MPI Gatherv 不工作（消息被截断）
MPI Gatherv not working (message truncated)
你好，我对 MPI_Gatherv 有疑问，它无法 'gather' 值，因为它 returns:
Fatal error in MPI_Gatherv: Message truncated, error stack:
MPI_Gatherv failed(sbuf=0x000001E0AAE36920, scount=16, MPI_INT,
rbuf=0x000001E0AAE367E0, rcnts=0x000001E0AAE18500,
displs=0x0000005A09F6F9D8, MPI_INT, root=0, MPI_COMM_WORLD) failed
Message truncated; 16 bytes received but buffer size is 16
代码使用 C 语言。
我的代码：
#include "stdio.h"
#include "mpi.h"
#include <stdlib.h>
int* multiply(int* x, int xLength, int* y, int yLength) {
int* resultMatrix = (int *) malloc(xLength*yLength * sizeof(int));
int r = 0;
for (int i = 0; i < xLength; i++) {
for (int j = 0; j < yLength; j++) {
resultMatrix[r] = x[i] * y[j];
printf("\nresult[%d]: %d", r, resultMatrix[r]);
r++;
}
}
return resultMatrix;
}
int* countOfValuesOfProcess(int matrixLength, int numOfProcesses) {
int* countOfValuesOfProcess = (int*) malloc (numOfProcesses);
for (int i = 0; i < numOfProcesses; i++) {
if (i == numOfProcesses - 1) {
countOfValuesOfProcess[i] = (matrixLength / numOfProcesses) + (matrixLength % numOfProcesses);
}else countOfValuesOfProcess[i] = matrixLength / numOfProcesses;
}
return countOfValuesOfProcess;
}
int main(argc, argv)
int argc; char *argv[];
{
int x[] = { 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16};
int y[] = { 2,3, -1, 4 };
int* result;
int size, rank;
int* recieveInt;
MPI_Status status;
MPI_Init(NULL, NULL);
MPI_Comm_size(MPI_COMM_WORLD, &size);
int xSize = sizeof(x) / sizeof(x[0]);
int ySize = sizeof(y) / sizeof(y[0]);
result = (int *) malloc((xSize * ySize) * sizeof(int));
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
int* numOfValuesPerProcess = countOfValuesOfProcess(xSize, size);
int displs[4];
recieveInt = (int *) malloc(numOfValuesPerProcess[rank] * sizeof(int));
int* resultPart = (int *) malloc((numOfValuesPerProcess[rank] * ySize) * sizeof(int));
//displs count
if (rank == 0) {
displs[0] = 0;
for (int i = 1; i < size; i++) {
displs[i] = (displs[i - 1] + numOfValuesPerProcess[i - 1]);
}
}
MPI_Scatterv(x, numOfValuesPerProcess, displs, MPI_INT, recieveInt, numOfValuesPerProcess[rank], MPI_INT, 0, MPI_COMM_WORLD);
resultPart = multiply(recieveInt, numOfValuesPerProcess[rank], y, ySize);
MPI_Gatherv(resultPart, numOfValuesPerProcess[rank]*ySize, MPI_INT, result, numOfValuesPerProcess, displs, MPI_INT, 0, MPI_COMM_WORLD);
free(resultPart);
free(recieveInt);
}
MPI_Finalize();
return(0);
}
当我将 gatherv 部分中的 numOfValuesPerProcess[rank]*ySize 替换为仅 numOfValuesPerProcess[rank] 时，它会起作用，但结果将是：
gathered matrix[0]: 2
gathered matrix[1]: 3
gathered matrix[2]: -1
gathered matrix[3]: 4
gathered matrix[4]: 10
gathered matrix[5]: 15
gathered matrix[6]: -5
gathered matrix[7]: 20
gathered matrix[8]: 18
gathered matrix[9]: 27
gathered matrix[10]: -9
gathered matrix[11]: 36
gathered matrix[12]: 26
gathered matrix[13]: 39
gathered matrix[14]: -13
gathered matrix[15]: 52
gathered matrix[16]: -842150451
gathered matrix[17]: -842150451
gathered matrix[18]: -842150451
gathered matrix[19]: -842150451
gathered matrix[20]: -842150451
gathered matrix[21]: -842150451
gathered matrix[22]: -842150451
gathered matrix[23]: -842150451
gathered matrix[24]: -842150451
gathered matrix[25]: -842150451
gathered matrix[26]: -842150451
gathered matrix[27]: -842150451
gathered matrix[28]: -842150451
gathered matrix[29]: -842150451
gathered matrix[30]: -842150451
gathered matrix[31]: -842150451
gathered matrix[32]: -842150451
gathered matrix[33]: -842150451
