Introduction to MPI

Aveuh

326.6K views

GitHub

Open Source Your Knowledge, Become a Contributor

Technology knowledge has to be shared and made accessible for free. Join the movement.

Create Content

Previous: Probing - exercise Next: Point-to-point communications - conclusion

Communication modes

We have seen almost everything important about point-to-point communication. There is one thing I have not talked about yet and that might be interesting to understand regarding MPI.

If you have experimented a bit yet, you might have remarked something awkward when doing blocking communications. Sometimes, it seems that the blocking communication is not blocking, while at other times, it seems to be. How strange.

For those of you who have not stumbled across that strange behaviour, here is a small example to convince you :

Blocking awkwardness

#include <iostream>
#include <unistd.h>
#include <mpi.h>
int main(int argc, char **argv) {
  MPI_Init(&argc, &argv);
  int size, rank;
  MPI_Comm_size(MPI_COMM_WORLD, &size);
  MPI_Comm_rank(MPI_COMM_WORLD, &rank);
  // We have two buffers, one big and one small
  constexpr int small_count = 50;
  constexpr int big_count   = 100000;
  int buff1[small_count];
  int buff2[big_count];
  // Waiting time, play with this value if you think the difference is due to the quantity of data to send
  constexpr int waiting_time = 5; // in seconds
  if (rank == 0) {
    // We initialize both buffers
    for (int i=0; i < big_count; ++i) {
      if (i < small_count)
  buff1[i] = i;
      buff2[i] = i;
    }
    // Synchronisation 1
    MPI_Barrier(MPI_COMM_WORLD);
    // We register the current time
    double time = -MPI_Wtime();
    // We send the buffer immediately
    MPI_Send(buff1, small_count, MPI_INT, 1, 0, MPI_COMM_WORLD);
    // We print the time it took us to complete this send
    std::cout << "Time elapsed to complete blocking send 1 : " << time + MPI_Wtime() << "s" << std::endl;
    // Synchronisation 2
    MPI_Barrier(MPI_COMM_WORLD);
    time = -MPI_Wtime();
    // We send the second buffer
    MPI_Send(buff2, big_count, MPI_INT, 1, 1, MPI_COMM_WORLD);
    // We print the time again
    std::cout << "Time elapsed to complete blocking send 2 : " << time + MPI_Wtime() << "s" << std::endl;
  }
  else {

This example shows two blocking-communication. One is performed on a small buffer (50 elements) while the second is performed on a large buffer (100000 elements). Process 0, only sends the buffers, and prints when the send has been performed as well as the time the send as taken to be completed. Process 1 is going to wait (to demonstrate the blocking properties) then receive and warn the user when the receive has been completed.

You can easily notice the awkward behaviour : the first send is completed way before the first receive, while for the large buffer, both operation seem to be happening at the same time.

The reason for that is that the first send is performed in synchronous mode while the second is not. There are actually four different communication modes you can trigger while communicating : standard, buffered, synchronous and ready. Each of these mode has a corresponding Send function (The Recv function will always be the same).

The standard mode (MPI_Send) is actually a "non-mode" that lets the MPI implementation choose which communication mode is preferable. This might be heavily dependent on the implementation. In OpenMPI, the observed behaviour is that for short messages, the send is automatically buffered while for long messages, the message will be sent using a mode somewhat close to the synchronous mode.

The buffered mode (MPI_Bsend) stores all the data to be sent in a temporary buffer and returns to the execution, just as a non-blocking send would do. The advantage here is that execution continues immediatly even if the corresponding blocking Recv has not been called yet. On the other hand, buffered mode copies all the data of your buffer into another region of memory, duplicating the data. This might be dangerous memory wise if you are transferring large amounts of data.

The ready mode (MPI_Rsend) can start only if the corresponding receive has already been called. This allows your program to gain time from some additional overhead in the initialization of messages. If the corresponding Recv has not yet been called, the message that will be received //might// be ill-defined so you have to make sure that the receiving process has asked for a Recv before calling the ready mode.

Finally, the synchronous mode (MPI_Ssend) will wait for the corresponding Recv to complete. The data transfer will occur at that exact moment, ensuring that both processes are ready for transfer.

Try to play a bit with these mode and the previous example to see the difference in times to return to the program.

Open Source Your Knowledge: become a Contributor and help others learn. Create New Content

Open Source Your Knowledge, Become a Contributor

12/30 Communication modes

Communication modes

PYTHON: BEGINNER QUIZ (10 Questions)

Simple Python Test

PYTHON STRING METHODS: center(), rjust() and ljust().

Stored Procedures con T-SQL