Known issues

Julia module precompilation

If multiple MPI ranks trigger Julia's module precompilation, then a race condition can result in an error such as:

ERROR: LoadError: IOError: mkdir: file already exists (EEXIST)
Stacktrace:
 [1] uv_error at ./libuv.jl:97 [inlined]
 [2] mkdir(::String; mode::UInt16) at ./file.jl:177
 [3] mkpath(::String; mode::UInt16) at ./file.jl:227
 [4] mkpath at ./file.jl:222 [inlined]
 [5] compilecache_path(::Base.PkgId) at ./loading.jl:1210
 [6] compilecache(::Base.PkgId, ::String) at ./loading.jl:1240
 [7] _require(::Base.PkgId) at ./loading.jl:1029
 [8] require(::Base.PkgId) at ./loading.jl:927
 [9] require(::Module, ::Symbol) at ./loading.jl:922
 [10] include(::Module, ::String) at ./Base.jl:377
 [11] exec_options(::Base.JLOptions) at ./client.jl:288
 [12] _start() at ./client.jl:484

See julia issue #30174 for more discussion of this problem. There are similar issues with Pkg operations, see Pkg issue #1219.

This can be worked around be either:

  1. Triggering precompilation before launching MPI processes, for example:

    julia --project -e 'using Pkg; pkg"instantiate"'
    julia --project -e 'using Pkg; pkg"precompile"'
    mpiexec julia --project script.jl
  2. Launching julia with the --compiled-modules=no option. This can result in much longer package load times.

Open MPI

Segmentation fault when loading the library

When attempting to use a system-provided Open MPI implementation, you may encounter a segmentation fault upon loading the library, or whenever the value of an environment variable is requested. This can be fixed by setting the environment variable ZES_ENABLE_SYSMAN=1. See Open MPI issue #10142 for more details.

Segmentation fault in HCOLL

If Open MPI was built with support for HCOLL, you may encounter a segmentation fault in certain operations involving custom datatypes. The stacktrace may look something like

hcoll_create_mpi_type at /opt/mellanox/hcoll/lib/libhcoll.so.1 (unknown line)
ompi_dtype_2_hcoll_dtype at /lustre/software/openmpi/llvm14/4.1.4/lib/openmpi/mca_coll_hcoll.so (unknown line)
mca_coll_hcoll_allgather at /lustre/software/openmpi/llvm14/4.1.4/lib/openmpi/mca_coll_hcoll.so (unknown line)
MPI_Allgather at /lustre/software/openmpi/llvm14/4.1.4/lib/libmpi.so (unknown line)

This is due to a bug in HCOLL, see Open MPI issue #11201 for more details. You can disable HCOLL by exporting the environment variable

export OMPI_MCA_coll_hcoll_enable="0"

before starting the MPI process.

UCX

UCX is a communication framework used by several MPI implementations.

Memory cache

When used with CUDA, UCX intercepts cudaMalloc so it can determine whether the pointer passed to MPI is on the host (main memory) or the device (GPU). Unfortunately, there are several known issues with how this works with Julia:

By default, MPI.jl disables this by setting

ENV["UCX_MEMTYPE_CACHE"] = "no"

at __init__ which may result in reduced performance, especially for smaller messages.

Multi-threading and signal handling

When using Julia multi-threading, the Julia garbage collector internally uses SIGSEGV to synchronize threads.

By default, UCX will error if this signal is raised (#337), resulting in a message such as:

Caught signal 11 (Segmentation fault: invalid permissions for mapped object at address 0xXXXXXXXX)

This signal interception can be controlled by setting the environment variable UCX_ERROR_SIGNALS: if not already defined, MPI.jl will set it as:

ENV["UCX_ERROR_SIGNALS"] = "SIGILL,SIGBUS,SIGFPE"

at __init__. If set externally, it should be modified to exclude SIGSEGV from the list.

CUDA-aware MPI

Memory pool

Using CUDA-aware MPI on multi-GPU nodes with recent CUDA.jl may trigger (see here)

The call to cuIpcGetMemHandle failed. This means the GPU RDMA protocol
cannot be used.
  cuIpcGetMemHandle return value:   1

in the MPI layer, or fail on a segmentation fault (see here) with

[1642930332.032032] [gcn19:4087661:0] gdr_copy_md.c:122 UCX ERROR gdr_pin_buffer failed. length :65536 ret:22

This is due to the MPI implementation using legacy cuIpc* APIs, which are incompatible with stream-ordered allocator, now default in CUDA.jl, see UCX issue #7110.

To circumvent this, one has to ensure the CUDA memory pool to be set to none:

export JULIA_CUDA_MEMORY_POOL=none

More about CUDA.jl memory environment-variables.

Hints to ensure CUDA-aware MPI to be functional

Make sure to:

  • Have MPI and CUDA on path (or module loaded) that were used to build the CUDA-aware MPI
  • Set the following environment variables: export JULIA_CUDA_MEMORY_POOL=none export JULIA_CUDA_USE_BINARYBUILDER=false
  • Add CUDA, MPIPreferences, and MPI packages in Julia. Switch to using the system binary
    julia --project -e 'using Pkg; Pkg.add(["CUDA", "MPIPreferences", "MPI"]); using MPIPreferences; MPIPreferences.use_system_binary()'
  • Then in Julia, upon loading MPI and CUDA modules, you can check
    • CUDA version: CUDA.versioninfo()
    • If MPI has CUDA: MPI.has_cuda()
    • If you are using correct MPI library: MPI.libmpi

After that, it may be preferred to run the Julia MPI script (as suggested here) launching it from a shell script (as suggested here).

ROCm-aware MPI

Hints to ensure ROCm-aware MPI to be functional

Make sure to:

  • Have MPI and ROCm on path (or module loaded) that were used to build the ROCm-aware MPI
  • Add AMDGPU, MPIPreferences, and MPI packages in Julia:
    julia --project -e 'using Pkg; Pkg.add(["AMDGPU", "MPIPreferences", "MPI"]); using MPIPreferences; MPIPreferences.use_system_binary()'
  • Then in Julia, upon loading MPI and CUDA modules, you can check
    • AMDGPU version: AMDGPU.versioninfo()
    • If you are using correct MPI implementation: MPI.identify_implementation()

After that, this script can be used to verify if ROCm-aware MPI is functional (modified after the CUDA-aware version from here). It may be preferred to run the Julia ROCm-aware MPI script launching it from a shell script (as suggested here).

Custom reduction operators

It is not possible to use custom reduction operators with 32-bit Microsoft MPI on Windows and on ARM CPUs with any operating system. These issues are due to due how custom operators are currently implemented in MPI.jl, that is by using closure cfunctions. However they have two limitations:

  • Julia's C-compatible function pointers cannot be used where the stdcall calling convention is expected, which is the case for 32-bit Microsoft MPI,
  • closure cfunctions in Julia are based on LLVM trampolines, which are not supported on ARM architecture.