在mpi4py中使用运算符归约列表并求和每个元素

Question

我正在写一个mpi的Python代码。例如，四个进程有以下数据：

data on procs0: [1, 2, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0]
data on procs1: [0, 0, 0, 4, 5, 6, 0, 0, 0, 0, 0, 0]
data on procs2: [0, 0, 0, 0, 0, 0, 7, 8, 9, 0, 0, 0]
data on procs3: [0, 0, 0, 0, 0, 0, 0, 0, 0, 10, 11, 12]

我想使用mpi4py库中的reduce函数，把数据汇总到procs0上，得到的结果是：

result on procs0: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]

我该如何使用mpi4py库的函数来实现呢？

编辑：上面是一个简单的特殊情况，不能使用集合，下面是另一个案例：

data on procs0: [1,0,0,0,0,0]
data on procs1: [0,2,0,0,0,0]
data on procs2: [0,0,0,3,0,0]
data on procs3: [0,0,0,0,4,5]

理想的结果必须是：

result on procs0: [1,2,0,3,4,5]

Answer 1

使用列表推导式和zip函数，来存储每一列的最大值。

In [1]: procs0=[1, 2, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0]    
In [2]: procs1=[0, 0, 0, 4, 5, 6, 0, 0, 0, 0, 0, 0]
In [3]: procs2=[0, 0, 0, 0, 0, 0, 7, 8, 9, 0, 0, 0]
In [4]: procs3=[0, 0, 0, 0, 0, 0, 0, 0, 0, 10, 11, 12]

In [5]: [max(i) for i in zip(procs0, procs1, procs2, procs3)]
Out[5]: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]

Answer 2

我不太确定你问的问题是想要数据的总和，还是想要最大值。我写了一个简单的例子，使用了mpi的Reduce函数，这个函数可以计算总和。

#!/usr/bin/env python
import numpy as np
from mpi4py import MPI
comm = MPI.COMM_WORLD

comm.Barrier()
t_start = MPI.Wtime()

# this array lives on each processor
data = np.zeros(5)
for i in xrange(comm.rank, len(data), comm.size):
    # set data in each array that is different for each processor
    data[i] = i

# print out the data arrays for each processor
print '[%i]'%comm.rank, data
comm.Barrier()

# the 'totals' array will hold the sum of each 'data' array
if comm.rank==0:
    # only processor 0 will actually get the data
    totals = np.zeros_like(data)
else:
    totals = None

# use MPI to get the totals 
comm.Reduce(
    [data, MPI.DOUBLE],
    [totals, MPI.DOUBLE],
    op = MPI.SUM,
    root = 0
)

# print out the 'totals'
# only processor 0 actually has the data
print '[%i]'%comm.rank, totals

comm.Barrier()
t_diff = MPI.Wtime() - t_start
if comm.rank==0: print t_diff

把这个代码保存到一个叫reduce_test.py的文件里，然后用命令mpirun -np 3 ./reduce_test.py运行，结果在我的机器上是这样的：

[0] [ 0.  0.  0.  3.  0.]
[1] [ 0.  1.  0.  0.  4.]
[2] [ 0.  0.  2.  0.  0.]
[1] None
[2] None
[0] [ 0.  1.  2.  3.  4.]
0.00260496139526

注意，如果你把调用comm.Reduce时的参数op = MPI.SUM改成op = MPI.MAX，那么就会计算最大值，而不是总和。