MPI：生成树分段故障问题
问题描述 投票：0回答：1
我目前正在解决一个 MPI 程序中的问题，我正在开发该程序来构建跨多个进程的生成树。尽管进行了几天的故障排除，我还是遇到了障碍 - 我的程序始终因分段错误错误而崩溃。作为 MPI 编程相对较新的人，我正在寻求社区的帮助来帮助我诊断问题并找到解决方案。
[DESKTOP:18804] *** Process received signal ***
[DESKTOP:18804] Signal: Segmentation fault (11)
[DESKTOP:18804] Signal code: Address not mapped (1)
[DESKTOP:18804] Failing at address: 0xffffffffffffffff
*** stack smashing detected ***: terminated
[DESKTOP:18809] *** Process received signal ***
[DESKTOP:18809] Signal: Aborted (6)
[DESKTOP:18809] Signal code: (-6)
[DESKTOP:18809] *** Process received signal ***
[DESKTOP:18809] Signal: Segmentation fault (11)
[DESKTOP:18809] Signal code: Address not mapped (1)
[DESKTOP:18809] Failing at address: 0xffffffffffffffff
--------------------------------------------------------------------------
Primary job terminated normally, but 1 process returned
a non-zero exit code. Per user-direction, the job has been aborted.
--------------------------------------------------------------------------
--------------------------------------------------------------------------
mpirun noticed that process rank 7 with PID 0 on node DESKTOP-HUHQT64 exited on signal 11 (Segmentation fault).
--------------------------------------------------------------------------
[DESKTOP:18784] *** Process received signal ***
[DESKTOP:18784] Signal: Segmentation fault (11)
[DESKTOP:18784] Signal code: Address not mapped (1)
[DESKTOP:18784] Failing at address: (nil)
[DESKTOP:18784] [ 0] /lib/x86_64-linux-gnu/libc.so.6(+0x42520)[0x7f5008733520]
[DESKTOP:18784] [ 1] /lib/x86_64-linux-gnu/libpmix.so.2(PMIx_server_finalize+0x928)[0x7f5005eee5b8]
[DESKTOP:18784] [ 2] /usr/lib/x86_64-linux-gnu/openmpi/lib/openmpi3/mca_pmix_ext3x.so(ext3x_server_finalize+0x386)[0x7f5006052ff6]
[DESKTOP:18784] [ 3] /lib/x86_64-linux-gnu/libopen-rte.so.40(pmix_server_finalize+0xa0)[0x7f5008a4d5e0]
[DESKTOP:18784] [ 4] /usr/lib/x86_64-linux-gnu/openmpi/lib/openmpi3/mca_ess_hnp.so(+0x66b4)[0x7f50084f26b4]
[DESKTOP:18784] [ 5] /lib/x86_64-linux-gnu/libopen-rte.so.40(orte_finalize+0x65)[0x7f5008a23385]
[DESKTOP:18784] [ 6] mpirun(+0x1350)[0x55d954018350]
[DESKTOP:18784] [ 7] /lib/x86_64-linux-gnu/libc.so.6(+0x29d90)[0x7f500871ad90]
[DESKTOP:18784] [ 8] /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0x80)[0x7f500871ae40]
[DESKTOP:18784] [ 9] mpirun(+0x1415)[0x55d954018415]
[DESKTOP:18784] *** End of error message ***
该程序使用 MPI 进程实现分布式生成树算法。每个进程代表树中的一个节点。该算法在遍历树结构时使用父信息更新每个节点的表。当节点从其子节点接收到更新的表时，它会将这些信息与自己的信息合并，并将更新的表发送回其父节点。这个过程一直持续到到达根节点，从而形成一个完整的生成树，每个节点的表都填充了父节点信息。
#include <mpi.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#define MAX_TIME_TO_WAIT 30
#define TABLE_SIZE 100
// tag 0 : parent -> children (children updates table with the parent)
// tag 1: children -> parent (this is the new table)
// tag 2 : children -> node (already have a parent)
// Run: mpirun -np 12 ./a.out
// DO NOT USE DIRECTLY. USE getNeighbors() INSTEAD
int graph[][2] = {{0, 1}, {1, 2}, {2, 3}, {3, 4}, {4, 11}, {11, 5}, {5, 6}, {6, 7}, {7, 8}, {8, 9}, {9, 10}, {10, 9}, {9, 8}, {8, 7}, {7, 6}, {6, 5}, {5, 11}, {11, 4}, {4, 3}, {3, 2}, {2, 1}, {1, 0}, {9, 5}, {5, 9}, {5, 3}, {3, 5}, {0, 2}, {2, 0}, {9, 7}, {7, 9}};
// getNeighbors returns a 100 int vector.