gathered matrix[34]: -842150451
gathered matrix[35]: -842150451
gathered matrix[36]: -842150451
gathered matrix[37]: -842150451
gathered matrix[38]: -842150451
gathered matrix[39]: -842150451
gathered matrix[40]: -842150451
gathered matrix[41]: -842150451
gathered matrix[42]: -842150451
gathered matrix[43]: -842150451
gathered matrix[44]: -842150451
gathered matrix[45]: -842150451
gathered matrix[46]: -842150451
gathered matrix[47]: -842150451
gathered matrix[48]: -842150451
gathered matrix[49]: -842150451
gathered matrix[50]: -842150451
gathered matrix[51]: -842150451
gathered matrix[52]: -842150451
gathered matrix[53]: -842150451
gathered matrix[54]: -842150451
gathered matrix[55]: -842150451
gathered matrix[56]: -842150451
gathered matrix[57]: -842150451
gathered matrix[58]: -842150451
gathered matrix[59]: -842150451
gathered matrix[60]: -842150451
gathered matrix[61]: -842150451
gathered matrix[62]: -842150451
gathered matrix[63]: -842150451
正如我们所看到的，收集了前 16 个数字，但缺少其余数字（因为我们只想要完整结果的一部分）我不知道问题出在哪里我试图为变量 int* 结果设置更大的内存分配，但没有奏效。
哪里有问题？
谢谢大家的建议
在您的 MPI_Gatherv 调用中，您从每个等级发送 numOfValuesPerProcess[rank]*ySize 元素，但仅在接收端为 numOfValuesPerProcess[rank] 元素保留 space。在你做乘法之后，你是 sending/receiving ySize 倍的数据，所以 recvcounts 和 displs[= MPI_Gatherv 调用的 25=] 参数需要考虑 ySize 因素。
顺便说一句，您似乎也有很多内存泄漏，frees 不足以满足 mallocs 的数量。学习使用 valgrind 等工具来帮助查找和修复这些问题。
更新代码：
#include "stdio.h"
#include "mpi.h"
#include <stdlib.h>
int* multiply(int* x, int xLength, int* y, int yLength) {
int* resultMatrix = malloc(xLength*yLength * sizeof(int));
int r = 0;
for (int i = 0; i < xLength; i++) {
for (int j = 0; j < yLength; j++) {
resultMatrix[r] = x[i] * y[j];
//printf("\nresult[%d]: %d", r, resultMatrix[r]);
r++;
}
}
return resultMatrix;
}
int* countOfValuesOfProcess(int matrixLength, int numOfProcesses) {
int* countOfValuesOfProcess = malloc(numOfProcesses * sizeof(int));
for (int i = 0; i < numOfProcesses; i++)
{
if (i == numOfProcesses - 1) {
countOfValuesOfProcess[i] = (matrixLength / numOfProcesses) + (matrixLength % numOfProcesses);
}
else
{
countOfValuesOfProcess[i] = matrixLength / numOfProcesses;
}
}
return countOfValuesOfProcess;
}
int main(int argc, char *argv[])
{
int x[] = { 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16};
int y[] = { 2, 3, -1, 4 };
int* result;
int size, rank;
int* recieveInt;
MPI_Status status;
MPI_Init(NULL, NULL);
MPI_Comm_size(MPI_COMM_WORLD, &size);
int xSize = sizeof(x) / sizeof(x[0]);
int ySize = sizeof(y) / sizeof(y[0]);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
int* numOfValuesPerProcess = countOfValuesOfProcess(xSize, size);
int displs[size];
//displs count
if (rank == 0) {
displs[0] = 0;
for (int i = 1; i < size; i++) {
displs[i] = (displs[i - 1] + numOfValuesPerProcess[i - 1]);
}
}
recieveInt = malloc(numOfValuesPerProcess[rank] * sizeof(int));
MPI_Scatterv(x, numOfValuesPerProcess, displs, MPI_INT, recieveInt, numOfValuesPerProcess[rank], MPI_INT, 0, MPI_COMM_WORLD);
int* resultPart = multiply(recieveInt, numOfValuesPerProcess[rank], y, ySize);
for (int i = 0; i < size; i++)
{
numOfValuesPerProcess[i] *= ySize;
displs[i] *= ySize;
}
result = (int *) malloc((xSize * ySize) * sizeof(int));
MPI_Gatherv(resultPart, numOfValuesPerProcess[rank], MPI_INT, result, numOfValuesPerProcess, displs, MPI_INT, 0, MPI_COMM_WORLD);
if (rank == 0)
{
for (int i = 0; i < xSize*ySize; i++)
printf("result[%d]: %d\n", i, result[i]);
}
free(resultPart);
free(recieveInt);
free(numOfValuesPerProcess);
free(result);
