MPI race condition


When a buffer, which is involved in an uncompleted non-blocking communication, is used in another communication, there will be a race to read and write the buffer. The outcome of this situation is undefined. That is why the MPI standard states that a buffer must not be reused unless its communication is completed. It might look acceptable that multiple communications only read a buffer, however, it is still illegal in the MPI.


In the example below, process 0 sends buff0 in a non-blocking way to another process, it immediately moves forward to fill the buffer with data coming from another process. MPI_Isend can complete before, during or after the MPI_recv, therefore there is a race of reading and writing on buff0. To avoid this, we have two solutions: first using different buffers for MPI_Isend and MPI_Recv, and second, waiting for MPI_Isend to complete with the aid of MPI_Wait.

MPI_Isend and MPI_Recv use same buffer

Always check non-blocking communication

Let’s modify the previous example in a way that after MPI_Isend a very time-consuming task happens. Let’s say we are sure that MPI_Isend has time to complete during the task, is it OK to reuse the buffer in another communication afterward? Nope, we Must check the completion of MPI_Isend using MPI_Wait or MPI_Test.

MPI_Isend runs during high computational task


Let’s create a code to examine the previous rule. Rank 0 sends buffer=2 to rank 1 which receives it using MPI_Irecv. Then the threads are put in sleep for 2 seconds to mimic a time-consuming task.

#include <mpi.h>
#include <chrono>
#include <thread>
#include <stdio.h>
using namespace std;
int main(){
   MPI_Init(NULL, NULL);
   int rank;
   MPI_Comm_rank(MPI_COMM_WORLD, &rank);
   MPI_Request req;
   MPI_Status stat;
   int sendbuf = 2;
   int recvbuf = 3;
   if (rank == 0){
       MPI_Isend( &sendbuf, 1, MPI_INT, 1, 0, MPI_COMM_WORLD, &req);
       cout<< "Rank"<< rank<< " sendbuf="<< sendbuf<< endl;
   if (rank==1){
       MPI_Irecv( &recvbuf, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &req);
       int result = 0;
      // MPI_Test( &req, &result , &stat);
       cout<<"Rank"<< rank <<
           " recvbuf="<< recvbuf <<
           " success:"<< (result?"Yes":"No")

We expect within two seconds the data transfer happens, and the content of recvbuf to be 2, but we see the below outcome:

Rank0 sendbuf=2 
Rank1 recvbuf=3 success:No

Now we uncomment the MPI_Test line, and run the code, we get

Rank0 sendbuf=2
Rank1 recvbuf=2 success:Yes

Therefore, always call MPI_Test or MPI_Wait before reusing the buffer of non-blocking communications without any exception.


I got ideas and codes from the below website(s)

OpenMPI CodingGame Stackoverflow

Tags ➡ HPC MPI C++


I notify you of my new posts

Latest Posts


0 comment