// position 0 represents the number of neighbors + 1
// neigh[i] for i > 0 represent the ranks of the neighbors
// your program should only send messages to processes on this list
int *getNeighbors(int myRank)
{
int *neigh = malloc(sizeof(int) * 100);
int i;
int j = 1;
int sizeOfGraph = sizeof(graph) / sizeof(int) / 2;
for (i = 0; i < sizeOfGraph; i++) {
if (graph[i][0] == myRank) {
neigh[j] = graph[i][1];
j++;
}
}
neigh[0] = j;
return neigh;
}
void printNeighbors(int *neigh, int rank, int neighSize)
{
char aux[1000];
aux[0] = 0;
sprintf(aux + strlen(aux), "Rank %2i neighbors: ", rank);
for (int i = 1; i < neighSize; i++) {
sprintf(aux + strlen(aux), " %2i,", neigh[i]);
}
sprintf(aux + strlen(aux) - 1, "\n"); // also deletes the last comma
printf("%s", aux);
}
int getInitRank(int argc, char **argv)
{
if (argc < 2) {
printf("Usage: ./a.out <Init Node> (usually 3)\n");
exit(1);
}
int initRank = atoi(argv[1]);
return initRank;
}
void updateTable(int **table, int *childTable, int count)
{
for (int i = 0; i < count; i++) {
if (((*table)[i] == -1) && (childTable[i] != -1)) {
(*table)[i] = childTable[i];
}
}
}
void printTable(int **table, int rank, int count)
{
printf("\nRANK: %d\n", rank);
for (int i = 0; i < count; i++) {
printf("Node: %2d\tParent:%d\n", i, (*table)[i]);
}
}
void sendToChildren(int **table, int *neigh, int rank, int neighSize)
{
// sends to all children
MPI_Request req[neighSize];
for (int i = 1; i < neighSize; i++) {
if (neigh[i] != (*table)[rank]) {
MPI_Isend(*table, TABLE_SIZE, MPI_INT, neigh[i], 0, MPI_COMM_WORLD, &(req[i]));
printf("rank:%d sent to rank:%d\n", rank, neigh[i]);
}
}
}
void sendToParent(int **table, int rank)
{
// sends to parent
MPI_Send(*table, TABLE_SIZE, MPI_INT, (*table)[rank], 1, MPI_COMM_WORLD);
printf("rank:%d sent to parent rank:%d\n", rank, (*table)[rank]);
}
void recvParent(int **table, int *neigh, int rank, int neighSize)
{
MPI_Status status;
MPI_Request req[neighSize];
int flags[neighSize];
int timeWaited = 0;
int stillWait = 1;
int tmpTables[neighSize][TABLE_SIZE];
for (int i = 1; i < neighSize; i++) {
flags[i] = 0;
}
// waits for a parent
for (int i = 1; i < neighSize; i++) {
MPI_Irecv(tmpTables[i], TABLE_SIZE, MPI_INT, neigh[i], 0, MPI_COMM_WORLD, &(req[i]));
}
while ((timeWaited < MAX_TIME_TO_WAIT) && stillWait) {
for (int i = 1; i < neighSize; i++) {
MPI_Test(&(req[i]), &(flags[i]), &status);
if (flags[i]) {
printf("rank:%d recived from parent rank:%d\n", rank, neigh[i]);
memcpy(*table, tmpTables[i], sizeof(int) * TABLE_SIZE);
(*table)[rank] = neigh[i];
stillWait = 0;
break;
}
}
if (stillWait == 1) {
timeWaited++;
sleep(1);
} else {
break;
}
}
// tells other neighbours that already have a parent
for (int i = 1; i < neighSize; i++) {
if (neigh[i] != (*table)[rank]) {
MPI_Isend(*table, TABLE_SIZE, MPI_INT, neigh[i], 2, MPI_COMM_WORLD, &(req[i]));
printf("rank:%d informed rank:%d that already has a parent\n", rank, neigh[i]);
}
}
}
void recvChildren(int **table, int *neigh, int rank, int neighSize)
{
MPI_Status status;
MPI_Request req[neighSize];
int flags[neighSize];
int timeWaited = 0;
int stillWait = 1;
int tmpTable[neighSize][TABLE_SIZE];
for (int i = 1; i < neighSize; i++) {
