I intend to achieve a simple task using MPI collective communication but being new to MPI, I have found the collective routines somewhat non-intuitive. I have 4 slaves, each of which must read a unique string and send the string to all the other slaves.
I looked into MPI_Bcast, MPI_Scatter, and MPI_Alltoall. I settled for MPI_Alltoall but the program ends with bad termination.
The program is:
int main(int argc,char *argv[])
{
int my_rank, num_workers;
MPI_Comm SLAVES_WORLD;
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &num_workers);
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
createSlavesCommunicator(&SLAVES_WORLD);
char send_msg[20], recv_buf[20];
sprintf(send_msg, "test string %d", my_rank);
MPI_Alltoall(send_buf, strlen(send_buf), MPI_CHAR, recv_buf, 20, MPI_CHAR, MPI_COMM_WORLD);
printf("slave %d recvd message %s\n", my_rank, recv_buf);
}
void createSlavesCommunicator(MPI_Comm *SLAVES_WORLD)
{
MPI_Group SLAVES_GROUP, MPI_COMM_GROUP;
int ranks_to_excl[1];
ranks_to_excl[0] = 0;
MPI_Comm_group(MPI_COMM_WORLD, &MPI_COMM_GROUP);
MPI_Group_excl(MPI_COMM_GROUP, 1, ranks_to_excl, &SLAVES_GROUP);
MPI_Comm_create(MPI_COMM_WORLD, SLAVES_GROUP, SLAVES_WORLD);
}
MPI_AlltoAll() sends messages from everyone to everyone. Input and output buffer should be much larger than 20 if each process sends 20 char. Starting from your code, here is how MPI_AlltoAll() works :
#include <stdio.h>
#include <time.h>
#include <stdlib.h>
#include <string.h>
#include "mpi.h"
int main(int argc,char *argv[])
{
int my_rank, num_workers;
MPI_Comm SLAVES_WORLD;
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &num_workers);
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
//createSlavesCommunicator(&SLAVES_WORLD);
char send_msg[20*num_workers], recv_buf[20*num_workers];
int i;
for(i=0;i<num_workers;i++){
sprintf(&send_msg[i*20], "test from %d to %d", my_rank,i);
}
MPI_Barrier(MPI_COMM_WORLD);
//MPI_Alltoall(send_msg, strlen(send_msg), MPI_CHAR, recv_buf, 20, MPI_CHAR, MPI_COMM_WORLD);
MPI_Alltoall(send_msg, 20, MPI_CHAR, recv_buf, 20, MPI_CHAR, MPI_COMM_WORLD);
MPI_Barrier(MPI_COMM_WORLD);
for(i=0;i<num_workers;i++){
printf("slave %d recvd message %s\n", my_rank, &recv_buf[20*i]);
}
MPI_Finalize();
return 0;
}
Looking at your question, it seems that MPI_AllGather() is the function that would do the trick...
"The block of data sent from the jth process is received by every process and placed in the jth block of the buffer recvbuf. "
http://www.mcs.anl.gov/research/projects/mpi/www/www3/MPI_Allgather.html
Bye,
Francis
Related
I have some code that for testing purposes, I removed all sends and only have non-blocking receives. You can imagine my surprise when using MPI_Test the flags were indicating some of the requests were actually being completed. I have my code setup in a cartesian grid, with a small replica below, although this doesn't reproduce the error:
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h> // for sleep
#include <mpi.h>
void test(int pos);
MPI_Comm comm_cart;
int main(int argc, char *argv[])
{
int i, j;
int rank, size;
MPI_Status status;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
/* code for mpi cartesian gird topology */
int dim[1];
dim[0] = 2;
int periods[1];
periods[0] = 0;
int reorder = 1;
int coords[1];
MPI_Cart_create(MPI_COMM_WORLD, 1, dim, periods, 1, &comm_cart);
MPI_Cart_coords(comm_cart, rank, 2, coords);
test(coords[0]);
MPI_Finalize();
return (0);
}
void test(int pos)
{
float placeholder[4];
int other = (pos+1) % 2;
MPI_Request reqs[8];
int flags[4];
for(int iter = 0; iter < 20; iter++){
// Test requests from previous time cycle
for(int i=0;i<4;i++){
if(iter == 0) break;
MPI_Test(&reqs[0], &flags[0] , MPI_STATUS_IGNORE);
printf("Flag: %d\n", flags[0]);
}
MPI_Irecv(&placeholder[0], 1, MPI_FLOAT, other, 0, comm_cart, &reqs[0]);
}
}
Any help would be appreciated.
