From fdbb4418500ed680c8d3a0d00af28df9584c920c Mon Sep 17 00:00:00 2001 From: Mikael Simberg Date: Fri, 14 Nov 2025 12:16:51 +0100 Subject: [PATCH 1/5] Update OpenMPI documentation --- docs/software/communication/openmpi.md | 60 +++++++++++++--------- docs/software/prgenv/index.md | 4 ++ docs/software/prgenv/prgenv-gnu-openmpi.md | 18 +++++++ docs/software/prgenv/prgenv-gnu.md | 2 + 4 files changed, 61 insertions(+), 23 deletions(-) create mode 100644 docs/software/prgenv/prgenv-gnu-openmpi.md diff --git a/docs/software/communication/openmpi.md b/docs/software/communication/openmpi.md index 9c45c0da..8d7302b4 100644 --- a/docs/software/communication/openmpi.md +++ b/docs/software/communication/openmpi.md @@ -4,21 +4,11 @@ [Cray MPICH][ref-communication-cray-mpich] is the recommended MPI implementation on Alps. However, [OpenMPI](https://www.open-mpi.org/) can be used as an alternative in some cases, with limited support from CSCS. -To use OpenMPI on Alps, it must be built against [libfabric][ref-communication-libfabric] with support for the [Slingshot 11 network][ref-alps-hsn]. +Support for the [Slingshot 11 network][ref-alps-hsn] is provided by the [libfabric][ref-communication-libfabric] library. ## Using OpenMPI -!!! warning - Building and using OpenMPI on Alps is still [work in progress](https://eth-cscs.github.io/cray-network-stack/). - The instructions found on this page may be inaccurate, but are a good starting point to using OpenMPI on Alps. - -!!! todo - Deploy experimental uenv. - -!!! todo - Document OpenMPI uenv next to prgenv-gnu, prgenv-nvfortran, and linalg? - -OpenMPI is provided through a [uenv][ref-uenv] similar to [`prgenv-gnu`][ref-uenv-prgenv-gnu]. +OpenMPI is provided in the [`prgenv-gnu-openmpi`][ref-uenv-prgenv-gnu-openmpi] uenv. Once the uenv is loaded, compiling and linking with OpenMPI and libfabric is transparent. At runtime, some additional options must be set to correctly use the Slingshot network. @@ -32,27 +22,51 @@ Additionally, the following environment variables should be set: ```bash export PMIX_MCA_psec="native" # (1)! export FI_PROVIDER="cxi" # (2)! -export OMPI_MCA_pml="^ucx" # (3)! +export OMPI_MCA_pml="cm" # (3)! export OMPI_MCA_mtl="ofi" # (4)! ``` 1. Ensures PMIx uses the same security domain as Slurm. Otherwise PMIx will print warnings at startup. 2. Use the CXI (Slingshot) provider. -3. Use anything except [UCX](https://openucx.org/documentation/) for [point-to-point communication](https://docs.open-mpi.org/en/v5.0.x/mca.html#selecting-which-open-mpi-components-are-used-at-run-time). The `^` signals that OpenMPI should exclude all listed components. +3. Use CM for [point-to-point communication](https://docs.open-mpi.org/en/v5.0.x/mca.html#selecting-which-open-mpi-components-are-used-at-run-time). 4. Use libfabric for the [Matching Transport Layer](https://docs.open-mpi.org/en/v5.0.x/mca.html#frameworks). !!! info "CXI provider does all communication through the network interface cards (NICs)" When using the libfabric CXI provider, all communication goes through NICs, including intra-node communication. This means that intra-node communication can not make use of shared memory optimizations and the maximum bandwidth will not be severely limited. - Libfabric has a new [LINKx](https://ofiwg.github.io/libfabric/v2.1.0/man/fi_lnx.7.html) provider, which allows using different libfabric providers for inter- and intra-node communication. - This provider is not as well tested, but can in theory perform better for intra-node communication, because it can use shared memory. - To use the LINKx provider, set the following, instead of `FI_PROVIDER=cxi`: +### Using the experimental LINKx provider + +The default configuration routes all communication through the NICs. +While performance may sometimes be acceptable, this mode does not make full use of the much higher intra-node bandwidth available on Grace-Hopper nodes. +In particular, GPU-GPU communication is