fc35.x86_64                            upower-0.99.13-1.fc35.x86_64                                  vulkan-loader-1.3.204.0-1.fc35.x86_64                   which-2.21-27.fc35.x86_64                                 
  xcb-util-0.4.0-18.fc35.x86_64                            xcb-util-image-0.4.0-18.fc35.x86_64                           xcb-util-keysyms-0.4.0-16.fc35.x86_64                   xcb-util-renderutil-0.3.9-19.fc35.x86_64                  
  xcb-util-wm-0.4.1-21.fc35.x86_64                         xkbcomp-1.4.5-2.fc35.x86_64                                   xkeyboard-config-2.33-2.fc35.noarch                     xml-common-0.6.3-57.fc35.noarch                           
  xorg-x11-drv-libinput-1.2.0-1.fc35.x86_64                xorg-x11-fonts-Type1-7.5-32.fc35.noarch                       xorg-x11-proto-devel-2021.5-1.fc35.noarch               xorg-x11-server-Xorg-1.20.14-9.fc35.x86_64                
  xorg-x11-server-common-1.20.14-9.fc35.x86_64             xz-5.2.5-7.fc35.x86_64                                       
Failed:
  nvidia-driver-cuda-3:515.43.04-1.fc35.x86_64                                                                          nvidia-persistenced-3:515.43.04-1.fc35.x86_64                                                                         

Error: Transaction failed

The CUDA toolkit installation will say transaction failed but this is okay. The necessary files were installed to provide what we need for building perftest.

Set the LD_LIBRARY_PATH and LIBRARY_PATH variables below.

sh-5.1# export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
sh-5.1# export LIBRARY_PATH=/usr/local/cuda/lib64:$LIBRARY_PATH

Next remove the existing /root/perftest directory in the pod and git clone down the perftest repository.

sh-5.1# rm -r -f perftest
sh-5.1# git clone https://github.com/linux-rdma/perftest.git
Cloning into 'perftest'...
remote: Enumerating objects: 6077, done.
remote: Counting objects: 100% (2157/2157), done.
remote: Compressing objects: 100% (398/398), done.
remote: Total 6077 (delta 1876), reused 1920 (delta 1747), pack-reused 3920 (from 1)
Receiving objects: 100% (6077/6077), 1.89 MiB | 43.11 MiB/s, done.
Resolving deltas: 100% (4826/4826), done.

Finally change into the perftest directory and build the binaries.

