我试图在不同节点上获取多个不同长度的列表，在一个节点中将它们收集在一起，并让主节点将它们放在一个集合中。此列表在每个节点中命名为rout_array。请注意，rout_array中的元素只是整数，而且在节点之间是唯一的。
Traceback (most recent call last):
File "prout.py", line 160, in <module>
main()
File "prout.py", line 153, in main
num = DetermineRoutingNumber(steps, goal, vertexSetSize)
File "prout.py", line 129, in DetermineRoutingNumber
comm.Gather(send_buffer, recv_buffer, root = 0)
File "MPI\Comm.pyx", line 589, in mpi4py.MPI.Comm.Gather (c:\projects\mpi4py\src\mpi4py.MPI.c:97806)
File "MPI\msgbuffer.pxi", line 525, in mpi4py.MPI._p_msg_cco.for_gather (c:\projects\mpi4py\src\mpi4py.MPI.c:34678)
File "MPI\msgbuffer.pxi", line 446, in mpi4py.MPI._p_msg_cco.for_cco_send (c:\projects\mpi4py\src\mpi4py.MPI.c:33938)
File "MPI\msgbuffer.pxi", line 148, in mpi4py.MPI.message_simple (c:\projects\mpi4py\src\mpi4py.MPI.c:30349)
File "MPI\msgbuffer.pxi", line 93, in mpi4py.MPI.message_basic (c:\projects\mpi4py\src\mpi4py.MPI.c:29448)
KeyError: 'O'
当我的代码中没有字符串时，我不知道如何获得'O'的KeyError。所有列表都包含整数，numpy数组包含整数，此处唯一活动的字典只有键的整数。应该注意，每个节点都输出这个错误。
import numpy, math
from mpi4py import MPI
from sympy.combinatorics import Permutation as Perm
def GetEdges(size,file):
"""This function takes in a file of edges in a graph in the form 'u,v'
without quotes, where u and v are vertices of the graph. It then
generates a permutation that swaps those vertices, and returns these
transpositions."""
edgeFile = open(file, "r")
edges = []
for line in edgeFile:
line = line.strip()
line = line.split(",")
for vertex in line:
line[line.index(vertex)] = int(vertex)
edges.append(Perm([line], size = size))
edgeFile.close()
edges.append(Perm([[size - 1]], size = size))
return edges
def AreDisjoint(p1,p2):
"""This function determines whether or not two permutations move any
common elements, and returns the appropriate boolean."""
v1 = set(p1.support())
v2 = set(p2.support())
return len(v1 & v2) == 0
def GetMatchings(edges, maxMatching, size):
"""This function takes in a set of edges given by GetEdges(), and
generates all possible matchings in the given graph. It then converts
each matching into its rank given by lexicographical order, and appends
that rank to a set, which is then returned."""
stepDict = {1:set(edges)}
steps = set(edges)
for i in range(1,maxMatching):
temp = set()
for p1 in stepDict[1]:
for p2 in stepDict[i]:
newPerm = p1 * p2
if AreDisjoint(p1,p2) and newPerm not in steps:
temp.add(newPerm)
steps.add(newPerm)
stepDict[i+1] = temp
newSteps = set()
for step in steps:
newSteps.add(step.rank())
return newSteps
def FromRank(rank,level):
"""This function takes in a rank and size of a permutation, then returns
the permutation that lies at the rank according to lexicographical
ordering. """
lst = list(range(level + 1))
perm = []
while lst:
fact = math.factorial(len(lst) - 1)
index, rank = divmod(rank, fact)
perm.append(lst.pop(index))
assert rank == 0
return perm
def SplitArrayBetweenNodes(rank, rem, length):
"""This function takes in the rank of a node and any remainder after
dividing up an array between all the nodes. It then returns a starting
and ending partition index unique to each node."""
if rem != 0:
if rank in list(range(rem)):
if rank == 0:
part_start = 0
part_end = length
else:
part_start = rank * (length + 1)
part_end = part_start + length
else:
part_start = rank * length + rem
part_end = part_start + length - 1
else:
part_start = rank * length
part_end = part_start + length - 1
return part_start, part_end
def DetermineRoutingNumber(steps, goal, vertexSetSize):
"""This function takes in the matchings created by GetMatchings(),
and calculates all possible products between its own elements. It then
takes all unique products, and calculates all possible prducts between
the matching set and the previous output. This repeats until all
permutations of a given type are found. The level at which this occurs
is then returned."""