significantly faster when using the appropriate intra-node links. + +The experimental [LINKx](https://ofiwg.github.io/libfabric/v2.3.1/man/fi_lnx.7.html) libfabric provider allows composing multiple libfabric providers for inter- and intra-node communication. +The CXI provider can be used for inter-node communication while the shared memory (`shm`) provider can be used to take advantage of xpmem for CPU-CPU communication and GDRCopy for GPU-GPU communication. + +!!! warning "The LINKx provider is experimental and may contain bugs, in particular for intra-node communication" + +To use the LINKx provider, set `--mpi=pmix`, exactly as without the LINKx provider. +Additionally, set the following environment variables: + +```bash +export PMIX_MCA_psec="native" +export FI_PROVIDER="lnx" # (1)! +export FI_LNX_PROV_LINKS="shm+cxi:cxi0|shm+cxi:cxi1|shm+cxi:cxi2|shm+cxi:cxi3" # (2)! +export FI_SHM_USE_XPMEM=1 # (3)! +export OMPI_MCA_pml="cm" +export OMPI_MCA_mtl="ofi" +export OMPI_MCA_mtl_ofi_av=table # (4)! +``` + +1. Use the libfabric LINKx provider, to allow using different libfabric providers for inter- and intra-node communication. +2. Specify which providers LINKx should use. + Use the shared memory provider for intra-node communication and the CXI (Slingshot) provider for inter-node communication. + Choose one of the four available NICs on a node in a round-robin fashion. +3. Explicitly use xpmem for CPU-CPU communication. + The default is to use CMA. +4. The LINKx provider requires this option to be set. TODO: Better explanation? - ```bash - export FI_PROVIDER="lnx" # (1)! - export FI_LNX_PROV_LINKS="shm+cxi" # (2)! - ``` +## Known issues - 1. Use the libfabric LINKx provider, to allow using different libfabric providers for inter- and intra-node communication. - 2. Use the shared memory provider for intra-node communication and the CXI (Slingshot) provider for inter-node communication. +TODO: async collectives don't work. diff --git a/docs/software/prgenv/index.md b/docs/software/prgenv/index.md index 62b53a6c..847e7d81 100644 --- a/docs/software/prgenv/index.md +++ b/docs/software/prgenv/index.md @@ -10,6 +10,10 @@ CSCS provides "programming environments" on Alps vClusters that provide compiler Provides compilers, MPI, tools and libraries built around the GNU compiler toolchain. It is the go to programming environment on all systems and target node types, that is it is the first that you should try out when starting to compile an application or create a python virtual environment. +- :fontawesome-solid-layer-group: [__prgenv-gnu-openmpi__][ref-uenv-prgenv-gnu-openmpi] uenv + + Same as [`prgenv-gnu`][ref-uenv-prgenv-gnu] but with OpenMPI instead of Cray MPICH. + - :fontawesome-solid-layer-group: [__prgenv-nvfortran__][ref-uenv-prgenv-nvfortran] uenv Provides a set of tools and libraries for building applications that need the NVIDIA Fortran compiler, commonly required for OpenACC and CUDA-Fortran applications. diff --git a/docs/software/prgenv/prgenv-gnu-openmpi.md b/docs/software/prgenv/prgenv-gnu-openmpi.md new file mode 100644 index 00000000..b4c43c31 --- /dev/null +++ b/docs/software/prgenv/prgenv-gnu-openmpi.md @@ -0,0 +1,18 @@ +[](){#ref-uenv-prgenv-gnu-openmpi} +# prgenv-gnu-openmpi + +The `prgenv-gnu-openmpi` uenv is otherwise similar to [`prgenv-gnu`][ref-uenv-prgenv-gnu] except it provides OpenMPI instead of Cray MPICH. + +!!! warning "OpenMPI is not officially supported on CSCS systems" + Cray MPICH is the preferred, and officially supported, MPI implementation on CSCS systems. + OpenMPI is provided on a best effort basis. + While most applications should work correctly with OpenMPI, there may be missing features, broken functionality, or bad performance compared to Cray MPICH. + Issues are best reported upstream, but CSCS is happy to help facilitate and coordinate issue reporting if you [get in touch][ref-get-in-touch]. + +Use of the uenv is similar to [`prgenv-gnu`][ref-uenv-prgenv-gnu]. +See