在将OpenMPI从1.8.4更新为2.0.2之后,我使用mpi_wtime()跑了错误的时间测量。使用版本1.8.4,结果与OMP_Get_wtime()计时器返回的结果相同,现在MPI_WTIME运行速度约2倍。
什么会导致这种行为?
我的示例代码:
#include <omp.h>
#include <mpi.h>
#include <stdio.h>
#include <stdlib.h>
int some_work(int rank, int tid){
int count = 10000;
int arr[count];
for( int i=0; i<count; i++)
arr[i] = i + tid + rank;
for( int val=0; val<4000000; val++)
for(int i=0; i<count-1; i++)
arr[i] = arr[i+1];
return arr[0];
}
int main (int argc, char *argv[]) {
MPI_Init(NULL, NULL);
int rank, size;
MPI_Comm_size(MPI_COMM_WORLD, &size);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
if (rank == 0)
printf("there are %d mpi processesn", size);
MPI_Barrier(MPI_COMM_WORLD);
double omp_time1 = omp_get_wtime();
double mpi_time1 = MPI_Wtime();
#pragma omp parallel
{
int tid = omp_get_thread_num();
if ( tid == 0 ) {
int nthreads = omp_get_num_threads();
printf("There are %d threads for process %dn", nthreads, rank);
int result = some_work(rank, tid);
printf("result for process %d thread %d is %dn", rank, tid, result);
}
}
MPI_Barrier(MPI_COMM_WORLD);
double mpi_time2 = MPI_Wtime();
double omp_time2 = omp_get_wtime();
printf("process %d omp time: %fn", rank, omp_time2 - omp_time1);
printf("process %d mpi time: %fn", rank, mpi_time2 - mpi_time1);
printf("process %d ratio: %fn", rank, (mpi_time2 - mpi_time1)/(omp_time2 - omp_time1) );
MPI_Finalize();
return EXIT_SUCCESS;
}
编译
g++ -O3 src/example_main.cpp -o bin/example -fopenmp -I/usr/mpi/gcc/openmpi-2.0.2/include -L /usr/mpi/gcc/openmpi-2.0.2/lib -lmpi
和运行
salloc -N2 -n2 mpirun --map-by ppr:1:node:pe=16 bin/example
给出
之类的东西there are 2 mpi processes
There are 16 threads for process 0
There are 16 threads for process 1
result for process 1 thread 0 is 10000
result for process 0 thread 0 is 9999
process 1 omp time: 5.066794
process 1 mpi time: 10.098752
process 1 ratio: 1.993125
process 0 omp time: 5.066816
process 0 mpi time: 8.772390
process 0 ratio: 1.731342
我首先写的比率并不一致,但仍然足够大。
OpenMPI 1.8.4的结果还可以:
g++ -O3 src/example_main.cpp -o bin/example -fopenmp -I/usr/mpi/gcc/openmpi-1.8.4/include -L /usr/mpi/gcc/openmpi-1.8.4/lib -lmpi -lmpi_cxx
给出
result for process 0 thread 0 is 9999
result for process 1 thread 0 is 10000
process 0 omp time: 4.655244
process 0 mpi time: 4.655232
process 0 ratio: 0.999997
process 1 omp time: 4.655335
process 1 mpi time: 4.655321
process 1 ratio: 0.999997
我的群集上有类似的行为(与您的OpenMPI版本相同,2.0.2),问题是CPU频率的默认调速器,即"保守"。将调速器设置为"绩效"后,MPI_WTIME()的输出与正确的时间对齐(在我的情况下,"时间"的输出)。看来,对于某些较旧的Xeon处理器(例如Xeon E5620),当使用过于激进的策略以进行动态频率调整时,某些时钟功能会偏向 - 相同的OpenMPI版本不会在同一集群中的较新Xeon上遇到此问题。
也许mpi_wtime()本身可能是一个昂贵的操作?如果您避免在OpenMP时间的一部分中测量MPI_WTIME()消耗的时间,结果是否会变得更加一致?例如:
double mpi_time1 = MPI_Wtime();
double omp_time1 = omp_get_wtime();
/* do something */
double omp_time2 = omp_get_wtime();
double mpi_time2 = MPI_Wtime();