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
size = comm.Get_size()
length = len(steps)
rem = length % size
part_len = length // size
part_start, part_end = SplitArrayBetweenNodes(rank,rem, part_len)
permDict = {1: steps}
i = 1
while True:
rout_array = set()
work_array = set(list(permDict[i])[part_start:part_end + 1])
#Calculate all possible products
for p1 in permDict[1]:
for p2 in work_array:
p2_perm = Perm(FromRank(p2,vertexSetSize - 1))
p1_perm = Perm(FromRank(p1,vertexSetSize - 1))
new = p2_perm * p1_perm
if new(0) == 0 or new(0) == 1:
order = new.rank()
rout_array.add(order)
#All nodes send their work to master node
comm.Barrier()
send_buffer = numpy.array(rout_array)
sendcounts = numpy.array(comm.gather(len(rout_array), root = 0))
if rank == 0:
recv_buffer = numpy.empty(sum(sendcounts), dtype = int)
else:
recv_buffer = None
comm.Gatherv(sendbuf = send_buffer, recvbuf = (recv_buffer, sendcounts), root = 0)
#Generate input for next level of the loop, and weed out repeats.
permDict[i+1] = rout_array
for j in range(1,i+1):
permDict[i+1] = permDict[i+1] - permDict[j]
def main():
file = "EdgesQ2.txt"
maxMatching = 2
vertexSetSize = 4
edges = GetEdges(vertexSetSize, file)
steps = GetMatchings(edges, maxMatching, vertexSetSize)
goal = 2 * math.factorial(vertexSetSize-1)
num = DetermineRoutingNumber(steps, goal, vertexSetSize)
print(num)
main()
请注意，此示例中为maxMatching = 2和vertexSetSize = 4。输出应为3。
0,1
1,2
2,3
0,3
请注意，此示例中为maxMatching = 4和vertexSetSize = 8。输出应为4。
0,1
0,3
0,4
1,2
1,5
2,3
2,6
3,7
4,5
4,7
5,6
6,7
答案 0 :(得分：2)
如果各个流程的长度不同，则需要使用向量变量Gatherv。使用该函数，您可以提供包含各种长度（recvcounts）的数组。
不幸的是，mpi4py文档目前没有描述如何使用Gatherv或任何其他矢量变体。这是一个简单的例子：
#!/usr/bin/env python3
import numpy as np
from mpi4py import MPI
import random
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
root = 0
local_array = [rank] * random.randint(2, 5)
print("rank: {}, local_array: {}".format(rank, local_array))
sendbuf = np.array(local_array)
# Collect local array sizes using the high-level mpi4py gather
sendcounts = np.array(comm.gather(len(sendbuf), root))
if rank == root:
print("sendcounts: {}, total: {}".format(sendcounts, sum(sendcounts)))
recvbuf = np.empty(sum(sendcounts), dtype=int)
else:
recvbuf = None
comm.Gatherv(sendbuf=sendbuf, recvbuf=(recvbuf, sendcounts), root=root)
if rank == root:
print("Gathered array: {}".format(recvbuf))
正如您所看到的，mpi4py不会将sendcounts或recvcounts作为额外参数，而是作为recvbuf参数的元组/列表。如果你传递(recvbuf, sendcounts)，它将从recvbuf派生出类型。将进行位移/偏移，使得来自所有等级的数据被连续存储并按等级排序。
基本上mpi4py会疯狂地猜测你可能对recvbuf参数的各种形式有什么意义。完整且明确的表单为(buffer, counts, displacements, type)。
KeyError：相当容易引起混淆的rout_array是set，它不是numpy.array的有效输入。 set既不是序列也不是数组接口。不幸的是，numpy.array创建了一个没有维度的非常奇怪的ndarray对象，而不是失败。您可以将数组创建包装在列表中：
send_buffer = numpy.array(list(rout_array))
集体有效，但考虑到return中的break循环中没有while true或DetermineRoutingNumber，循环不会终止，这并不奇怪。