Segfault fix

.github/workflows/CI.yml

@@ -1,17 +1,22 @@
 name: CI
+
 on:
   push:
     branches:
       - main
     tags: ['*']
   pull_request:
+
 concurrency:
   # Skip intermediate builds: always.
   # Cancel intermediate builds: only if it is a pull request build.
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: ${{ startsWith(github.ref, 'refs/pull/') }}
+
 jobs:
-  test:
+
+  # default test
+  test-MPICH-jll:
     name: Julia ${{ matrix.version }} - ${{ matrix.os }} - ${{ matrix.arch }} - ${{ github.event_name }}
     runs-on: ${{ matrix.os }}
     strategy:
@@ -27,7 +32,7 @@ jobs:
         arch:
           - x64
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v3
       - uses: actions/setup-java@v3
         with:
           distribution: 'temurin'
@@ -43,13 +48,155 @@ jobs:
       - uses: codecov/codecov-action@v2
         with:
           files: lcov.info
+
+
+  # test OpenMPI by requesting it with MPIPreferences
+  # adapted from MPI.jl
+  test-OpenMPI-jll:
+    name: Julia OpenMPI_jll - ${{ github.event_name }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        version:
+          - '1.8'
+        os:
+          - ubuntu-latest
+        arch:
+          - x64
+
+      fail-fast: false
+    env:
+      JULIA_MPI_TEST_BINARY: OpenMPI_jll
+      JULIA_MPI_TEST_ABI: OpenMPI
+    steps:
+    - name: Checkout
+      uses: actions/checkout@v3
+
+    - uses: julia-actions/setup-julia@latest
+      with:
+        version: ${{ matrix.version }}
+        arch: ${{ matrix.arch }}
+    - uses: julia-actions/cache@v1
+
+    - name: use OpenMPI_jll
+      shell: julia --color=yes --project=test {0}
+      run: |
+        using Pkg
+        Pkg.instantiate()
+        using MPIPreferences
+        MPIPreferences.use_jll_binary("OpenMPI_jll")
+        rm("test/Manifest.toml")
+
+    - uses: julia-actions/julia-runtest@latest
+
+  # test system MPI using Brew in macOS
+  # adapted from MPI.jl
+  test-system-MPI-brew:
+    name: Julia ${{ matrix.version }} - ${{ matrix.os }} - ${{ matrix.mpi }} - ${{ github.event_name }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        version:
+          - '1.8'
+        os:
+          - macos-latest
+        mpi:
+          - mpich
+          - openmpi
+    env:
+      JULIA_MPI_TEST_BINARY: system
+      ZES_ENABLE_SYSMAN: 1  # https://github.com/open-mpi/ompi/issues/10142
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Install MPI via homebrew
+        run: brew install $MPI
+        env:
+          MPI: ${{ matrix.mpi }}
+
+      - uses: actions/setup-java@v3
+        with:
+          distribution: 'temurin'
+          java-version: '11'
+
+      - uses: julia-actions/setup-julia@v1
+        with:
+          version: ${{ matrix.version }}
+          arch: x64
+
+      - uses: julia-actions/cache@v1
+
+      - name: use system MPI
+        shell: julia --color=yes --project=test {0}
+        run: |
+          using Pkg
+          Pkg.instantiate()
+          using MPIPreferences
+          MPIPreferences.use_system_binary()
+          run(`sed -i.bu 's/unknown/MPICH/' test/LocalPreferences.toml`) # fix wrong abi detection for mpich
+          rm("test/Manifest.toml")
+
+      - uses: julia-actions/julia-runtest@v1
+
+  # # test system MPI using apt in ubuntu
+  # # adapted from MPI.jl
+  # # TODO: commented out because apt has older versions of MPICH and OMPI that
+  # # segfault with multithreading. Re-introduce them when apt pkgs are upgraded 
+  # test-system-MPI-apt:
+  #   name: Julia ${{ matrix.version }} - ${{ matrix.os }} - ${{ matrix.mpi }} - ${{ github.event_name }}
+  #   runs-on: ${{ matrix.os }}
+  #   strategy:
+  #     fail-fast: false
+  #     matrix:
+  #       version:
+  #         - '1.8'
+  #       os:
+  #         - ubuntu-latest
+  #       mpi:
+  #         - libmpich-dev
+  #   env:
+  #     JULIA_MPI_TEST_BINARY: system
+  #   steps:
+  #     - uses: actions/checkout@v3
+
+  #     - name: Install MPI via apt
+  #       run: |
+  #         sudo apt-get update
+  #         sudo apt-get install $MPI
+  #       env:
+  #         MPI: ${{ matrix.mpi }}
+
+  #     - uses: actions/setup-java@v3
+  #       with:
+  #         distribution: 'temurin'
+  #         java-version: '11'
+
+  #     - uses: julia-actions/setup-julia@v1
+  #       with:
+  #         version: ${{ matrix.version }}
+  #         arch: x64
+
+  #     - uses: julia-actions/cache@v1
+
+  #     - name: use system MPI
+  #       shell: julia --color=yes --project=test {0}
+  #       run: |
+  #         using Pkg
+  #         Pkg.instantiate()
+  #         using MPIPreferences
+  #         MPIPreferences.use_system_binary()
+  #         rm("test/Manifest.toml")
+
+  #     - uses: julia-actions/julia-runtest@v1
+
   docs:
     name: Documentation
     runs-on: ubuntu-latest
     permissions:
       contents: write
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v3
       - uses: actions/setup-java@v3
         with:
           distribution: 'temurin'

