Merge pull request #17 from JuliaParallel/sb/mpiexec

use MPI.mpiexec function

Author: simonbyrne

.appveyor.yml

@@ -1,8 +1,7 @@
 environment:
   matrix:
-  - julia_version: 1.0
-  - julia_version: 1.1
-  - julia_version: 1.2
+  - julia_version: 1.3
+  - julia_version: 1.4
   - julia_version: nightly
 
 platform:
@@ -26,14 +25,6 @@ notifications:
 
 install:
   - ps: iex ((new-object net.webclient).DownloadString("https://raw.githubusercontent.com/JuliaCI/Appveyor.jl/version-1/bin/install.ps1"))
-  - ps: (new-object net.webclient).DownloadFile(
-        "https://download.microsoft.com/download/A/E/0/AE002626-9D9D-448D-8197-1EA510E297CE/msmpisetup.exe",
-        "C:\projects\MSMpiSetup.exe")
-  - C:\projects\MSMpiSetup.exe -unattend -minimal
-  - set PATH=C:\Program Files\Microsoft MPI\Bin;%PATH% # can be removed in new MPI.jl
-  # This shouldn't typicallybe needed: the installer adds the directory to
-  # the PATH. For some reason this doesn't work on Appveyor
-  - set JULIA_MPIEXEC=C:\Program Files\Microsoft MPI\Bin\mpiexec.exe
 
 build_script:
   - echo "%JL_BUILD_SCRIPT%"

.travis.yml

@@ -1,15 +1,13 @@
 language: julia
-sudo: required
 dist: trusty
 
 os:
   - linux
   - osx
 
 julia:
-  - 1.0
-  - 1.1
-  - 1.2
+  - 1.3
+  - 1.4
   - nightly
 
 branches:
@@ -25,41 +23,9 @@ cache:
   - $HOME/.ccache
   - $HOME/.julia/registries # can we cache all of .julia?
 
-env:
-  - MPI_IMPL=mpich
-  - MPI_IMPL=openmpi
-#  - MPI_IMPL=intelmpi
-
 matrix:
   allow_failures:
   - julia: nightly
-  - os: osx # MPI.jl issue 262
-    env: MPI_IMPL=openmpi
-  exclude:
-  - os: osx
-    env: MPI_IMPL=intelmpi
-
-before_install:
-  - sh ./conf/travis-install-mpi.sh $MPI_IMPL
-  - |
-    if [ "$TRAVIS_OS_NAME" == "linux" ]; then
-      if [ "$MPI_IMPL" == "intelmpi" ]; then
-        source $HOME/intel/compilers_and_libraries/linux/mpi/intel64/bin/mpivars.sh release
-      else
-        export JULIA_MPI_PATH=$HOME/$MPI_IMPL;
-      fi
-    fi
-# following can be removed with next MPI.jl release.
-  - export PATH=$HOME/$MPI_IMPL/bin:$PATH
-  - export CC=mpicc
-  - export FC=mpif90
-# Work around OpenMPI attempting to create overly long temporary
-# file names - and erroring as a result
-  - export TMPDIR=/tmp
-# Work around MacOS/OpenMPI issue:
-#   https://github.com/open-mpi/ompi/issues/6518
-#   https://github.com/open-mpi/ompi/issues/5798
-  - export OMPI_MCA_btl=self,tcp
 
 jobs:
   include:

Project.toml

@@ -10,6 +10,7 @@ Serialization = "9e88b42a-f829-5b0c-bbe9-9e923198166b"
 Sockets = "6462fe0b-24de-5631-8697-dd941f90decc"
 
 [compat]
+MPI = "0.14"
 julia = "1"
 
 [extras]

src/MPIClusterManagers.jl

@@ -4,7 +4,6 @@ export MPIManager, launch, manage, kill, procs, connect, mpiprocs, @mpi_do, Tran
 
 using Distributed, Serialization
 import MPI
-const mpiexec = isdefined(MPI, :mpiexec_path) ? MPI.mpiexec_path : "mpiexec"
 
 include("mpimanager.jl")
 

src/mpimanager.jl

@@ -29,7 +29,6 @@ mutable struct MPIManager <: ClusterManager
     mode::TransportMode
 
     launched::Bool # Are the MPI processes running?
-    mpirun_cmd::Cmd # How to start MPI processes
     launch_timeout::Int # seconds
 
     initialized::Bool # All workers registered with us
@@ -51,10 +50,9 @@ mutable struct MPIManager <: ClusterManager
     receiving_done::Channel{Nothing}
 
