我有一个MPI程序,用于从包含文件名列表的文件中读取多个进程,并基于读取的文件名 - 它读取相应的文件并计数单词的频率。当一个MPI进程执行MPI_Barrier()时,其他进程挂起
如果其中一个进程完成并返回 - 阻止执行MPI_Barrier(),则其他进程也会挂起。在调试时,可以看到函数没有输入process_files() readFile()函数无法找出原因。请看以下代码:
#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>
#include <ctype.h>
#include <string.h>
#include "hash.h"
void process_files(char*, int* , int, hashtable_t*);
void initialize_word(char *c,int size)
{
int i;
for(i=0;i<size;i++)
c[i]=0;
return;
}
char* readFilesList(MPI_File fh, char* file,int rank, int nprocs, char* block, const int overlap, int* length)
{
char *text;
int blockstart,blockend;
MPI_Offset size;
MPI_Offset blocksize;
MPI_Offset begin;
MPI_Offset end;
MPI_Status status;
MPI_File_open(MPI_COMM_WORLD,file,MPI_MODE_RDONLY,MPI_INFO_NULL,&fh);
MPI_File_get_size(fh,&size);
/*Block size calculation*/
blocksize = size/nprocs;
begin = rank*blocksize;
end = begin+blocksize-1;
end+=overlap;
if(rank==nprocs-1)
end = size;
blocksize = end-begin+1;
text = (char*)malloc((blocksize+1)*sizeof(char));
MPI_File_read_at_all(fh,begin,text,blocksize,MPI_CHAR, &status);
text[blocksize+1]=0;
blockstart = 0;
blockend = blocksize;
if(rank!=0)
{
while(text[blockstart]!='\n' && blockstart!=blockend) blockstart++;
blockstart++;
}
if(rank!=nprocs-1)
{
blockend-=overlap;
while(text[blockend]!='\n'&& blockend!=blocksize) blockend++;
}
blocksize = blockend-blockstart;
block = (char*)malloc((blocksize+1)*sizeof(char));
block = memcpy(block, text + blockstart, blocksize);
block[blocksize]=0;
*length = strlen(block);
MPI_File_close(&fh);
return block;
}
void calculate_term_frequencies(char* file, char* text, hashtable_t *hashtable,int rank)
{
printf("Start File %s, rank %d \n\n ",file,rank);
fflush(stdout);
if(strlen(text)!=0||strlen(file)!=0)
{
int i,j;
char w[100];
i=0,j=0;
while(text[i]!=0)
{
if((text[i]>=65&&text[i]<=90)||(text[i]>=97&&text[i]<=122))
{
w[j]=text[i];
j++; i++;
}
else
{
w[j] = 0;
if(j!=0)
{
//ht_set(hashtable, strcat(strcat(w,"#"),file),1);
}
j=0;
i++;
initialize_word(w,100);
}
}
}
return;
}
void readFile(char* filename, hashtable_t *hashtable,int rank)
{
MPI_Status stat;
MPI_Offset size;
MPI_File fx;
char* textFromFile=0;
printf("Start File %d, rank %d \n\n ",strlen(filename),rank);
fflush(stdout);
if(strlen(filename)!=0)
{
MPI_File_open(MPI_COMM_WORLD,filename,MPI_MODE_RDONLY,MPI_INFO_NULL,&fx);
MPI_File_get_size(fx,&size);
printf("Start File %s, rank %d \n\n ",filename,rank);
fflush(stdout);
textFromFile = (char*)malloc((size+1)*sizeof(char));
MPI_File_read_at_all(fx,0,textFromFile,size,MPI_CHAR, &stat);
textFromFile[size]=0;
calculate_term_frequencies(filename, textFromFile, hashtable,rank);
MPI_File_close(&fx);
}
printf("Done File %s, rank %d \n\n ",filename,rank);
fflush(stdout);
return;
}
void process_files(char* block, int* length, int rank,hashtable_t *hashtable)
{
char s[2];
s[0] = '\n';
s[1] = 0;
char *file;
if(*length!=0)
{
/* get the first file */
file = strtok(block, s);
/* walk through other tokens */
while(file != NULL)
{
readFile(file,hashtable,rank);
file = strtok(NULL, s);
}
}
return;
}
void execute_process(MPI_File fh, char* file, int rank, int nprocs, char* block, const int overlap, int * length, hashtable_t *hashtable)
{
block = readFilesList(fh,file,rank,nprocs,block,overlap,length);
process_files(block,length,rank,hashtable);
}
int main(int argc, char *argv[]){
/*Initialization*/
MPI_Init(&argc, &argv);
MPI_File fh=0;
int rank,nprocs,namelen;
char *block=0;
const int overlap = 70;
char* file = "filepaths.txt";
int *length = (int*)malloc(sizeof(int));
hashtable_t *hashtable = ht_create(65536);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
char processor_name[MPI_MAX_PROCESSOR_NAME];
MPI_Get_processor_name(processor_name, &namelen);
printf("Rank %d is on processor %s\n",rank,processor_name);
fflush(stdout);
execute_process(fh,file,rank,nprocs,block,overlap,length,hashtable);
printf("Rank %d returned after processing\n",rank);
MPI_Barrier(MPI_COMM_WORLD);
MPI_Finalize();
return 0;
}
的filepaths.txt是包含普通文本文件的绝对文件名的文件:
如:
/home/mpiuser/mpi/MPI_Codes/code/test1.txt
/home/mpiuser/mpi/MPI_Codes/code/test2.txt
/home/mpiuser/mpi/MPI_Codes/code/test3.txt
这个readFilesList看起来相当复杂,你确定它在那里生成正确的块大小吗?我认为你不会从并行化这部分代码中获得很多收益。读取单个文本文件(与实际数据相比,该文件相对较小,您希望从这些文件读取)在单个进程中更容易完成,甚至可能更快。因此,我会在单个进程中读取该列表并进行广播或分散结果列表的文件。 – haraldkl
在我看来,随后你继续让每个进程读取其中一个文件,而不是所有的读取部分。如果是这种情况,您不要在这里使用MPI_IO! MPI_read_all操作要求所有进程参与此文件的调用。 – haraldkl
readFilesList正在生成文件列表的非重叠块。但是,我会尝试单个进程的建议来读取它,并使用scatter将它们分配给进程。 – Dhanashree