MPI_Finalize();
return(0);
}
你好，我对 MPI_Gatherv 有疑问，它无法 'gather' 值，因为它 returns:
Fatal error in MPI_Gatherv: Message truncated, error stack: MPI_Gatherv failed(sbuf=0x000001E0AAE36920, scount=16, MPI_INT, rbuf=0x000001E0AAE367E0, rcnts=0x000001E0AAE18500, displs=0x0000005A09F6F9D8, MPI_INT, root=0, MPI_COMM_WORLD) failed Message truncated; 16 bytes received but buffer size is 16
代码使用 C 语言。 我的代码：
#include "stdio.h"
#include "mpi.h"
#include <stdlib.h>
int* multiply(int* x, int xLength, int* y, int yLength) {
int* resultMatrix = (int *) malloc(xLength*yLength * sizeof(int));
int r = 0;
for (int i = 0; i < xLength; i++) {
for (int j = 0; j < yLength; j++) {
resultMatrix[r] = x[i] * y[j];
printf("\nresult[%d]: %d", r, resultMatrix[r]);
r++;
}
}
return resultMatrix;
}
int* countOfValuesOfProcess(int matrixLength, int numOfProcesses) {
int* countOfValuesOfProcess = (int*) malloc (numOfProcesses);
for (int i = 0; i < numOfProcesses; i++) {
if (i == numOfProcesses - 1) {
countOfValuesOfProcess[i] = (matrixLength / numOfProcesses) + (matrixLength % numOfProcesses);
}else countOfValuesOfProcess[i] = matrixLength / numOfProcesses;
}
return countOfValuesOfProcess;
}
int main(argc, argv)
int argc; char *argv[];
{
int x[] = { 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16};
int y[] = { 2,3, -1, 4 };
int* result;
int size, rank;
int* recieveInt;
MPI_Status status;
MPI_Init(NULL, NULL);
MPI_Comm_size(MPI_COMM_WORLD, &size);
int xSize = sizeof(x) / sizeof(x[0]);
int ySize = sizeof(y) / sizeof(y[0]);
result = (int *) malloc((xSize * ySize) * sizeof(int));
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
int* numOfValuesPerProcess = countOfValuesOfProcess(xSize, size);
int displs[4];
recieveInt = (int *) malloc(numOfValuesPerProcess[rank] * sizeof(int));
int* resultPart = (int *) malloc((numOfValuesPerProcess[rank] * ySize) * sizeof(int));
//displs count
if (rank == 0) {
displs[0] = 0;
for (int i = 1; i < size; i++) {
displs[i] = (displs[i - 1] + numOfValuesPerProcess[i - 1]);
}
}
MPI_Scatterv(x, numOfValuesPerProcess, displs, MPI_INT, recieveInt, numOfValuesPerProcess[rank], MPI_INT, 0, MPI_COMM_WORLD);
resultPart = multiply(recieveInt, numOfValuesPerProcess[rank], y, ySize);
MPI_Gatherv(resultPart, numOfValuesPerProcess[rank]*ySize, MPI_INT, result, numOfValuesPerProcess, displs, MPI_INT, 0, MPI_COMM_WORLD);
free(resultPart);
free(recieveInt);
}
MPI_Finalize();
return(0);
}
当我将 gatherv 部分中的 numOfValuesPerProcess[rank]*ySize 替换为仅 numOfValuesPerProcess[rank] 时，它会起作用，但结果将是：
gathered matrix[0]: 2
gathered matrix[1]: 3
gathered matrix[2]: -1
gathered matrix[3]: 4
gathered matrix[4]: 10
gathered matrix[5]: 15
gathered matrix[6]: -5
gathered matrix[7]: 20
gathered matrix[8]: 18
gathered matrix[9]: 27
gathered matrix[10]: -9
gathered matrix[11]: 36
gathered matrix[12]: 26
gathered matrix[13]: 39
gathered matrix[14]: -13
gathered matrix[15]: 52
gathered matrix[16]: -842150451
gathered matrix[17]: -842150451
gathered matrix[18]: -842150451
gathered matrix[19]: -842150451
gathered matrix[20]: -842150451
gathered matrix[21]: -842150451
gathered matrix[22]: -842150451
gathered matrix[23]: -842150451
gathered matrix[24]: -842150451
gathered matrix[25]: -842150451
gathered matrix[26]: -842150451
gathered matrix[27]: -842150451
gathered matrix[28]: -842150451
gathered matrix[29]: -842150451
gathered matrix[30]: -842150451
gathered matrix[31]: -842150451
gathered matrix[32]: -842150451
gathered matrix[33]: -842150451
gathered matrix[34]: -842150451