     function MPIManager(; np::Integer = Sys.CPU_THREADS,
-                          mpirun_cmd::Cmd = `$mpiexec -n $np`,
                           launch_timeout::Real = 60.0,
                           mode::TransportMode = MPI_ON_WORKERS,
-                          master_tcp_interface::String="" )
+                        master_tcp_interface::String="" )
         mgr = new()
         mgr.np = np
         mgr.mpi2j = Dict{Int,Int}()
@@ -64,7 +62,6 @@ mutable struct MPIManager <: ClusterManager
         # Only start MPI processes for MPI_ON_WORKERS
         mgr.launched = mode != MPI_ON_WORKERS
         @assert MPI.Initialized() == mgr.launched
-        mgr.mpirun_cmd = mpirun_cmd
         mgr.launch_timeout = launch_timeout
 
         mgr.initialized = false
@@ -141,8 +138,10 @@ function Distributed.launch(mgr::MPIManager, params::Dict,
             end
             cookie = string(":cookie_",Distributed.cluster_cookie())
             setup_cmds = `import MPIClusterManagers\;MPIClusterManagers.setup_worker'('$(mgr.ip),$(mgr.port),$cookie')'`
-            mpi_cmd = `$(mgr.mpirun_cmd) $(params[:exename]) -e $(Base.shell_escape(setup_cmds))`
-            open(detach(mpi_cmd))
+            MPI.mpiexec() do mpiexec_cmd
+                mpi_cmd = `$mpiexec_cmd -n $(mgr.np) $(params[:exename]) -e $(Base.shell_escape(setup_cmds))`
+                open(detach(mpi_cmd))
+            end
             mgr.launched = true
         end
 

test/runtests.jl

@@ -1,42 +1,23 @@
 using Pkg
 pkg"precompile"
 
-using MPIClusterManagers: mpiexec
-using Test
+using Test, MPI
 
-# Code coverage command line options; must correspond to src/julia.h
-# and src/ui/repl.c
-const JL_LOG_NONE = 0
-const JL_LOG_USER = 1
-const JL_LOG_ALL = 2
-const coverage_opts =
-    Dict{Int, String}(JL_LOG_NONE => "none",
-                      JL_LOG_USER => "user",
-                      JL_LOG_ALL => "all")
+nprocs = clamp(Sys.CPU_THREADS, 2, 4)
 
-function runtests()
-    nprocs = clamp(Sys.CPU_THREADS, 2, 4)
-    exename = joinpath(Sys.BINDIR, Base.julia_exename())
-    extra_args = []
-    @static if !Sys.iswindows()
-        if occursin( "OpenRTE", read(`$mpiexec --version`, String))
-            push!(extra_args,"--oversubscribe")
-        end
-    end
-
-    coverage_opt = coverage_opts[Base.JLOptions().code_coverage]
-
-    jlexec = `$exename --code-coverage=$coverage_opt`
-    mpijlexec = `$mpiexec $extra_args -n $nprocs $jlexec`
-
-    @info "Testing: test_cman_julia.jl"
-    run(`$jlexec $(joinpath(@__DIR__, "test_cman_julia.jl"))`)
+@info "Testing: test_cman_julia.jl"
+run(`$(Base.julia_cmd()) $(joinpath(@__DIR__, "test_cman_julia.jl")) $nprocs`)
 
+if VERSION >= v"1.5"
     @info "Testing: test_cman_mpi.jl"
-    run(`$mpijlexec $(joinpath(@__DIR__, "test_cman_mpi.jl"))`)
-
-    @info "Testing: test_cman_tcp.jl"
-    run(`$mpijlexec $(joinpath(@__DIR__, "test_cman_tcp.jl"))`)
+    mpiexec() do cmd
+        run(`$cmd -n $nprocs $(Base.julia_cmd()) $(joinpath(@__DIR__, "test_cman_mpi.jl"))`)
+    end
+else
+    @warn "MPI_TRANSPORT_ALL broken on Julia versions < 1.5\nSee https://github.com/JuliaParallel/MPIClusterManagers.jl/issues/9"
 end
 
-runtests()
+@info "Testing: test_cman_tcp.jl"
+mpiexec() do cmd
+    run(`$cmd -n $nprocs $(Base.julia_cmd()) $(joinpath(@__DIR__, "test_cman_tcp.jl"))`)
+end

test/test_cman_julia.jl

@@ -2,16 +2,13 @@ using Test
 using MPIClusterManagers
 using Distributed
 
-using MPIClusterManagers: mpiexec
 import MPI
 
 # Start workers via `mpiexec` that communicate among themselves via MPI;
 # communicate with the workers via TCP
-if !Sys.iswindows() && occursin( "OpenRTE", open(f->read(f, String),`$mpiexec --version`))
-    mgr = MPIManager(np=4, mpirun_cmd=`$mpiexec --oversubscribe -n 4`)
-else
-    mgr = MPIManager(np=4)
-end
+nprocs = parse(Int, ARGS[1])
+
+mgr = MPIManager(np=nprocs)
 addprocs(mgr)
 
 refs = []