Python MPI.reduce不起作用

Question

from mpi4py import MPI
import json
import numpy as np
comm = MPI.COMM_WORLD
comm_rank = comm.Get_rank()
comm_size = comm.Get_size()
comm.Barrier()
FINAL_RESULT =[]
#Parent process
if (comm_rank == 0):
    print "There are %d cores" % comm_size
    with open('tinyTwitter.json','r') as f:
        coor = json.load(f)
    with open('melbGrid.json','r') as g:
        grid = json.load(g)

    ## Broadcast the coordinate
    Coordinate = []
    ## Loop through each item 
    n = len(coor)
    for i in range(0,n):
        ##Determine which grid the item is sent from
        y = coor[i]['json']['geo']['coordinates'][0] ## latitude
        x = coor[i]['json']['geo']['coordinates'][1] ## longtitude
        Coordinate.append([x,y])
    local_Coordinate = np.array(Coordinate)

    local_Grid_Info = [] ## A List to store the grid information
    ##Extrating the boundary coordinates of each grid
    size = len(grid['features'])
    for i in range(0,size):
        Grid_dict = {}
        Grid_id = grid['features'][i]['properties']['id']
        xmin = grid['features'][i]['properties']['xmin']
        xmax = grid['features'][i]['properties']['xmax']
        ymin = grid['features'][i]['properties']['ymin']
        ymax = grid['features'][i]['properties']['ymax']
        Grid_dict[Grid_id]={'xmin': xmin ,'xmax' : xmax , 'ymin': ymin , 'ymax' : ymax} ## [ {A1:{'xmin':...}},A2:{'xmin'},...]
        local_Grid_Info.append(Grid_dict)
    local_Grid_Info = np.array(local_Grid_Info)

#Broadcast the data to slaves   
local_Coordinate = comm.bcast(local_Coordinate if comm_rank ==0 else None,root = 0)
local_grid = comm.bcast(local_Grid_Info if comm_rank == 0 else None,root = 0)

#child process
if(comm_rank > 0):  
#print local_Coordinate       
    ## The result table
    local_Result = {}
    ## Initialize the result dictionary
    for i in local_grid:
        for grid_id in i:
            local_Result[grid_id] = 0          
    local_Result['A-Row'] = 0
    local_Result['B-Row'] = 0
    local_Result['C-Row'] = 0
    local_Result['D-Row'] = 0
    local_Result['Col-1'] = 0
    local_Result['Col-2'] = 0
    local_Result['Col-3'] = 0
    local_Result['Col-4'] = 0
    local_Result['Col-5'] = 0
##    print local_Result
    #Evenly split the input coordinate to different processes based on their rank
    start = (comm_rank - 1)*len(local_Coordinate)/(comm_size-1)
    end = (comm_rank)*len(local_Coordinate)/(comm_size-1)

    for i in range(start,end):
        x = local_Coordinate[i][0]
        y = local_Coordinate[i][1]
        #Determine the location of the point
        for one_grid in local_grid:  #index in a list to fetch a dic
            for (k,v) in one_grid.items(): # a dictionary
            # Single  Grid Calculation
                if(y >= v['ymin'] and y <= v['ymax'] and
                   x >= v['xmin'] and x <= v['xmax']):
                    ## Append to the local_Result dictionary
                    local_Result[k] = local_Result[k] + 1
                    break;
    local_Result['A-Row'] = local_Result['A1']+local_Result['A2']+local_Result['A3']+local_Result['A4']
    local_Result['B-Row'] = local_Result['B1']+local_Result['B2']+local_Result['B3']+local_Result['B4']
    local_Result['C-Row'] = local_Result['C1']+local_Result['C2']+local_Result['C3']+local_Result['C4']+local_Result['C4']
    local_Result['D-Row'] = local_Result['D3']+local_Result['D4']+ local_Result['D5']
    local_Result['Col-1'] = local_Result['A1'] + local_Result['B1'] + local_Result['C1']
    local_Result['Col-2'] = local_Result['A2'] + local_Result['B2'] + local_Result['C2']
    local_Result['Col-3'] = local_Result['A3'] + local_Result['B3'] + local_Result['C3'] + local_Result['D3']
    local_Result['Col-4'] = local_Result['A4'] + local_Result['B4'] + local_Result['C4'] + local_Result['D4']
    local_Result['Col-5'] = local_Result['C5'] + local_Result['D5']
    print local_Result
    r = [k for (v,k) in local_Result.items()]
    r = np.asarray(r)
    print r
    comm.Reduce(r, FINAL_RESULT, op=MPI.SUM, root = 0)
comm.Barrier()
print FINAL_RESULT


##
##for key in Result:
##    print key +":" + str(Result[key])   
##
##if (comm_rank == 0):
##    exit(0)

您好，我正在从 JSON 文件进行坐标提取，然后根据它们在地图上的地理位置进行分类。每个过程都旨在计算特定网格中有多少点，我希望主过程总结所有结果。但是，当我尝试在最后做comm.Reduce()时，我发现FINAL_RESULT是空白的。
例如，我运行mpiexec -np 2 python p.py， 我会得到 r 作为 [ 4 53 0 2 2 0 15 50 1 1 7 7 80 8 88 29 11 2 2 7 36 8 2 1 4]而mpiexec -np 3 python p.py给

[ 2 25  0  1  1  0  9 22  1  0  2  3 36  6 35 18  4  1  2  5 12  6  0  1  2]
[ 2 28  0  1  1  0  6 28  0  1  5  4 44  2 53 11  7  1  0  2 24  2  2  0  2].

FINAL_RESULT是[].我想知道我应该如何将两个过程的结果组合在一起，或者至少在FINAL_RESULT发回给主人谢谢

Answer 1

缩减调用位于if(comm_rank > 0):条件的主体内，这意味着至少有一个进程不参与集体缩减调用。碰巧这个进程不是别的，而是指定的根秩 0，因此没有任何东西存储在FINAL_RESULT中，因为根永远不会从工作进程接收任何内容。comm.Reduce语句应位于条件块之外。