如何在mpi4py中分散和收集对象的python列表

Question

I have a list of 100,000 python objects that I would like to scatter and gather in mpi4py .

When I try with 8 processors I get:

SystemError: Negative size passed to PyBytes_FromStringAndSize

on the scattering.

When I try with 64 processors I get the same error but on the gather.

When I try making an array of objects out of the list and use Gather and Scatter, I get an error which basically states the dtype of the array cannot be object.

Any way I can get this to work? Or anything else I can use other than MPI?
I'm running this on an 8-node, 64-ppn computer.

Answer 1

Using scatter and gather, an example of splitting a numpy array with 100000 items.

import numpy as np
from mpi4py import MPI
from pprint import pprint
comm = MPI.COMM_WORLD

pprint("-" * 78)
pprint(" Running on %d cores" % comm.size)
pprint("-" * 78)

N = 100000
my_N = N // 8

if comm.rank == 0:
    A = np.arange(N, dtype=np.float64)
else:
    A = np.empty(N, dtype=np.float64)

my_A = np.empty(my_N, dtype=np.float64)

# Scatter data 
comm.Scatter([A, MPI.DOUBLE], [my_A, MPI.DOUBLE])

pprint("After Scatter:")
for r in range(comm.size):
    if comm.rank == r:
        print("[%d] %s" % (comm.rank, len(my_A)))
    comm.Barrier()

# Allgather data into A
comm.Allgather([my_A, MPI.DOUBLE], [A, MPI.DOUBLE])

pprint("After Allgather:")
for r in range(comm.size):
    if comm.rank == r:
        print("[%d] %s" % (comm.rank, len(A)))
    comm.Barrier()

Also you could check scatterv and gatherv , more examples here and here .

Answer 2

I'm not sure this is the answer, and also I'm not sure you are still looking for the answer, but...

So you have 100,000 python objects . If these objects are regular data (data sets), not an instance of some class, pass data as json string. Something like this:

#!/usr/bin/env python

import json
import numpy as np
from mpi4py import MPI


comm = MPI.COMM_WORLD

if comm.rank == 0:
    tasks = [
        json.dumps( { 'a':1,'x':2,'b':3 } ),
        json.dumps( { 'a':3,'x':1,'b':2 } ),
        json.dumps( { 'a':2,'x':3,'b':1 } )
    ]
else:
    tasks = None


# Scatter paramters arrays
unit = comm.scatter(tasks, root=0)

p = json.loads(unit)
print "-"*18
print("-- I'm rank %d in %d size task" % (comm.rank,comm.size) )
print("-- My paramters are: {}".format(p))
print "-"*18

comm.Barrier()

calc = p['a']*p['x']**2+p['b']

# gather results
result = comm.gather(calc, root=0)
# do something with result

if comm.rank == 0:
    print "the result is ", result
else:
    result = None

note, that if you have only 8 nodes/cores, you have to create 8 records in the tasks list and sequentially scatter and gather all 100,000 data sets. If all your data set is in ALLDATA list, the code could look like this:

def calc(a=0,x=0,b=0):
    return a*x**2+b

if comm.rank == 0: collector = []
for xset in zip(*(iter(ALLDATA),) * comm.size):
    task = [ json.dumps(s) for s in xset ]
    comm.Barrier()
    unit = comm.scatter(task if comm.rank == 0 else None, root=0)
    p = json.loads(unit)
    res = json.dumps( calc(**p) )
    totres = comm.gather(res, root=0)
    if comm.rank == 0:
        collector += [ json.loads(x) for x in  totres  ]



if comm.rank == 0:
    print "the result is ", collector