The issue is with MPI_Test and MPI_PROC_NULLs. Quite often when using MPI_Cart_shift, you end up with MPI_PROC_NULLs as if you're on the edge of the grid, a neighbouring cell simply doesn't exist in some directions.
I can't find any documentation for this anywhere, so I had to discover it myself, but when you do an MPI_Irecv with an MPI_PROC_NULL source, it will instantly complete and when tested using MPI_Test, the flag will return true for a completed request. Example code below:
#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>
int main(int argc, char *argv[])
{
MPI_Init(&argc, &argv);
int t;
int flag;
MPI_Request req;
MPI_Irecv(&t, 1, MPI_INT, MPI_PROC_NULL, 0, MPI_COMM_WORLD, &req);
MPI_Test(&req, &flag, MPI_STATUS_IGNORE);
printf("Flag: %d\n", flag);
MPI_Finalize();
return (0);
}
Which returns the following when run:
Flag: 1
Flag: 1
When I run this code with one process, it works fine:
#include <iostream>
#include "mpi.h"
using namespace std;
int main(int argc, char* argv[])
{
MPI_Comm SubWorld;
MPI_Init(&argc, &argv);
int Rank;
MPI_Comm_rank(MPI_COMM_WORLD, &Rank);
if (Rank == 0)
MPI_Comm_split(MPI_COMM_WORLD, 1, Rank, &SubWorld);
cout << Rank << endl;
MPI_Finalize();
return 0;
}
Output: 0
But when I run it with more than one process, the output is "1" only!!.
This code works fine:
#include <iostream>
#include "mpi.h"
using namespace std;
int main(int argc, char* argv[])
{
MPI_Comm SubWorld;
MPI_Init(&argc, &argv);
int Rank;
MPI_Comm_rank(MPI_COMM_WORLD, &Rank);
MPI_Comm_split(MPI_COMM_WORLD, 1, Rank, &SubWorld);
cout << Rank << endl;
MPI_Finalize();
return 0;
}
Outputs two lines with "0" and "1" in some order.
So, the question is what the wrong in the first code when I run it with more than one process?
MPI_Comm_split() is a collective operation, and it must be invoked by all the MPI tasks of the communicator (MPI_COMM_WORLD here).
So what really happens in your first example running on multiple nodes is that MPI task 0 hangs in MPI_Comm_split()
#include <stdio.h>
#include <gmp.h>
#include <mpi.h>
int main(int *argc, char **argv)
{
int rank;
mpz_t a;
mpz_t b;
mpz_init(a);
mpz_init(b);
mpz_set_ui(a,23000000000000000000000000001);
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
if(rank==0)
{
MPI_Send(&a,1,MPI_INT,1,100,MPI_COMM_WORLD);
}
if(rank==1)
{
MPI_Recv(&b,1,MPI_INT,0,100,MPI_COMM_WORLD,MPI_STATUS_IGNORE);
}
MPI_Finalize();
return 1;
}
This code produce error so please help me how to send mpz_t variable through MPI_Send function.
I found a presentation www.macs.hw.ac.uk/~hwloidl/Courses/F21DP/gph_milan15_handout.pdf where, on slides 86 and 88, they send a multiprecision integer by marshalling it to/from a string using gmp_sprintf() and gmp_sscanf().
Actually, I notice they're using rationals of type mpq_t but I guess the same approach would work for mpz_t.
I'm trying to implement the following scenario using mpi_comm_spawn & scatter :
1- Master spawns 2 processes with a job.
2- He scatters an array to those spawned processes.
3- The spawned processes receive the scattered array sort it then send it back.
4- The master receives the sorted parts of the array.
I'd like to know how to do the step 2, so far i've tried with send and receives, they work perfectly but i want to do it with the scatter function.
