cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: FFTW 3.3.5: Avoiding MPI Deadlocks cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127:
cannam@127:

cannam@127: Next: , Previous: , Up: Distributed-memory FFTW with MPI   [Contents][Index]

cannam@127:
cannam@127:
cannam@127: cannam@127:

6.9 Avoiding MPI Deadlocks

cannam@127: cannam@127: cannam@127:

An MPI program can deadlock if one process is waiting for a cannam@127: message from another process that never gets sent. To avoid deadlocks cannam@127: when using FFTW’s MPI routines, it is important to know which cannam@127: functions are collective: that is, which functions must cannam@127: always be called in the same order from every cannam@127: process in a given communicator. (For example, MPI_Barrier is cannam@127: the canonical example of a collective function in the MPI standard.) cannam@127: cannam@127: cannam@127:

cannam@127: cannam@127:

The functions in FFTW that are always collective are: every cannam@127: function beginning with ‘fftw_mpi_plan’, as well as cannam@127: fftw_mpi_broadcast_wisdom and fftw_mpi_gather_wisdom. cannam@127: Also, the following functions from the ordinary FFTW interface are cannam@127: collective when they are applied to a plan created by an cannam@127: ‘fftw_mpi_plan’ function: fftw_execute, cannam@127: fftw_destroy_plan, and fftw_flops. cannam@127: cannam@127: cannam@127: cannam@127:

cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: