Skip to content

Latest commit

 

History

History
83 lines (54 loc) · 3.51 KB

README.md

File metadata and controls

83 lines (54 loc) · 3.51 KB

GPUDirect RDMA

The latest advancement in GPU-GPU communications is GPUDirect RDMA. This new technology provides a direct P2P (Peer-to-Peer) data path between the GPU Memory directly to/from the NVIDIA HCA/NIC devices. This provides a significant decrease in GPU-GPU communication latency and completely offloads the CPU, removing it from all GPU-GPU communications across the network.

Mellanox Product Family

General

MLNX_OFED 2.1 introduces an API between IB CORE to peer memory clients, such as NVIDIA Kepler class GPU's, (e.g. GPU cards), also known as GPUDirect RDMA. It provides access for the HCA to read/write peer memory data buffers, as a result it allows RDMA-based applications to use the peer device computing power with the RDMA interconnect without the need for copying data to host memory.

This capability is supported with Mellanox ConnectX-3 VPI or Connect-IB InfiniBand adapters. It will also seamlessly work using RoCE technology with the Mellanox ConnectX-3 VPI adapters.

This README describes the required steps to completing the installation for the NVIDIA peer memory client with Mellanox OFED.

A kernel module with comparable functionalities has been integrated into the GPU driver, starting from the release R470, under the name nvidia-peermem.

Installation

Starting from version 1.2, nv_peer_mem requires a MLNX_OFED containing a fix for “Peer-direct patch may cause deadlock due to lock inversion" (tracked by the Internal Ref. #2696789).

nv_peer_mem version 1.1 is the last one to support MLNX_OFED LTS 4.9.

Pre-requisites:

  1. NVIDIA compatible driver is installed and up.
  2. MLNX_OFED 5.1 or newer (with the fix of bug #2696789) is installed and up.

Failure to have the proper configuration (as described above) will result in build failure.

For the required NVIDIA driver and other relevant details in that area please check with NVIDIA support.

To build source packages (src.rpm for RPM based OS and tarball for DEB based OS), use the build_module.sh script.

For example, to build on RPM based OS:

$ ./build_module.sh
Building source rpm for nvidia_peer_memory...

Built: /tmp/nvidia_peer_memory-1.3-0.src.rpm

To install run on RPM based OS:
# rpmbuild --rebuild /tmp/nvidia_peer_memory-1.3-0.src.rpm
# rpm -ivh <path to generated binary rpm file>

To build on DEB based OS:

Building debian tarball for nvidia-peer-memory...

Built: /tmp/nvidia-peer-memory_1.3.orig.tar.gz

To install on DEB based OS:
# cd /tmp
# tar xzf /tmp/nvidia-peer-memory_1.3.orig.tar.gz
# cd nvidia-peer-memory-1.3
# dpkg-buildpackage -us -uc
# dpkg -i <path to generated deb files>            

To install run (excluding Ubuntu):

rpmbuild --rebuild <path to srpm>.
rpm -ivh <path to generated binary rpm file.> [On SLES add --nodeps].

To install on Ubuntu run:

dpkg-buildpackage -us -uc
dpkg -i <path to generated deb files.>

(e.g. dpkg -i nvidia-peer-memory_1.3-0_all.deb
      dpkg -i nvidia-peer-memory-dkms_1.3-0_all.deb)

After successful installation:

  1. nv_peer_mem.ko is installed
  2. service file /etc/init.d/nv_peer_mem to be used for start/stop/status for that kernel module was added.
  3. /etc/infiniband/nv_peer_mem.conf to control whether kernel module will be loaded on boot (default is YES) was added.

Notes

To achieve good performance both the NIC and the GPU must physically sit on same i/o root complex, use lspci -tv to make sure that this is the case.