sh-5.1# cd perftest/
sh-5.1# ./autogen.sh && ./configure CUDA_H_PATH=/usr/local/cuda/include/cuda.h && make -j
libtoolize: putting auxiliary files in AC_CONFIG_AUX_DIR, 'config'.
libtoolize: copying file 'config/ltmain.sh'
libtoolize: putting macros in AC_CONFIG_MACRO_DIRS, 'm4'.
libtoolize: copying file 'm4/libtool.m4'
libtoolize: copying file 'm4/ltoptions.m4'
libtoolize: copying file 'm4/ltsugar.m4'
libtoolize: copying file 'm4/ltversion.m4'
libtoolize: copying file 'm4/lt~obsolete.m4'
libtoolize: 'AC_PROG_RANLIB' is rendered obsolete by 'LT_INIT'
configure.ac:55: installing 'config/compile'
configure.ac:59: installing 'config/config.guess'
configure.ac:59: installing 'config/config.sub'
configure.ac:36: installing 'config/install-sh'
configure.ac:36: installing 'config/missing'
Makefile.am: installing 'config/depcomp'
configure: loading site script /usr/share/config.site
checking for a BSD-compatible install... /usr/bin/install -c
checking whether build environment is sane... yes
checking for a thread-safe mkdir -p... /usr/bin/mkdir -p
checking for gawk... gawk
checking whether make sets $(MAKE)... yes
checking whether make supports nested variables... yes
checking whether make supports nested variables... (cached) yes
checking for gcc... gcc
checking whether the C compiler works... yes
checking for C compiler default output file name... a.out
checking for suffix of executables... 
checking whether we are cross compiling... no
checking for suffix of object files... o
checking whether we are using the GNU C compiler... yes
checking whether gcc accepts -g... yes
checking for gcc option to accept ISO C89... none needed
checking whether gcc understands -c and -o together... yes
checking whether make supports the include directive... yes (GNU style)
checking dependency style of gcc... gcc3
checking for g++... g++
checking whether we are using the GNU C++ compiler... yes
checking whether g++ accepts -g... yes
checking dependency style of g++... gcc3
checking dependency style of gcc... gcc3
checking build system type... x86_64-pc-linux-gnu
checking host system type... x86_64-pc-linux-gnu
checking how to print strings... printf
checking for a sed that does not truncate output... /usr/bin/sed
checking for grep that handles long lines and -e... /usr/bin/grep
checking for egrep... /usr/bin/grep -E
checking for fgrep... /usr/bin/grep -F
checking for ld used by gcc... /usr/bin/ld
checking if the linker (/usr/bin/ld) is GNU ld... yes
checking for BSD- or MS-compatible name lister (nm)... /usr/bin/nm -B
checking the name lister (/usr/bin/nm -B) interface... BSD nm
checking whether ln -s works... yes
checking the maximum length of command line arguments... 1572864
checking how to convert x86_64-pc-linux-gnu file names to x86_64-pc-linux-gnu format... func_convert_file_noop
checking how to convert x86_64-pc-linux-gnu file names to toolchain format... func_convert_file_noop
checking for /usr/bin/ld option to reload object files... -r
checking for objdump... objdump
checking how to recognize dependent libraries... pass_all
checking for dlltool... no
checking how to associate runtime and link libraries... printf %s\n
checking for ar... ar
checking for archiver @FILE support... @
checking for strip... strip
checking for ranlib... ranlib
checking command to parse /usr/bin/nm -B output from gcc object... ok
checking for sysroot... no
checking for a working dd... /usr/bin/dd
checking how to truncate binary pipes... /usr/bin/dd bs=4096 count=1
checking for mt... no
checking if : is a manifest tool... no
checking how to run the C preprocessor... gcc -E
checking for ANSI C header files... yes
checking for sys/types.h... yes
checking for sys/stat.h... yes
checking for stdlib.h... yes
checking for string.h... yes
checking for memory.h... yes
checking for strings.h... yes
checking for inttypes.h... yes
checking for stdint.h... yes
checking for unistd.h... yes
checking for dlfcn.h... yes
checking for objdir... .libs
checking if gcc supports -fno-rtti -fno-exceptions... no
checking for gcc option to produce PIC... -fPIC -DPIC
checking if gcc PIC flag -fPIC -DPIC works... yes
checking if gcc static flag -static works... no
checking if gcc supports -c -o file.o... yes
checking if gcc supports -c -o file.o... (cached) yes
checking whether the gcc linker (/usr/bin/ld -m elf_x86_64) supports shared libraries... yes
checking whether -lc should be explicitly linked in... no
checking dynamic linker characteristics... GNU/Linux ld.so
checking how to hardcode library paths into programs... immediate
checking whether stripping libraries is possible... yes
checking if libtool supports shared libraries... yes
checking whether to build shared libraries... yes
checking whether to build static libraries... yes
checking how to run the C++ preprocessor... g++ -E
checking for ld used by g++... /usr/bin/ld -m elf_x86_64
checking if the linker (/usr/bin/ld -m elf_x86_64) is GNU ld... yes
checking whether the g++ linker (/usr/bin/ld -m elf_x86_64) supports shared libraries... yes
checking for g++ option to produce PIC... -fPIC -DPIC
checking if g++ PIC flag -fPIC -DPIC works... yes
checking if g++ static flag -static works... no
checking if g++ supports -c -o file.o... yes
checking if g++ supports -c -o file.o... (cached) yes
checking whether the g++ linker (/usr/bin/ld -m elf_x86_64) supports shared libraries... yes
checking dynamic linker characteristics... (cached) GNU/Linux ld.so
checking how to hardcode library paths into programs... immediate
checking for ranlib... (cached) ranlib
checking for ANSI C header files... (cached) yes
checking infiniband/verbs.h usability... yes
checking infiniband/verbs.h presence... yes
checking for infiniband/verbs.h... yes
checking for ibv_get_device_list in -libverbs... yes
checking for rdma_create_event_channel in -lrdmacm... yes
checking for umad_init in -libumad... yes
checking for log in -lm... yes
checking for ibv_reg_dmabuf_mr in -libverbs... yes
checking pci/pci.h usability... yes
checking pci/pci.h presence... yes
checking for pci/pci.h... yes
checking for pci_init in -lpci... yes
checking for cuMemGetHandleForAddressRange in -lcuda... yes
checking for efadv_create_qp_ex in -lefa... yes
checking for mlx5dv_create_qp in -lmlx5... yes
checking for hnsdv_query_device in -lhns... no
checking that generated files are newer than configure... done
configure: creating ./config.status
config.status: creating Makefile
config.status: creating config.h
config.status: executing depfiles commands
config.status: executing libtool commands
config.status: executing man commands
make  all-am
make[1]: Entering directory '/root/perftest'
ln -s .././man/perftest.1 man/ib_read_bw.1
ln -s .././man/perftest.1 man/ib_write_bw.1
ln -s .././man/perftest.1 man/ib_send_bw.1
ln -s .././man/perftest.1 man/ib_atomic_bw.1
ln -s .././man/perftest.1 man/ib_write_lat.1
ln -s .././man/perftest.1 man/ib_read_lat.1
ln -s .././man/perftest.1 man/ib_send_lat.1
ln -s .././man/perftest.1 man/ib_atomic_lat.1
ln -s .././man/perftest.1 man/raw_ethernet_bw.1
ln -s .././man/perftest.1 man/raw_ethernet_lat.1
  CC       src/send_bw.o
ln -s .././man/perftest.1 man/raw_ethernet_burst_lat.1
ln -s .././man/perftest.1 man/raw_ethernet_fs_rate.1
  CC       src/multicast_resources.o
  CC       src/get_clock.o
  CC       src/perftest_communication.o
  CC       src/perftest_parameters.o
  CC       src/perftest_resources.o
  CC       src/perftest_counters.o
  CC       src/host_memory.o
  CC       src/mmap_memory.o
  CC       src/cuda_memory.o
  CC       src/raw_ethernet_resources.o
  CC       src/send_lat.o
  CC       src/write_lat.o
  CC       src/write_bw.o
  CC       src/read_lat.o
  CC       src/read_bw.o
  CC       src/atomic_lat.o
  CC       src/atomic_bw.o
  CC       src/raw_ethernet_send_bw.o
  CC       src/raw_ethernet_send_lat.o
  CC       src/raw_ethernet_send_burst_lat.o
  CC       src/raw_ethernet_fs_rate.o
  AR       libperftest.a
  CCLD     ib_send_bw
  CCLD     ib_write_lat
  CCLD     ib_send_lat
  CCLD     ib_write_bw
  CCLD     ib_read_lat
  CCLD     ib_read_bw
  CCLD     ib_atomic_lat
  CCLD     ib_atomic_bw
  CCLD     raw_ethernet_bw
  CCLD     raw_ethernet_lat
  CCLD     raw_ethernet_burst_lat
  CCLD     raw_ethernet_fs_rate
make[1]: Leaving directory '/root/perftest'

