在不使用 mpi4py 收集和分散的情况下检查所有排名是否正确

Question

我正在尝试在进程之间进行通信，以便在所有其他进程都准备好时通知每个进程。 下面的代码片段就是这样做的。 有没有更优雅的方法来做到这一点？

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

如果所有进程都需要知道哪些其他进程已准备好，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 在评论中指出的那样：

如果所有进程只需要知道是否所有进程都准备好了，那么 MPI_Allreduce() 更合适。

这个想法是理论上Allreduce应该比Allgather更快，因为前者可以使用树通信模式，并且因为它需要分配和通信更少的 memory。 更多信息可以在这里找到。

在您的情况下，您使用MPI.LAND （即逻辑与）作为 Allreduce 操作运算符。

一个例子：

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“第二次运行”之前），结果为FALSE ，因为具有偶数等级的进程未准备好（即False ），而具有奇数等级的进程已准备好。 因此， False & True => False 。 在第二部分中，结果为True ，因为所有进程都已准备好。