the [OpenMPI documentation][ref-communication-openmpi] for important information on configuring OpenMPI to take advantage of the Slingshot network. + +### Versions + +=== "25.11" + TODO diff --git a/docs/software/prgenv/prgenv-gnu.md b/docs/software/prgenv/prgenv-gnu.md index c07fd0df..3e978e95 100644 --- a/docs/software/prgenv/prgenv-gnu.md +++ b/docs/software/prgenv/prgenv-gnu.md @@ -10,6 +10,8 @@ It is the go to programming environment on all systems and target node types, th The [`linalg`][ref-uenv-linalg] environment is similar to prgenv-gnu, with additional linear algebra and mesh partitioning algorithms. + The [`prgenv-gnu-openmpi`][ref-uenv-prgenv-gnu-openmpi] is otherwise similar to `prgenv-gnu`, but provides OpenMPI instead of Cray MPICH. + [](){#ref-uenv-prgenv-gnu-versioning} ## Versioning From 21f78e02a11a8b5703010cfd46d7d98a11e419c9 Mon Sep 17 00:00:00 2001 From: Mikael Simberg Date: Fri, 14 Nov 2025 17:02:53 +0100 Subject: [PATCH 2/5] Add link to missing GPU-GPU synchronization issue --- docs/software/communication/openmpi.md | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/docs/software/communication/openmpi.md b/docs/software/communication/openmpi.md index 8d7302b4..b66e23dc 100644 --- a/docs/software/communication/openmpi.md +++ b/docs/software/communication/openmpi.md @@ -46,7 +46,11 @@ The CXI provider can be used for inter-node communication while the shared memor !!! warning "The LINKx provider is experimental and may contain bugs, in particular for intra-node communication" -To use the LINKx provider, set `--mpi=pmix`, exactly as without the LINKx provider. + A patch has been included in the [`prgenv-gnu-openmpi`][ref-uenv-prgenv-gnu-openmpi] uenv for [this LINKx issue](https://github.com/ofiwg/libfabric/issues/11231). + However, the patch may be incomplete and other issues may still be present. + Always validate your results to ensure MPI is working correctly. + +To use the LINKx provider, set `--mpi=pmix`, as without the LINKx provider. Additionally, set the following environment variables: ```bash From ece72a414db04d739c8824f3470642902da88f56 Mon Sep 17 00:00:00 2001 From: Mikael Simberg Date: Fri, 14 Nov 2025 17:09:08 +0100 Subject: [PATCH 3/5] Remove word --- docs/software/communication/openmpi.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/software/communication/openmpi.md b/docs/software/communication/openmpi.md index b66e23dc..6b369da4 100644 --- a/docs/software/communication/openmpi.md +++ b/docs/software/communication/openmpi.md @@ -33,7 +33,7 @@ export OMPI_MCA_mtl="ofi" # (4)! !!! info "CXI provider does all communication through the network interface cards (NICs)" When using the libfabric CXI provider, all communication goes through NICs, including intra-node communication. - This means that intra-node communication can not make use of shared memory optimizations and the maximum bandwidth will not be severely limited. + This means that intra-node communication can not make use of shared memory optimizations and the maximum bandwidth will be severely limited. ### Using the experimental LINKx provider From dea803214d83157a26a26f615e03204f3cc96c6a Mon Sep 17 00:00:00 2001 From: Mikael Simberg Date: Wed, 17 Dec 2025 13:00:00 +0100 Subject: [PATCH 4/5] Refactor lnx vs cxi openmpi sections --- docs/software/communication/openmpi.md | 28 +++++++++++++++++++------- 1 file changed, 21 insertions(+), 7 deletions(-) diff --git a/docs/software/communication/openmpi.md b/docs/software/communication/openmpi.md index 6b369da4..06355320 100644 --- a/docs/software/communication/openmpi.md +++ b/docs/software/communication/openmpi.md @@ -18,7 +18,21 @@ This is done with the `--mpi` flag of `srun`: srun --mpi=pmix ... ``` -Additionally, the following environment variables should be set: +There are two primary ways to configure OpenMPI and libfabric to use the Slingshot network: + +1. [Only using the CXI provider][ref-communication-openmpi-cxi]. + This