With the binaries built we can move onto running our validation tests.

Running RDMA Validation Tests

We already should have our workload pods running on the cluster in the default namespace.

$ oc get pods
NAME                        READY   STATUS    RESTARTS   AGE
rdma-eth-a100-01-workload   1/1     Running   0          15m
rdma-eth-a100-02-workload   1/1     Running   0          15m

Next we will need to open two rsh connections one into each pod.

$ oc rsh rdma-eth-a100-01-workload
sh-5.1# 

$ oc rsh rdma-eth-a100-02-workload
sh-5.1#

Then in the rsh connection into rdma-eth-a100-01-workload we will run the following ib_write_bw command.

sh-5.1# /root/perftest/ib_write_bw -R -T 41 -s 65536 -F -x 3 -m 4096 --report_gbits -q 16 -D 60  -d mlx5_1 -p 10000 --source_ip 172.16.0.1
 WARNING: BW peak won't be measured in this run.

************************************
* Waiting for client to connect... *
************************************

Then in the second rsh connection into rdma-eth-a100-02-workload we will run the following ib_write_bw command. Note this test is without cuda and will take a few minutes.

sh-5.1# /root/perftest/ib_write_bw -R -T 41 -s 65536 -F -x 3 -m 4096 --report_gbits -q 16 -D 60  -d mlx5_1 -p 10000 --source_ip 172.16.0.2 172.16.0.1
 WARNING: BW peak won't be measured in this run.
---------------------------------------------------------------------------------------
                    RDMA_Write BW Test
 Dual-port       : OFF        Device         : mlx5_1
 Number of qps   : 16        Transport type : IB
 Connection type : RC        Using SRQ      : OFF
 PCIe relax order: ON        Lock-free      : OFF
 ibv_wr* API     : ON        Using DDP      : OFF
 TX depth        : 128
 CQ Moderation   : 1
 Mtu             : 4096[B]
 Link type       : Ethernet
 GID index       : 3
 Max inline data : 0[B]
 rdma_cm QPs     : ON
 Data ex. method : rdma_cm     TOS    : 41
---------------------------------------------------------------------------------------
 local address: LID 0000 QPN 0x00bd PSN 0x6e902d
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00be PSN 0xdf3b13
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00bf PSN 0x14ba61
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00c0 PSN 0xd9209c
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00c1 PSN 0xc07f
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00c2 PSN 0xf06575
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00c3 PSN 0x481230
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00c4 PSN 0xc1a69
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00c5 PSN 0x7c6e59
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00c6 PSN 0xf16f67
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00c7 PSN 0xe82e7f
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00c8 PSN 0xf0a6a6
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00c9 PSN 0x41069a
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00ca PSN 0xe2153f
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00cb PSN 0xed2a91
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00cc PSN 0x2f3581
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 remote address: LID 0000 QPN 0x00c7 PSN 0xc8665d
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00c8 PSN 0xce8d83
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00c9 PSN 0x4b7411
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00ca PSN 0x3a508c
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00cb PSN 0xb3c9af
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00cc PSN 0xac6ee5
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00cd PSN 0x12b6e0
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00ce PSN 0x8a5959
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00cf PSN 0x8da89
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00d0 PSN 0x2e9fd7
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00d1 PSN 0xba6e2f
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00d2 PSN 0xede496
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00d3 PSN 0xfa05ca
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00d4 PSN 0x2bdcaf
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00d5 PSN 0xc5b541
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00d6 PSN 0x3c6271
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
---------------------------------------------------------------------------------------
 #bytes     #iterations    BW peak[Gb/sec]    BW average[Gb/sec]   MsgRate[Mpps]
 65536      5296615          0.00               92.56               0.176539
---------------------------------------------------------------------------------------

