带有对 mpirun 的嵌入式调用的 Python

Question

我正在尝试使用 PyOpt 运行一些并行优化。棘手的部分是，在我的目标函数中，我还想使用 mpi 运行 C++ 代码。

我的 python 脚本如下:

#!/usr/bin/env python    
# Standard Python modules
import os, sys, time, math
import subprocess


# External Python modules
try:
    from mpi4py import MPI
    comm = MPI.COMM_WORLD
    myrank = comm.Get_rank()
except:
    raise ImportError('mpi4py is required for parallelization')

# Extension modules
from pyOpt import Optimization
from pyOpt import ALPSO

# Predefine the BashCommand
RunCprogram = "mpirun -np 2 CProgram" # Parallel C++ program


######################### 
def objfunc(x):

    f = -(((math.sin(2*math.pi*x[0])**3)*math.sin(2*math.pi*x[1]))/((x[0]**3)*(x[0]+x[1])))

    # Run CProgram 
    os.system(RunCprogram) #where the mpirun call occurs

    g = [0.0]*2
    g[0] = x[0]**2 - x[1] + 1
    g[1] = 1 - x[0] + (x[1]-4)**2

    time.sleep(0.01)
    fail = 0
    return f,g, fail

# Instantiate Optimization Problem 
opt_prob = Optimization('Thermal Conductivity Optimization',objfunc)
opt_prob.addVar('x1','c',lower=5.0,upper=1e-6,value=10.0)
opt_prob.addVar('x2','c',lower=5.0,upper=1e-6,value=10.0)
opt_prob.addObj('f')
opt_prob.addCon('g1','i')
opt_prob.addCon('g2','i')

# Solve Problem (DPM-Parallelization)
alpso_dpm = ALPSO(pll_type='DPM')
alpso_dpm.setOption('fileout',0)
alpso_dpm(opt_prob)
print opt_prob.solution(0)

我使用以下代码运行该代码:

mpirun -np 20 python Script.py

但是，我收到以下错误:

[user:28323] *** Process received signal ***
[user:28323] Signal: Segmentation fault (11)
[user:28323] Signal code: Address not mapped (1)
[user:28323] Failing at address: (nil)
[user:28323] [ 0] /lib64/libpthread.so.0() [0x3ccfc0f500]
[user:28323] *** End of error message ***

我发现 2 个不同的 mpirun 调用(调用 python 脚本的调用和脚本中的调用)相互冲突。关于如何解决这个问题的任何线索？

谢谢!!

Answer 1

参见 Calling mpi binary in serial as subprocess of mpi application : 最安全的方法是使用 MPI_Comm_spawn() .看看this manager-worker example例如。

一个快速的解决方法是使用 subprocess.Popen 作为@EdSmith 的信号。然而，请注意 subprocess.Popen 的默认行为使用父环境。我的猜测是 os.system() 也是一样的。不幸的是，一些环境变量是由 mpirun 添加的，具体取决于 MPI 实现，例如 OMPI_COMM_WORLD_RANK 或 OMPI_MCA_orte_ess_num_procs。要查看这些环境变量，请键入 import os ；在 mpi4py 代码和基本的 python shell 中打印 os.environ。这些环境变量可能会导致子流程失败。所以我不得不添加一行来摆脱它们......这很脏......归结为:

    args = shlex.split(RunCprogram)
    env=os.environ
    # to remove all environment variables with "MPI" in it...rather dirty...
    new_env = {k: v for k, v in env.iteritems() if "MPI" not in k}

    #print new_env
    # shell=True : watch for security issues...
    p = subprocess.Popen(RunCprogram,shell=True, env=new_env,stdout=subprocess.PIPE, stdin=subprocess.PIPE)
    p.wait()
    result="process myrank "+str(myrank)+" got "+p.stdout.read()
    print result

完整的测试代码，由 mpirun -np 2 python opti.py 运行:

#!/usr/bin/env python    
# Standard Python modules
import os, sys, time, math
import subprocess
import shlex


# External Python modules
try:
    from mpi4py import MPI
    comm = MPI.COMM_WORLD
    myrank = comm.Get_rank()
except:
    raise ImportError('mpi4py is required for parallelization')

# Predefine the BashCommand
RunCprogram = "mpirun -np 2 main" # Parallel C++ program


######################### 
def objfunc(x):

    f = -(((math.sin(2*math.pi*x[0])**3)*math.sin(2*math.pi*x[1]))/((x[0]**3)*(x[0]+x[1])))

    # Run CProgram 
    #os.system(RunCprogram) #where the mpirun call occurs
    args = shlex.split(RunCprogram)
    env=os.environ
    new_env = {k: v for k, v in env.iteritems() if "MPI" not in k}

    #print new_env
    p = subprocess.Popen(RunCprogram,shell=True, env=new_env,stdout=subprocess.PIPE, stdin=subprocess.PIPE)
    p.wait()
    result="process myrank "+str(myrank)+" got "+p.stdout.read()
    print result



    g = [0.0]*2
    g[0] = x[0]**2 - x[1] + 1
    g[1] = 1 - x[0] + (x[1]-4)**2

    time.sleep(0.01)
    fail = 0
    return f,g, fail

print objfunc([1.0,0.0])

Basic worker，由mpiCC main.cpp -o main编译:

#include "mpi.h"

int main(int argc, char* argv[]) { 
    int rank, size;

    MPI_Init (&argc, &argv);    
    MPI_Comm_rank (MPI_COMM_WORLD, &rank);  
    MPI_Comm_size (MPI_COMM_WORLD, &size);  

    if(rank==0){
        std::cout<<" size "<<size<<std::endl;
    }
    MPI_Finalize();

    return 0;

}