InfiniBand Basic Setup

This guide provides a basic setup for enabling InfiniBand on your nodes, ensuring proper network configuration, and running NCCL and Torch distributed training over IB.

⚠️ Warning: The following setup guide is a general reference and may not work for every situation. Please adjust configurations based on your specific environment and workload requirements.

---

Introduction

InfiniBand (IB) is a high-speed networking technology often used for distributed computing, AI workloads, and high-performance computing (HPC) applications. This guide provides a basic setup for enabling InfiniBand on your nodes, ensuring proper network configuration, and running NCCL and Torch distributed training over IB.

1. Verify InfiniBand Device Availability

To check if InfiniBand devices are detected on your system, run:

lspci | grep Mellanox

If detected, check if the IB kernel modules are loaded:

lsmod | grep mlx

To list InfiniBand interfaces:

ibstat

If ibstat is not available, install rdma-core:

sudo apt install -y rdma-core

To list available InfiniBand devices:

ls /sys/class/infiniband/

If ibdev2netdev is missing, install the required package:

sudo apt install -y mlnx-tools

Then run:

ibdev2netdev

2. Install Required Packages

2.1 Install Mellanox OFED Drivers

If the MLNX_OFED driver is not installed, follow these steps:

wget https://www.mellanox.com/downloads/DOCA/DOCA_v2.9.1/host/doca-host_2.9.1-018000-24.10-ubuntu2204_amd64.deb
sudo dpkg -i doca-host_2.9.1-018000-24.10-ubuntu2204_amd64.deb
sudo apt install -y doca-host

Verify installation:

dpkg -l | grep -i ofed
ofed_info -s
lsmod | grep -E 'mlx|ib|rdma'

2.2 Install Essential Networking Tools

To install essential tools for InfiniBand communication:

sudo apt update && sudo apt install -y \
    rdma-core \
    mlnx-tools \
    infiniband-diags \
    ibverbs-utils \
    perftest \
    iproute2 \
    net-tools

2.3 Install CUDA and NCCL Support

For CUDA support:

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt-get update
sudo apt install -y \
    cuda-toolkit-12-8 \
    nvidia-cuda-toolkit \
    libnccl2 \
    libnccl-dev

3. Running a Basic InfiniBand Test

3.1 Interactive Pingpong Test

To verify connectivity between two nodes, use the following interactive scripts:

Server Script:

#!/bin/bash
echo "Starting InfiniBand pingpong test server..."
while true; do
  for n in 0 3 4 5 6 9 10 11; do
    ibv_rc_pingpong -s 65536 -d mlx5_$n
  done
done

How to use: Run this script on the server node to continuously listen for incoming pingpong tests.

Client Script:

#!/bin/bash
echo -n "Enter the IP address of the server: "
read SERVER_IP

if ping -c 1 "$SERVER_IP" > /dev/null 2>&1; then
  echo "$SERVER_IP is reachable. Continuing..."
else
  echo "Error: $SERVER_IP is not reachable."
  exit 1
fi

for n in 0 3 4 5 6 9 10 11; do
  echo "Testing device: mlx5_${n} against ${SERVER_IP}"
  sudo ibv_rc_pingpong -s 65536 -d mlx5_${n} ${SERVER_IP}
done

How to use: Run this script on the client node, enter the server IP when prompted, and it will test IB connectivity.

4. Running NCCL Tests with InfiniBand

Clone and build NCCL tests:

git clone https://github.com/NVIDIA/nccl-tests.git
cd nccl-tests
make -j

Run an NCCL all-reduce performance test:

export LD_LIBRARY_PATH=/usr/mpi/gcc/openmpi-4.1.7rc1/lib:$LD_LIBRARY_PATH
/usr/mpi/gcc/openmpi-4.1.7rc1/bin/mpirun -np 2 -bind-to none --verbose  \
    --hostfile ./mpihosts  \
    -x LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda/libnvvp:/usr/local/lib \
    -x NCCL_IB_HCA=mlx5 \
    -x UCX_NET_DEVICES=mlx5_0:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_9:1,mlx5_10:1,mlx5_11:1 \
    -x SHARP_COLL_ENABLE_PCI_RELAXED_ORDERING=1 \
    -x NCCL_COLLNET_ENABLE=0 \
    /home/ubuntu/nccl-tests/build/all_reduce_perf -b 512M -e 8G -f 2 -g 1

5. Debugging Common InfiniBand Issues

• If ibdev2netdev is missing, install mlnx-tools.

• If training hangs, ensure MLNX_OFED is installed and all IB devices show "Up".

• If NCCL falls back to Ethernet, verify IB interfaces with:

  ibstat

• Use lsmod | grep mlx to verify IB drivers are loaded.

Conclusion

InfiniBand provides a powerful interconnect for distributed training, but correct configuration is key to ensuring optimal performance. If you encounter issues, confirm driver installation, check network device mappings, and validate NCCL environment variable settings.

For further assistance, please contact CX Team.