如何在循环中重用MPI_Scatter和MPI_Gather
我试图学习如何多次使用MPI_Scatter和MPI_Gather ，并在等待这两个MPI函数完成后打印出结果。 在进程0，程序的顶部，我想使用一个调用Scatter和Gather的while循环。 完成所有计算后，我想将此数组发送回这些函数以进行更多计算。 我在下面的代码中解释了我想要做什么。 /*.....*/中的注释是我想要实现的任务。
以下代码使用4个处理器运行。
:$ mpicc test.c -o test
:$ mpirun -np 4 test
#include #include #include int main(int argc, char **argv) { int size, rank; MPI_Init(&argc, &argv); MPI_Comm_size(MPI_COMM_WORLD, &size); MPI_Comm_rank(MPI_COMM_WORLD, &rank); int globaldata[8]; int localdata[2]; int counter, i; if (rank == 0) { for (i=0; i<size*2; i++)//initializing array to all zeros, one time globaldata[i] = 0; /*counter=0; do { counter++; */ printf("Sending at Processor %d has data: ", rank); for (i=0; i<size*2; i++) printf("%d ", globaldata[i]); printf("\n"); /*After MPI_Gather is done, I want to get the newly assined array here. Now the globaldata array should hold values: 0 0 1 1 2 2 3 3 Therefore, in the next iteration of this while loop, these array values need to be send for a new calculation with Scatter & Gather }while(counter<2);*/ //Following need to be executed after all the scatter and gather has completed printf("Finally at Processor %d has data: ", rank); for (i=0; i<size*2; i++)//Here the result should be: 0 0 2 2 3 3 4 4 printf("%d ", globaldata[i]); printf("\n"); } MPI_Scatter(globaldata, 2, MPI_INT, &localdata, 2, MPI_INT, 0, MPI_COMM_WORLD); localdata[0]= localdata[0]+rank; localdata[1]= localdata[1]+rank; MPI_Gather(&localdata, 2, MPI_INT, globaldata, 2, MPI_INT, 0, MPI_COMM_WORLD); if (rank == 0) {//Currently I can only see the newly assinged array values if I print out the result at the bottom printf("At the bottom, Processor %d has data: ", rank); for (i=0; i<size*2; i++) printf("%d ", globaldata[i]); printf("\n"); } MPI_Finalize(); return 0; }
关于我上面尝试做什么的更多解释：我想将globaldata数组发送到所有处理器。 然后获取更新的globaldata数组。 一旦我得到更新的数组，我想再次将此数组重新发送回所有其他进程以进行更多计算。 我编写了以下代码，使用MPI_Send和MPI_Recv执行类似的工作。 这里我使用MPI_Send将我的arrays发送到所有处理器。 然后这个数组将改变它并发送回根/进程0.一旦我得到修改后的数组， do while循环将再次运行并执行更多计算。 我想要做的是：以类似的方式使用MPI_Scatter和MPI_Gather。 我在哪里获得更新的globaldata数组并将其发送回MPI_Scatter和MPI_Gather以再次更改该数组
#include #include #include int main(int argc, char **argv) { int size, rank; MPI_Init(&argc, &argv); MPI_Comm_size(MPI_COMM_WORLD, &size); MPI_Comm_rank(MPI_COMM_WORLD, &rank); int globaldata[8]; int counter, i; if (rank == 0) { for (i=0; i<size*2; i++) globaldata[i] = 0; counter=0; do { /*becase of this do while loop "globaldata" array will always be updated and resend for more caculations*/ counter++; printf("Sending at Processor %d has data: ", rank); for (i=0; i<size*2; i++) printf("%d ", globaldata[i]); printf("\n"); for(i = 0; i < 4; i++) { MPI_Send(&globaldata, 8, MPI_INT, i, 0, MPI_COMM_WORLD); } for(i = 1; i < 4; i++) { MPI_Recv(&globaldata, 8, MPI_INT, i, 99, MPI_COMM_WORLD, MPI_STATUS_IGNORE); } }while(counter<2); /*Following executes after all the above calculations has completed*/ printf("Finally at Processor %d has data: ", rank); for (i=0; i<size*2; i++) printf("%d ", globaldata[i]); printf("\n"); } counter=0; do { counter++; MPI_Recv(&globaldata, 8, MPI_INT, 0, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE); globaldata[rank]=globaldata[rank]+rank; globaldata[rank+1]=globaldata[rank+1]+rank; MPI_Send(&globaldata, 8, MPI_INT, 0, 99, MPI_COMM_WORLD); }while(counter<2); MPI_Finalize(); return 0; }
只需将散布和聚集与循环中的本地处理放在一起：
if (rank == 0) { for (i = 0; i < size*2; i++) globaldata[i] = 0; } for (counter = 0; counter < 2; counter++) { // if (rank == 0) // { // pre-process globaldata // } MPI_Scatter(globaldata, 2, MPI_INT, localdata, 2, MPI_INT, 0, MPI_COMM_WORLD); localdata[0] += rank; localdata[1] += rank; MPI_Gather(localdata, 2, MPI_INT, globaldata, 2, MPI_INT, 0, MPI_COMM_WORLD); // if (rank == 0) // { // post-process globaldata // } } if (rank == 0) { printf("Finally at Processor %d has data: ", rank); for (i=0; i
或者，如果您希望将“主”流程的逻辑分开：
if (rank == 0) { for (i = 0; i < size*2; i++) globaldata[i] = 0; for (counter = 0; counter < 2; counter++) { // pre-process globaldata MPI_Scatter(globaldata, 2, MPI_INT, localdata, 2, MPI_INT, 0, MPI_COMM_WORLD); // Not really useful as rank == 0 and it changes nothing localdata[0] += rank; localdata[1] += rank; MPI_Gather(localdata, 2, MPI_INT, globaldata, 2, MPI_INT, 0, MPI_COMM_WORLD); // post-process globaldata } printf("Finally at Processor %d has data: ", rank); for (i=0; i
确保代码的两个部分中的循环具有相同的迭代次数。 另请注意， MPI_Scatter将一大块MPI_Scatter发送到根级别，并且globaldata会从根级收集一大块数据，因此主进程也需要执行一些数据处理。