.gitignore

@@ -7,7 +7,8 @@
 .vscode/settings.json
 build
 .interfaces.md
-
+*.log
+*.err
 machines.txt
 results
 .includes_bu.jl

Project.toml

@@ -52,9 +52,3 @@ SpecialFunctions = "2"
 SplittableRandoms = "0.1"
 StatsBase = "0.33"
 julia = "1.6"
-
-[extras]
-Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
-
-[targets]
-test = ["Test"]

src/Pigeons.jl

@@ -8,7 +8,7 @@ import MPI: Comm, Allreduce, Comm_rank,
             Comm_dup, Request, Waitall,
             RequestSet, mpiexec, Allreduce, 
             Allgather, Comm_split, isend, recv,
-            bcast, tag_ub
+            bcast, tag_ub 
 
 
 using Base: Forward

src/mpi_utils/Entangler.jl

@@ -73,7 +73,7 @@ mutable struct Entangler
                 println("Entangler initialized 1 process (without MPI); $(Threads.nthreads())")
             end
         else
-            Init(threadlevel = :funneled) 
+            init_mpi()
             comm = Comm_dup(parent_communicator)
             transmit_counter_bound = ceil(Int, tag_ub() / n_global_indices - 2)
             my_process_index = Comm_rank(comm) + 1
@@ -101,10 +101,12 @@ mpi_active() =
     if silence_mpi[] 
         false
     else 
-        Init(threadlevel = :funneled)
+        init_mpi()
         Comm_size(COMM_WORLD) > 1
     end
 
+init_mpi() = Init(threadlevel = :funneled)
+
 """
 $SIGNATURES
 
@@ -155,6 +157,8 @@ function transmit!(e::Entangler, source_data::AbstractVector{T}, to_global_indic
     # indicators of whether each local index is to be received over MPI
     e.current_received_bits .= true 
     at_least_one_mpi = false
+
+    requests = RequestSet() # non-blocking requests that will be waited on
 
     # send (or copy if local)
     for local_index in 1:myload
@@ -172,13 +176,15 @@ function transmit!(e::Entangler, source_data::AbstractVector{T}, to_global_indic
             source_view = Ref{T}(source_datum)
             mpi_rank = process_index - 1
             # asynchronously (non-blocking) send over MPI:
-            Isend(source_view, e.communicator, dest = mpi_rank, tag = tag(e, transmit_index, global_index))
+            # note: we wait for the Isend request to avoid the application 
+            # terminating in the last iteration without completing its request.
+            request = Isend(source_view, e.communicator, dest = mpi_rank, tag = tag(e, transmit_index, global_index))
+            push!(requests, request)
         end
     end
 
     # receive
     if at_least_one_mpi
-        requests = RequestSet()
         my_globals = my_global_indices(e.load)
         for local_index in 1:myload
             if e.current_received_bits[local_index]
@@ -238,6 +244,8 @@ function reduce_deterministically(operation, source_data::AbstractVector{T}, e::
     # outer loop is over the levels of a binary tree over the global indices
     iteration = 1
 