gathered matrix[35]: -842150451
gathered matrix[36]: -842150451
gathered matrix[37]: -842150451
gathered matrix[38]: -842150451
gathered matrix[39]: -842150451
gathered matrix[40]: -842150451
gathered matrix[41]: -842150451
gathered matrix[42]: -842150451
gathered matrix[43]: -842150451
gathered matrix[44]: -842150451
gathered matrix[45]: -842150451
gathered matrix[46]: -842150451
gathered matrix[47]: -842150451
gathered matrix[48]: -842150451
gathered matrix[49]: -842150451
gathered matrix[50]: -842150451
gathered matrix[51]: -842150451
gathered matrix[52]: -842150451
gathered matrix[53]: -842150451
gathered matrix[54]: -842150451
gathered matrix[55]: -842150451
gathered matrix[56]: -842150451
gathered matrix[57]: -842150451
gathered matrix[58]: -842150451
gathered matrix[59]: -842150451
gathered matrix[60]: -842150451
gathered matrix[61]: -842150451
gathered matrix[62]: -842150451
gathered matrix[63]: -842150451
正如我们所看到的，收集了前 16 个数字，但缺少其余数字（因为我们只想要完整结果的一部分）我不知道问题出在哪里我试图为变量 int* 结果设置更大的内存分配，但没有奏效。
哪里有问题？ 谢谢大家的建议
在您的 MPI_Gatherv 调用中，您从每个等级发送 numOfValuesPerProcess[rank]*ySize 元素，但仅在接收端为 numOfValuesPerProcess[rank] 元素保留 space。在你做乘法之后，你是 sending/receiving ySize 倍的数据，所以 recvcounts 和 displs[= MPI_Gatherv 调用的 25=] 参数需要考虑 ySize 因素。
顺便说一句，您似乎也有很多内存泄漏，frees 不足以满足 mallocs 的数量。学习使用 valgrind 等工具来帮助查找和修复这些问题。
更新代码：
#include "stdio.h"
#include "mpi.h"
#include <stdlib.h>
int* multiply(int* x, int xLength, int* y, int yLength) {
int* resultMatrix = malloc(xLength*yLength * sizeof(int));
int r = 0;
for (int i = 0; i < xLength; i++) {
for (int j = 0; j < yLength; j++) {
resultMatrix[r] = x[i] * y[j];
//printf("\nresult[%d]: %d", r, resultMatrix[r]);
r++;
}
}
return resultMatrix;
}
int* countOfValuesOfProcess(int matrixLength, int numOfProcesses) {
int* countOfValuesOfProcess = malloc(numOfProcesses * sizeof(int));
for (int i = 0; i < numOfProcesses; i++)
{
if (i == numOfProcesses - 1) {
countOfValuesOfProcess[i] = (matrixLength / numOfProcesses) + (matrixLength % numOfProcesses);
}
else
{
countOfValuesOfProcess[i] = matrixLength / numOfProcesses;
}
}
return countOfValuesOfProcess;
}
int main(int argc, char *argv[])
{
int x[] = { 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16};
int y[] = { 2, 3, -1, 4 };
int* result;
int size, rank;
int* recieveInt;
MPI_Status status;
MPI_Init(NULL, NULL);
MPI_Comm_size(MPI_COMM_WORLD, &size);
int xSize = sizeof(x) / sizeof(x[0]);
int ySize = sizeof(y) / sizeof(y[0]);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
int* numOfValuesPerProcess = countOfValuesOfProcess(xSize, size);
int displs[size];
//displs count
if (rank == 0) {
displs[0] = 0;
for (int i = 1; i < size; i++) {
displs[i] = (displs[i - 1] + numOfValuesPerProcess[i - 1]);
}
}
recieveInt = malloc(numOfValuesPerProcess[rank] * sizeof(int));
MPI_Scatterv(x, numOfValuesPerProcess, displs, MPI_INT, recieveInt, numOfValuesPerProcess[rank], MPI_INT, 0, MPI_COMM_WORLD);
int* resultPart = multiply(recieveInt, numOfValuesPerProcess[rank], y, ySize);
for (int i = 0; i < size; i++)
{
numOfValuesPerProcess[i] *= ySize;
displs[i] *= ySize;
}
result = (int *) malloc((xSize * ySize) * sizeof(int));
MPI_Gatherv(resultPart, numOfValuesPerProcess[rank], MPI_INT, result, numOfValuesPerProcess, displs, MPI_INT, 0, MPI_COMM_WORLD);
if (rank == 0)
{
for (int i = 0; i < xSize*ySize; i++)
printf("result[%d]: %d\n", i, result[i]);
}
free(resultPart);
free(recieveInt);
free(numOfValuesPerProcess);
free(result);
MPI_Finalize();
return(0);
}