Now we are going to repeat the test but include the GPU with the --use_cuda switch on the command. So in the first rsh connection run.

sh-5.1# /root/perftest/ib_write_bw -R -T 41 -s 65536 -F -x 3 -m 4096 --report_gbits -q 16 -D 60  -d mlx5_1 -p 10000 --source_ip 172.16.0.1 --use_cuda=0
 WARNING: BW peak won't be measured in this run.
Perftest doesn't supports CUDA tests with inline messages: inline size set to 0

************************************
* Waiting for client to connect... *
************************************

Then in the second rsh connection run the following.

sh-5.1# /root/perftest/ib_write_bw -R -T 41 -s 65536 -F -x 3 -m 4096 --report_gbits -q 16 -D 60  -d mlx5_1 -p 10000 --source_ip 172.16.0.2 --use_cuda=0 172.16.0.1
 WARNING: BW peak won't be measured in this run.
Perftest doesn't supports CUDA tests with inline messages: inline size set to 0
initializing CUDA
Listing all CUDA devices in system:
CUDA device 0: PCIe address is 0F:00

Picking device No. 0
[pid = 4488, dev = 0] device name = [NVIDIA A100-SXM4-80GB]
creating CUDA Ctx
making it the current CUDA Ctx
CUDA device integrated: 0
cuMemAlloc() of a 2097152 bytes GPU buffer
allocated GPU buffer address at 00007fbebf200000 pointer=0x7fbebf200000
---------------------------------------------------------------------------------------
                    RDMA_Write BW Test
 Dual-port       : OFF        Device         : mlx5_1
 Number of qps   : 16        Transport type : IB
 Connection type : RC        Using SRQ      : OFF
 PCIe relax order: ON        Lock-free      : OFF
 ibv_wr* API     : ON        Using DDP      : OFF
 TX depth        : 128
 CQ Moderation   : 1
 Mtu             : 4096[B]
 Link type       : Ethernet
 GID index       : 3
 Max inline data : 0[B]
 rdma_cm QPs     : ON
 Data ex. method : rdma_cm     TOS    : 41
---------------------------------------------------------------------------------------
 local address: LID 0000 QPN 0x00ce PSN 0x282aa6
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00cf PSN 0x3ab698
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00d0 PSN 0x9dd002
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00d1 PSN 0x11fc29
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00d2 PSN 0x72e988
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00d3 PSN 0xb5f44a
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00d4 PSN 0x1540e1
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00d5 PSN 0x8801c6
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00d6 PSN 0xd77ef2
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00d7 PSN 0xacf68c
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00d8 PSN 0x47f740
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00d9 PSN 0x286d3
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00da PSN 0xc1e7c3
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00db PSN 0xd8c9b4
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00dc PSN 0xf51e62
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 local address: LID 0000 QPN 0x00dd PSN 0x4bcb7e
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:100
 remote address: LID 0000 QPN 0x00d8 PSN 0x533e76
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00d9 PSN 0x1f0628
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00da PSN 0xf1052
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00db PSN 0xd23e39
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00dc PSN 0x696a58
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00dd PSN 0x25acda
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00de PSN 0x383631
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00df PSN 0x9054d6
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00e0 PSN 0xc33cc2
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00e1 PSN 0x55a81c
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00e2 PSN 0x62f190
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00e3 PSN 0x22fae3
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00e4 PSN 0x99b293
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00e5 PSN 0xb10444
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00e6 PSN 0x636db2
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
 remote address: LID 0000 QPN 0x00e7 PSN 0x45708e
 GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:03:200
---------------------------------------------------------------------------------------
 #bytes     #iterations    BW peak[Gb/sec]    BW average[Gb/sec]   MsgRate[Mpps]
 65536      3994687          0.00               69.79               0.133122
---------------------------------------------------------------------------------------
deallocating GPU buffer 00007fbebf200000
destroying current CUDA Ctx

Once the tests complete we have confirmed the RDMA is working and we can now add our real-world workload.

Hopefully this blog was useful in showing RDMA with CUDA testing on an OpenShift environment.