在不使用 mpi4py 收集和分散的情況下檢查所有排名是否正確

Question

我正在嘗試在進程之間進行通信，以便在所有其他進程都准備好時通知每個進程。 下面的代碼片段就是這樣做的。 有沒有更優雅的方法來做到這一點？

def get_all_ready_status(ready_batch):
    all_ready= all(ready_batch)
    return [all_ready for _ in ready_batch]

ready_batch= comm.gather(ready_agent, root=0)
if rank == 0:
    all_ready_batch = get_all_ready_status(ready_batch)
all_ready_flag = comm.scatter(all_ready_batch , root=0)                

Answer 1

如果所有進程都需要知道哪些其他進程已准備好，那么您可以使用comm.Allgather例程：

from mpi4py import MPI
import numpy


comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()

sendBuffer = numpy.ones(1, dtype=bool)
recvBuffer = numpy.zeros(size, dtype=bool)

print("Before Allgather => Process %s | sendBuffer %s | recvBuffer %s" % (rank, sendBuffer, recvBuffer))
comm.Allgather([sendBuffer,  MPI.BOOL],[recvBuffer, MPI.BOOL])
print("After Allgather  => Process %s | sendBuffer %s | recvBuffer %s" % (rank, sendBuffer, recvBuffer))

Output：

Before Allgather => Process 0 | sendBuffer [ True] | recvBuffer [False False]
Before Allgather => Process 1 | sendBuffer [ True] | recvBuffer [False False]
After Allgather  => Process 0 | sendBuffer [ True] | recvBuffer [ True  True]
After Allgather  => Process 1 | sendBuffer [ True] | recvBuffer [ True  True]

正如@Gilles Gouaillardet 在評論中指出的那樣：

如果所有進程只需要知道是否所有進程都准備好了，那么 MPI_Allreduce() 更合適。

這個想法是理論上Allreduce應該比Allgather更快，因為前者可以使用樹通信模式，並且因為它需要分配和通信更少的 memory。 更多信息可以在這里找到。

在您的情況下，您使用MPI.LAND （即邏輯與）作為 Allreduce 操作運算符。

一個例子：

from mpi4py import MPI
import numpy


comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()

sendBuffer =  numpy.ones(1, dtype=bool) if rank % 2 ==  0 else numpy.zeros(1, dtype=bool)
recvBuffer = numpy.zeros(1, dtype=bool)

print("Before Allreduce => Process %s | sendBuffer %s | recvBuffer %s" % (rank, sendBuffer, recvBuffer))
comm.Allreduce([sendBuffer,  MPI.BOOL],[recvBuffer, MPI.BOOL], MPI.LAND)
print("After Allreduce  => Process %s | sendBuffer %s | recvBuffer %s" % (rank, sendBuffer, recvBuffer))

comm.Barrier()
if rank == 0:
   print("Second RUN")
comm.Barrier()

sendBuffer =  numpy.ones(1, dtype=bool)
recvBuffer = numpy.zeros(1, dtype=bool)

print("Before Allreduce => Process %s | sendBuffer %s | recvBuffer %s" % (rank, sendBuffer, recvBuffer))
comm.Allreduce([sendBuffer,  MPI.BOOL],[recvBuffer, MPI.BOOL], MPI.LAND)
print("After Allreduce  => Process %s | sendBuffer %s | recvBuffer %s" % (rank, sendBuffer, recvBuffer))

Output：

Before Allreduce => Process 1 | sendBuffer [False] | recvBuffer [False]
Before Allreduce => Process 0 | sendBuffer [ True] | recvBuffer [False]
After Allreduce  => Process 1 | sendBuffer [False] | recvBuffer [False]
After Allreduce  => Process 0 | sendBuffer [ True] | recvBuffer [False]
Second RUN
Before Allreduce => Process 0 | sendBuffer [ True] | recvBuffer [False]
Before Allreduce => Process 1 | sendBuffer [ True] | recvBuffer [False]
After Allreduce  => Process 0 | sendBuffer [ True] | recvBuffer [ True]
After Allreduce  => Process 1 | sendBuffer [ True] | recvBuffer [ True]

在 output 的第一部分（即“第二次運行”之前），結果為FALSE ，因為具有偶數等級的進程未准備好（即False ），而具有奇數等級的進程已准備好。 因此， False & True => False 。 在第二部分中，結果為True ，因為所有進程都已准備好。