+    requests = RequestSet()
+
     while n_remaining_to_reduce > 1
         transmit_index = next_transmit_index!(e)
         current_local = my_first_remaining_local
@@ -251,7 +259,8 @@ function reduce_deterministically(operation, source_data::AbstractVector{T}, e::
                 dest_global_index = current_global - spacing 
                 dest_process = find_process(e.load, dest_global_index)
                 dest_rank = dest_process - 1
-                isend(work_array[current_local], e.communicator; dest = dest_rank, tag = tag(e, transmit_index, iteration))
+                request = isend(work_array[current_local], e.communicator; dest = dest_rank, tag = tag(e, transmit_index, iteration))
+                push!(requests, request)
                 current_local += spacing           
                 did_send = true     
             elseif current_global + spacing ≤ e.load.n_global_indices
@@ -274,6 +283,7 @@ function reduce_deterministically(operation, source_data::AbstractVector{T}, e::
 
         if did_send 
             my_first_remaining_local += spacing
+            Waitall(requests)
         end
         n_global_indices_remaining_before = ceil(Int, n_global_indices_remaining_before/2)
         spacing = spacing * 2

src/submission/ChildProcess.jl

@@ -41,6 +41,11 @@ $FIELDS
     When wait is false, the process' I/O streams are directed to devnull.
     """
     wait = true
+
+    """
+    Extra arguments passed to mpiexec.
+    """
+    mpiexec_args::String = ""
 end 
 
 """
@@ -62,21 +67,28 @@ function pigeons(pt_arguments, new_process::ChildProcess)
         run(julia_cmd, wait = new_process.wait)
     else
         mpiexec() do exe
-            mpi_cmd = `$exe -n $(new_process.n_local_mpi_processes)`
-            cmd = `$mpi_cmd $julia_cmd`
+            args    = new_process.mpiexec_args
+            mpi_cmd = length(args)>0 ? `$exe $args` : `$exe` # need this because `$("")` == `''` != `` 
+            mpi_cmd = `$mpi_cmd -n $(new_process.n_local_mpi_processes)`
+            cmd     = `$mpi_cmd $julia_cmd`
             run(cmd, wait = new_process.wait)
         end
     end
     return Result{PT}(exec_folder)
 end
 
 function launch_cmd(pt_arguments, exec_folder, dependencies, n_threads::Int, silence_mpi::Bool)
-    julia_bin = Base.julia_cmd()
-    script_path = launch_script(pt_arguments, exec_folder, dependencies, silence_mpi)
-    return `$julia_bin 
-            --project   
-            --threads=$n_threads 
-            $script_path`
+    script_path  = launch_script(pt_arguments, exec_folder, dependencies, silence_mpi)
+    jl_cmd       = Base.julia_cmd()
+    project_file = Base.current_project()
+    if !isnothing(project_file)
+        # instantiate the project to make sure dependencies exist
+        # also, precompile to avoid issues with coordinating access to compile cache
+        project_dir = dirname(project_file)
+        jl_cmd      = `$jl_cmd --project=$project_dir`
+        run(`$jl_cmd -e "using Pkg; Pkg.instantiate(); Pkg.precompile()"`)
+    end
+    return `$jl_cmd --threads=$n_threads $script_path`
 end
 
 function launch_script(pt_arguments, exec_folder, dependencies, silence_mpi)
@@ -129,7 +141,6 @@ function launch_code(
 
     Pigeons.deserialize_immutables(raw"$path_to_serialized_immutables")
     pt_arguments = deserialize(raw"$path_to_serialized_pt_arguments")
-
     pt = PT(pt_arguments, exec_folder = raw"$exec_folder")
     pigeons(pt)
     """

src/submission/MPI.jl

@@ -41,6 +41,11 @@ $FIELDS
     process. 
     """
     dependencies::Vector{Module} = []
+
+    """
+    Extra arguments passed to mpiexec.
+    """
+    mpiexec_args::String = ""
 end
 
 """
@@ -96,7 +101,7 @@ function mpi_submission_script(exec_folder, mpi_submission::MPI, julia_cmd)
     #PBS -e $info_folder/stderr.txt
     cd \$PBS_O_WORKDIR
     $(modules_string(mpi_settings))
-    mpiexec --merge-stderr-to-stdout --output-filename $exec_folder $julia_cmd_str
+    mpiexec $(mpi_submission.mpiexec_args) --merge-stderr-to-stdout --output-filename $exec_folder $julia_cmd_str
     """
     script_path = "$exec_folder/.submission_script.sh"
     write(script_path, code)

src/utils/misc.jl

@@ -97,20 +97,6 @@ is attempted).
 """
 macro abstract() quote error("Attempted to call an abstract function.") end end
 
-function mpi_test(n_processes::Int, test_file::String; options = [])
-    project_folder = dirname(Base.current_project())
-    # handle 2 different "modes" that tests can be ran (for julia 1.0,1.1 vs. >1.1)
-    resolved_test_file = 
-        if isfile("$project_folder/$test_file")
-            "$project_folder/$test_file" 
-        else
-            "$project_folder/test/$test_file"
-        end
-    mpiexec() do exe
-        run(`$exe -n $n_processes $(Base.julia_cmd()) --project=$project_folder $resolved_test_file $options`)
-    end
-end
-
 
 """
     @weighted(w, x)