for (int j = 0; j < TABLE_SIZE; j++) {
tmpTable[i][j] = -1;
}
}
for (int i = 1; i < neighSize; i++) {
flags[i] = 0;
}
// waits for response
for (int i = 1; i < neighSize; i++) {
if (neigh[i] != (*table)[rank]) {
MPI_Irecv(tmpTable[i], TABLE_SIZE, MPI_INT, neigh[i], MPI_ANY_TAG, MPI_COMM_WORLD, &(req[i]));
}
}
while ((timeWaited < MAX_TIME_TO_WAIT) && stillWait) {
stillWait = 0;
for (int i = 1; i < neighSize; i++) {
if (neigh[i] != (*table)[rank]) {
if (flags[i] == 0) {
MPI_Test(&(req[i]), &(flags[i]), &status);
if (flags[i]) {
if (status.MPI_TAG == 1) {
printf("rank:%d recived from child rank:%d\n", rank, neigh[i]);
updateTable(table, tmpTable[i], TABLE_SIZE);
} else if (status.MPI_TAG == 2) {
printf("rank:%d recived from not his child rank:%d\n", rank, neigh[i]);
}
} else {
stillWait = 1;
}
}
}
}
if (stillWait == 1) {
timeWaited++;
sleep(1);
} else {
break;
}
}
}
int main(int argc, char *argv[])
{
int rank;
int nProcesses;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nProcesses);
int initRank = getInitRank(argc, argv);
int *neigh = getNeighbors(rank);
int neighSize = neigh[0];
// printNeighbors(neigh, rank);
int *table = (int *)malloc(TABLE_SIZE * sizeof(int));
if (table == NULL) {
printf("Error allocating memory\n");
exit(1);
}
for (int i = 0; i < TABLE_SIZE; i++) {
table[i] = -1;
}
// initRank is the root
if (rank == initRank) {
sendToChildren(&table, neigh, rank, neighSize);
recvChildren(&table, neigh, rank, neighSize);
printTable(&table, rank, nProcesses);
} else {
recvParent(&table, neigh, rank, neighSize);
sendToChildren(&table, neigh, rank, neighSize);
recvChildren(&table, neigh, rank, neighSize);
printTable(&table, rank, nProcesses);
sendToParent(&table, rank);
}
free(table);
free(neigh);
MPI_Finalize();
return 0;
}
尽管我努力使用 gdb、valgrind 和 AddressSanitizer 进行调试，但我还是走进了死胡同。由于对这些工具和调试技术相对较新，我正在寻求一些指导和帮助。任何帮助将不胜感激！
编辑： 我修改了帖子中的代码和错误输出，以反映建议的更改以及当前生成的错误。
我已经按照@SvenNilsson的建议更改了代码，它的效果更好，但在某些情况下，我仍然会遇到分段错误，例如当 init_node 为 7 时。以下是我的程序中用于测试的图形的直观表示：
另外，按照@GillesGouaillardet 的要求，我添加了 mpirun 命令和 mpicc。 该程序编译为：
mpicc spanningTree.c -o spanningTree -Wall -O3 -g3
我运行它：
mpirun --oversubscribe -n 12 ./spanningTree <init_node>
除了错误消息之外，输出的关键部分是初始化节点打印的不完整表。
RANK: 7
Node: 0 Parent:2
Node: 1 Parent:2
Node: 2 Parent:3
Node: 3 Parent:5
Node: 4 Parent:11
Node: 5 Parent:6
Node: 6 Parent:7
Node: 7 Parent:-1
Node: 8 Parent:-1
Node: 9 Parent:-1
Node: 10 Parent:-1
Node: 11 Parent:5
1个回答
0
投票
投票
问题是节点在等待父节点时为每个邻居设置接收操作，但它只处理一个。这可能会导致离开子例程时出现未定义的行为。解决方案是仅使用一个 MPI_Irecv 并以 MPI_ANY_SOURCE 作为参数：
MPI_Irecv(auxTable, TABLE_SIZE, MPI_INT, MPI_ANY_SOURCE, 0, MPI_COMM_WORLD, &req)最新问题
- 如何在 GNU GCC 上启用 c++23？
- React.js 中使用 useParams 的未定义参数
- 列表中其值总和最多为 K 的元素的最大数量，复杂度为 O(log n)
- JS：如何删除动画而不只是停止它们？
- 两个独立的 Java 桌面应用程序之间的通信
- Java 8 JavaFX - 如何设置单个选项卡标题文本的颜色
- python函数根据主数据帧设置值
- 验证贝宝账户
- 运行分布式 API 测试的工具
- IFogSim：GUI 和代码之间的链接
- Pytest Monkeypatch 用于测试的多进程函数
- 使用 HR (95%CI) 绘制森林图
- 无法在Android Studio中添加插件
- 如何配置nest/Bull redis连接
- 错误：包“imageio”需要不同的 Python：2.7.17 不在“>=3.5”中
- 内容安全策略阻止 Firefox 中的本地主机
- Microsoft SQL Server 删除前导零和所有字符字母
- Php 已安装，但在 Apache2 版本 24.04 中未加载
- 如何从 NSArray 中选择一个对象？
- Typescript 如何从字符串中提取多个文字
© www.soinside.com 2019 - 2024. All rights reserved.