method has been found to work in more applications but uses NICs for intra-node communication which can limit performance. +2. [Using the LINKx provider][ref-communication-openmpi-lnx] which combines the CXI provider for inter-node communication with the shared memory provider for intra-node communication. + This provider is newer, may not support all features, and more likely to contain bugs, but makes full use of intra-node bandwidth. + +We recommend trying the LINKx provider first as it provides better performance in the situations that it's supported. +If you encounter failures using the LINKx provider we ask you to [get in touch with us][ref-get-in-touch] so that we can evaluate whether upstream libfabric or OpenMPI need fixing. + +[](){#ref-communication-openmpi-cxi} +### Using the CXI provider + +To use the CXI provider the following environment variables should be set: + ```bash export PMIX_MCA_psec="native" # (1)! export FI_PROVIDER="cxi" # (2)! @@ -34,8 +48,10 @@ export OMPI_MCA_mtl="ofi" # (4)! !!! info "CXI provider does all communication through the network interface cards (NICs)" When using the libfabric CXI provider, all communication goes through NICs, including intra-node communication. This means that intra-node communication can not make use of shared memory optimizations and the maximum bandwidth will be severely limited. + Use the [LINKx][ref-communication-openmpi-lnx] provider to make full use of the available intra-node bandwidth. -### Using the experimental LINKx provider +[](){#ref-communication-openmpi-lnx} +### Using the LINKx provider The default configuration routes all communication through the NICs. While performance may sometimes be acceptable, this mode does not make full use of the much higher intra-node bandwidth available on Grace-Hopper nodes. @@ -44,14 +60,12 @@ In particular, GPU-GPU communication is significantly faster when using the appr The experimental [LINKx](https://ofiwg.github.io/libfabric/v2.3.1/man/fi_lnx.7.html) libfabric provider allows composing multiple libfabric providers for inter- and intra-node communication. The CXI provider can be used for inter-node communication while the shared memory (`shm`) provider can be used to take advantage of xpmem for CPU-CPU communication and GDRCopy for GPU-GPU communication. -!!! warning "The LINKx provider is experimental and may contain bugs, in particular for intra-node communication" +!!! danger "The LINKx provider is experimental" - A patch has been included in the [`prgenv-gnu-openmpi`][ref-uenv-prgenv-gnu-openmpi] uenv for [this LINKx issue](https://github.com/ofiwg/libfabric/issues/11231). - However, the patch may be incomplete and other issues may still be present. + While many basic tests work correctly using the LINKx provider we have had reports of applications failing to run with the LINKx provider. Always validate your results to ensure MPI is working correctly. -To use the LINKx provider, set `--mpi=pmix`, as without the LINKx provider. -Additionally, set the following environment variables: +To use the LINKx provider set the following environment variables: ```bash export PMIX_MCA_psec="native" From 77f8e2b4a54bd1001ad870f7f4323e71dc412867 Mon Sep 17 00:00:00 2001 From: Mikael Simberg Date: Wed, 17 Dec 2025 13:00:09 +0100 Subject: [PATCH 5/5] Add example of broken MPI_Iallreduce with GPU buffers using OpenMPI --- docs/software/communication/openmpi.md | 53 +++++++++++++++++++++++++- 1 file changed, 52 insertions(+), 1 deletion(-) diff --git a/docs/software/communication/openmpi.md b/docs/software/communication/openmpi.md index 06355320..512834c5 100644 --- a/docs/software/communication/openmpi.md +++ b/docs/software/communication/openmpi.md @@ -87,4 +87,55 @@ export OMPI_MCA_mtl_ofi_av=table # (4)! ## Known issues -TODO: async collectives don't work. +Some asynchronous collectives are known not to work with GPU buffers, independent of the libfabric provider used. +For example, `MPI_Iallreduce` will fail with a segmentation fault. +Running the `osu_iallreduce` benchmark with GPU buffers results in: + +```console +$ srun -u --mpi=pmix -n4 osu_iallreduce -d cuda + +# OSU MPI-CUDA Non-blocking Allreduce Latency Test v7.5 +# Overall = Coll. Init + Compute + MPI_Test + MPI_Wait + +# Datatype: MPI_INT. +# Size Overall(us) Compute(us) Pure Comm.