Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: FFTW 3.3.8: An improved replacement for MPI_Alltoall Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: Chris@82:
Chris@82:

Chris@82: Previous: , Up: FFTW MPI Transposes   [Contents][Index]

Chris@82:
Chris@82:
Chris@82: Chris@82:

6.7.3 An improved replacement for MPI_Alltoall

Chris@82: Chris@82:

We close this section by noting that FFTW’s MPI transpose routines can Chris@82: be thought of as a generalization for the MPI_Alltoall function Chris@82: (albeit only for floating-point types), and in some circumstances can Chris@82: function as an improved replacement. Chris@82: Chris@82:

Chris@82: Chris@82:

MPI_Alltoall is defined by the MPI standard as: Chris@82:

Chris@82:
Chris@82: 
int MPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype sendtype, 
Chris@82:                  void *recvbuf, int recvcnt, MPI_Datatype recvtype, 
Chris@82:                  MPI_Comm comm);
Chris@82:
Chris@82: Chris@82:

In particular, for double* arrays in and out, Chris@82: consider the call: Chris@82:

Chris@82:
Chris@82: 
MPI_Alltoall(in, howmany, MPI_DOUBLE, out, howmany MPI_DOUBLE, comm);
Chris@82:
Chris@82: Chris@82:

This is completely equivalent to: Chris@82:

Chris@82:
Chris@82: 
MPI_Comm_size(comm, &P);
Chris@82: plan = fftw_mpi_plan_many_transpose(P, P, howmany, 1, 1, in, out, comm, FFTW_ESTIMATE);
Chris@82: fftw_execute(plan);
Chris@82: fftw_destroy_plan(plan);
Chris@82:
Chris@82: Chris@82:

That is, computing a P × P Chris@82: transpose on P processes, Chris@82: with a block size of 1, is just a standard all-to-all communication. Chris@82:

Chris@82:

However, using the FFTW routine instead of MPI_Alltoall may Chris@82: have certain advantages. First of all, FFTW’s routine can operate Chris@82: in-place (in == out) whereas MPI_Alltoall can only Chris@82: operate out-of-place. Chris@82: Chris@82:

Chris@82: Chris@82:

Second, even for out-of-place plans, FFTW’s routine may be faster, Chris@82: especially if you need to perform the all-to-all communication many Chris@82: times and can afford to use FFTW_MEASURE or Chris@82: FFTW_PATIENT. It should certainly be no slower, not including Chris@82: the time to create the plan, since one of the possible algorithms that Chris@82: FFTW uses for an out-of-place transpose is simply to call Chris@82: MPI_Alltoall. However, FFTW also considers several other Chris@82: possible algorithms that, depending on your MPI implementation and Chris@82: your hardware, may be faster. Chris@82: Chris@82: Chris@82:

Chris@82: Chris@82: Chris@82: Chris@82: Chris@82: