MPI_Scatterv segfault - c

Question

I'm just starting out with MPI programming and decided to make a simple distributed qsort using OpenMPI. To distribute parts of the array I want to sort I'm trying to use MPI_Scatterv, however the following code segfaults on me:
#include <stdio.h>
#include <string.h>
#include <stdlib.h>
#include <time.h>
#include <mpi.h>
#define ARRAY_SIZE 26
#define BUFFER_SIZE 2048
int main(int argc, char** argv) {
int my_rank, nr_procs;
int* data_in, *data_out;
int* sizes;
int* offsets;
srand(time(0));
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &nr_procs);
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
// everybody generates the control tables
int nr_workers = nr_procs-1;
sizes = malloc(sizeof(int)*nr_workers);
offsets = malloc(sizeof(int)*nr_workers);
int nr_elems = ARRAY_SIZE/nr_workers;
// basic distribution
for (int i = 0; i < nr_workers; ++i) {
sizes[i] = nr_elems;
}
// distribute the remainder
int left = ARRAY_SIZE%nr_workers;
int curr_worker = 0;
while (left) {
++sizes[curr_worker];
curr_worker = (++curr_worker)%nr_workers;
--left;
}
// offsets
int curr_offset = 0;
for (int i = 0; i < nr_workers; ++i) {
offsets[i] = curr_offset;
curr_offset += sizes[i];
}
if (my_rank == 0) {
// root
data_in = malloc(sizeof(int)*ARRAY_SIZE);
data_out = malloc(sizeof(int)*ARRAY_SIZE);
for (int i = 0; i < ARRAY_SIZE; ++i) {
data_in[i] = rand();
}
for (int i = 0; i < nr_workers; ++i) {
printf("%d at %d\n", sizes[i], offsets[i]);
}
MPI_Scatterv (data_in, sizes, offsets, MPI_INT, data_out, ARRAY_SIZE, MPI_INT, 0, MPI_COMM_WORLD);
} else {
// worker
printf("%d has %d elements!\n",my_rank, sizes[my_rank-1]);
// alloc the input buffer
data_in = malloc(sizeof(int)*sizes[my_rank-1]);
MPI_Scatterv(NULL, NULL, NULL, MPI_INT, data_in, sizes[my_rank-1], MPI_INT, 0, MPI_COMM_WORLD);
printf("%d got:\n", my_rank);
for (int i = 0; i < sizes[my_rank-1]; ++i) {
printf("%d ", data_in[i]);
}
printf("\n");
}
MPI_Finalize();
return 0;
}
How would I go about using Scatterv? Am I doing something wrong with allocating my input buffer from inside the worker code?

Answer 1

I changed some part in your code to get something working.
MPI_Scatter() will send data to every processors, including himself. According to your program, processor 0 expects ARRAY_SIZE integers, but sizes[0] is much smaller.
There are other problems on other processus : MPI_Scatter will send sizes[my_rank] integers, but sizes[my_rank-1] will be expected...
Here is a code that scatters data_in from 0 to all processors, including 0. Therefore i added 1 to nr_workers :
#include <stdio.h>
#include <string.h>
#include <stdlib.h>
#include <time.h>
#include <mpi.h>
#define ARRAY_SIZE 26
#define BUFFER_SIZE 2048
int main(int argc, char** argv) {
int my_rank, nr_procs;
int* data_in, *data_out;
int* sizes;
int* offsets;
srand(time(0));
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &nr_procs);
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
// everybody generates the control tables
int nr_workers = nr_procs;
sizes = malloc(sizeof(int)*nr_workers);
offsets = malloc(sizeof(int)*nr_workers);
int nr_elems = ARRAY_SIZE/nr_workers;
// basic distribution
for (int i = 0; i < nr_workers; ++i) {
sizes[i] = nr_elems;
}
// distribute the remainder
int left = ARRAY_SIZE%nr_workers;
int curr_worker = 0;
while (left) {
++sizes[curr_worker];
curr_worker = (++curr_worker)%nr_workers;
--left;
}
// offsets
int curr_offset = 0;
for (int i = 0; i < nr_workers; ++i) {
offsets[i] = curr_offset;
curr_offset += sizes[i];
}
if (my_rank == 0) {
// root
data_in = malloc(sizeof(int)*ARRAY_SIZE);
for (int i = 0; i < ARRAY_SIZE; ++i) {
data_in[i] = rand();
printf("%d %d \n",i,data_in[i]);
}
for (int i = 0; i < nr_workers; ++i) {
printf("%d at %d\n", sizes[i], offsets[i]);
}
} else {
printf("%d has %d elements!\n",my_rank, sizes[my_rank]);
}
data_out = malloc(sizeof(int)*sizes[my_rank]);
MPI_Scatterv (data_in, sizes, offsets, MPI_INT, data_out, sizes[my_rank], MPI_INT, 0, MPI_COMM_WORLD);
printf("%d got:\n", my_rank);
for (int i = 0; i < sizes[my_rank]; ++i) {
printf("%d ", data_out[i]);
}
printf("\n");
free(data_out);
if(my_rank==0){
free(data_in);
}
MPI_Finalize();
return 0;
}
Regarding memory managment, data_in and data_out should be freed at the end of the code.
Is it what you wanted to do ? Good luck with qsort ! I think you are not the first one to sort integers using MPI. See parallel sort using mpi. Your way to generate random numbers on the 0 processus and then scatter them is the right way to go. I think you will be interrested by his TD_Trier() function for communication. Even if you change tri_fusion(T, 0, size - 1); for qsort(...)...
Bye,
Francis