(us) Overlap(%) +[nid006549:31808] *** Process received signal *** +[nid006549:31808] Signal: Segmentation fault (11) +[nid006549:31808] Signal code: Invalid permissions (2) +[nid006549:31808] Failing at address: 0x4002da000000 +[nid006550:188198] *** Process received signal *** +[nid006550:188198] Signal: Segmentation fault (11) +[nid006550:188198] Signal code: Invalid permissions (2) +[nid006550:188198] Failing at address: 0x40029a000000 +[nid006549:31808] [ 0] linux-vdso.so.1(__kernel_rt_sigreturn+0x0)[0x400027ce07dc] +[nid006549:31808] [ 1] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libmpi.so.40(+0x19f1c8)[0x400029b0f1c8] +[nid006549:31808] [ 2] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libmpi.so.40(+0x12836c)[0x400029a9836c] +[nid006549:31808] [ 3] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libmpi.so.40(NBC_Progress+0x164)[0x400029a97bd4] +[nid006549:31808] [ 4] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libmpi.so.40(ompi_coll_libnbc_progress+0x8c)[0x400029a96a0c] +[nid006549:31808] [ 5] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libopen-pal.so.80(opal_progress+0x3c)[0x40002a23737c] +[nid006549:31808] [ 6] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libmpi.so.40(ompi_request_default_wait+0x50)[0x4000299f3810] +[nid006549:31808] [ 7] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libmpi.so.40(MPI_Wait+0x64)[0x400029a3df24] +[nid006549:31808] [ 8] /user-environment/env/default/libexec/osu-micro-benchmarks/mpi/collective/osu_iallreduce[0x40424c] +[nid006549:31808] [ 9] /lib64/libc.so.6(__libc_start_main+0xe8)[0x40002a073fa0] +[nid006549:31808] [10] /user-environment/env/default/libexec/osu-micro-benchmarks/mpi/collective/osu_iallreduce[0x404e98] +[nid006549:31808] *** End of error message *** +[nid006550:188198] [ 0] linux-vdso.so.1(__kernel_rt_sigreturn+0x0)[0x4000026a07dc] +[nid006550:188198] [ 1] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libmpi.so.40(+0x19f1c8)[0x4000044cf1c8] +[nid006550:188198] [ 2] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libmpi.so.40(+0x12836c)[0x40000445836c] +[nid006550:188198] [ 3] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libmpi.so.40(NBC_Progress+0x164)[0x400004457bd4] +[nid006550:188198] [ 4] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libmpi.so.40(ompi_coll_libnbc_progress+0x8c)[0x400004456a0c] +[nid006550:188198] [ 5] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libopen-pal.so.80(opal_progress+0x3c)[0x400004bf737c] +[nid006550:188198] [ 6] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libmpi.so.40(ompi_request_default_wait+0x50)[0x4000043b3810] +[nid006550:188198] [ 7] /user-environment/linux-neoverse_v2/openmpi-5.0.9-leskuw5dyswfdw3eaybcyfmsrbid3uuq/lib/libmpi.so.40(MPI_Wait+0x64)[0x4000043fdf24] +[nid006550:188198] [ 8] /user-environment/env/default/libexec/osu-micro-benchmarks/mpi/collective/osu_iallreduce[0x40424c] +[nid006550:188198] [ 9] /lib64/libc.so.6(__libc_start_main+0xe8)[0x400004a33fa0] +[nid006550:188198] [10] /user-environment/env/default/libexec/osu-micro-benchmarks/mpi/collective/osu_iallreduce[0x404e98] +[nid006550:188198] *** End of error message *** +srun: error: nid006549: task 0: Segmentation fault (core dumped) +srun: Terminating StepId=2243671.21 +[2025-12-17T12:59:34.342] error: *** STEP 2243671.21 ON nid006549 CANCELLED AT 2025-12-17T12:59:34 DUE TO TASK FAILURE *** +srun: error: nid006550: task 2: Segmentation fault (core dumped) +srun: error: nid006550: task 3: Terminated +srun: error: nid006549: task 1: Terminated +srun: Force Terminated StepId=2243671.21 +```