3.10. Network Support

3.10.1. Main network support models

There are multiple MPI network models available in this release:

  • ob1 supports a variety of networks using BTL (“Byte Transfer Layer”) plugins that can be used in combination with each other:

    • self: Loopback (send-to-self)

    • sm: Shared memory, including single-copy technologies: XPMEM, Linux CMA, as Linux KNEM, as well as traditional copy-in/copy-out shared memory.

    • tcp: TCP

    • smcuda: CUDA-enabled shared memory

    • usnic: Cisco usNIC

    • ugni: uGNI (Cray Gemini, Aries)

  • cm supports a smaller number of networks (and they cannot be used together), but may provide better overall MPI performance by utilizing MTL (“Matching Transport Layer”) plugins:

    • OpenFabrics Interfaces (“libfabric” tag matching)

    • Intel Omni-Path PSM2 (version 11.2.173 or later)

    • Intel True Scale PSM (QLogic InfiniPath)

    • Portals 4

  • ucx uses the Unified Communication X (UCX) communication library. This is an open-source project developed in collaboration between industry, laboratories, and academia to create an open-source production grade communication framework for data centric and high-performance applications. The UCX library can be downloaded from repositories (e.g., Fedora/RedHat yum repositories). The UCX library is also part of Mellanox OFED and Mellanox HPC-X binary distributions.

    UCX currently supports:

    • OpenFabrics Verbs (including InfiniBand and RoCE)

    • Cray’s uGNI

    • TCP

    • Shared memory

    • NVIDIA CUDA drivers

While users can manually select any of the above transports at run time, Open MPI will select a default transport as follows:

  1. If InfiniBand devices are available, use the UCX PML.

  2. If PSM, PSM2, or other tag-matching-supporting Libfabric transport devices are available (e.g., Cray uGNI), use the cm PML and a single appropriate corresponding mtl module.

  3. Otherwise, use the ob1 PML and one or more appropriate btl modules.

Users can override Open MPI’s default selection algorithms and force the use of a specific transport if desired by setting the pml MCA parameter (and potentially the btl and/or mtl MCA parameters) at run-time:

shell$ mpirun --mca pml ob1 --mca btl [comma-delimted-BTLs] ...
# or
shell$ mpirun --mca pml cm --mca mtl [MTL] ...
# or
shell$ mpirun --mca pml ucx ...

There is a known issue when using UCX with very old Mellanox Infiniband HCAs, in particular HCAs preceding the introduction of the ConnectX product line, which can result in Open MPI crashing in MPI_Finalize. This issue is addressed by UCX release 1.9.0 and newer.

3.10.2. Miscellaneous network notes

  • The main OpenSHMEM network model is ucx; it interfaces directly with UCX.

  • In prior versions of Open MPI, InfiniBand and RoCE support was provided through the openib BTL and ob1 PML plugins. Starting with Open MPI 4.0.0, InfiniBand support through the openib plugin is both deprecated and superseded by the ucx PML component. The openib BTL was removed in Open MPI v5.0.0.

    While the openib BTL depended on libibverbs, the UCX PML depends on the UCX library.

    Once installed, Open MPI can be built with UCX support by adding --with-ucx to the Open MPI configure command. Once Open MPI is configured to use UCX, the runtime will automatically select the ucx PML if one of the supported networks is detected (e.g., InfiniBand). It’s possible to force using UCX in the mpirun or oshrun command lines by specifying any or all of the following mca parameters: --mca pml ucx for MPI point-to-point operations, --mca spml ucx for OpenSHMEM support, and --mca osc ucx for MPI RMA (one-sided) operations.

  • The usnic BTL is support for Cisco’s usNIC device (“userspace NIC”) on Cisco UCS servers with the Virtualized Interface Card (VIC). Although the usNIC is accessed via the OpenFabrics Libfabric API stack, this BTL is specific to Cisco usNIC devices.

  • uGNI is a Cray library for communicating over the Gemini and Aries interconnects.

  • Linux knem support is used when the sm (shared memory) BTL is compiled with knem support (see the --with-knem configure option) and the knem Linux module is loaded in the running kernel. If the knem Linux kernel module is not loaded, the knem support is (by default) silently deactivated during Open MPI jobs.

    See https://knem.gitlabpages.inria.fr/ for details on Knem.

  • Linux Cross-Memory Attach (CMA) or XPMEM is used by the sm shared memory BTL when the CMA/XPMEM libraries are installed, respectively. Linux CMA and XPMEM are similar (but different) mechanisms for Open MPI to utilize single-copy semantics for shared memory.

  • The OFI MTL does not support sending messages larger than the active Libfabric provider’s max_msg_size. If you receive an error message about sending too large of a message when using the OFI MTL, please reach out to your networking vendor to ask them to support a larger max_msg_size for tagged messages.