Edit : Here's what i'd like to do in the master code , i'm missing the part in the slave's where i receive the scattered array
/*Master Here*/
MPI_Comm_spawn(slave, MPI_ARGV_NULL, 2, MPI_INFO_NULL,0, MPI_COMM_WORLD, &inter_comm, array_of_errcodes);
printf("MASTER Sending a message to slaves \n");
MPI_Send(message, 50, MPI_CHAR,0 , tag, inter_comm);
MPI_Scatter(array, 10, MPI_INT, &array_r, 10, MPI_INT, MPI_ROOT, inter_comm);
Thanks.
master.c
#include "mpi.h"
int main(int argc, char *argv[])
{
int n_spawns = 2;
MPI_Comm intercomm;
MPI_Init(&argc, &argv);
MPI_Comm_spawn("worker_program", MPI_ARGV_NULL, n_spawns, MPI_INFO_NULL, 0, MPI_COMM_SELF, &intercomm, MPI_ERRCODES_IGNORE);
int sendbuf[2] = {3, 5};
int recvbuf; // redundant for master.
MPI_Scatter(sendbuf, 1, MPI_INT, &recvbuf, 1, MPI_INT, MPI_ROOT, intercomm);
MPI_Finalize();
return 0;
}
worker.c
#include "mpi.h"
#include <stdio.h>
int main(int argc, char *argv[])
{
MPI_Init(&argc, &argv);
MPI_Comm intercomm;
MPI_Comm_get_parent(&intercomm);
int sendbuf[2]; // redundant for worker.
int recvbuf;
MPI_Scatter(sendbuf, 1, MPI_INT, &recvbuf, 1, MPI_INT, 0, intercomm);
printf("recvbuf = %d\n", recvbuf);
MPI_Finalize();
return 0;
}
Command line
mpicc master.c -o master_program
mpicc worker.c -o worker_program
mpirun -n 1 master_program
I wonder if it is possible to send data to a 3rd party communicator by having its integer value.
In other words, I would like to send a MPI_Comm (int) to processes in a non-related communicator in order to establish a communication between them.
For example, this picture shows my goal:
For this purpose, I have developed testing codes which intend to transfer the MPI_Comm.
parent.c
#include <mpi.h>
#include <stdio.h>
int main(int argc, char** argv) {
MPI_Init(NULL, NULL);
int world_size, world_rank;
MPI_Comm_size(MPI_COMM_WORLD, &world_size);
MPI_Comm_rank(MPI_COMM_WORLD, &world_rank);
MPI_Comm children;
int err[world_size], msg;
MPI_Comm_spawn("./children", NULL, 2, MPI_INFO_NULL, 0, MPI_COMM_WORLD, &children, err);
if (world_rank == 0) {
MPI_Send(&children, 1, MPI_INT, 0, 0, children);
MPI_Recv(&msg, 1, MPI_INT, MPI_ANY_SOURCE, MPI_ANY_TAG, children, MPI_STATUS_IGNORE);
}
MPI_Finalize();
return (0);
}
children.c
#include <mpi.h>
#include <stdio.h>
int main(int argc, char** argv) {
MPI_Init(NULL, NULL);
int world_size, world_rank;
MPI_Comm_size(MPI_COMM_WORLD, &world_size);
MPI_Comm_rank(MPI_COMM_WORLD, &world_rank);
MPI_Comm parent;
MPI_Comm_get_parent(&parent);
int comm, msg = 123;
if (world_rank == 0) {
MPI_Recv(&comm, 1, MPI_INT, MPI_ANY_SOURCE, MPI_ANY_TAG, parent, MPI_STATUS_IGNORE);
MPI_Comm children = (MPI_Comm) comm;
MPI_Send(&msg, 1, MPI_INT, 0, 0, children);
}
MPI_Finalize();
return (0);
}
Of course, the code does not work due to:
Fatal error in MPI_Send: Invalid communicator, error stack
Is there any way two establish that connection?
PS: This is an example, and I know that if I used "parent" comm in the send of "children.c", it would work. But my intention is to send data to a 3rd party communicator with only its integer id.