Chris@19: Chris@19: Chris@19: Avoiding MPI Deadlocks - FFTW 3.3.4 Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19:
Chris@19: Chris@19:

Chris@19: Next: , Chris@19: Previous: FFTW MPI Wisdom, Chris@19: Up: Distributed-memory FFTW with MPI Chris@19:


Chris@19:
Chris@19: Chris@19:

6.9 Avoiding MPI Deadlocks

Chris@19: Chris@19:

Chris@19: An MPI program can deadlock if one process is waiting for a Chris@19: message from another process that never gets sent. To avoid deadlocks Chris@19: when using FFTW's MPI routines, it is important to know which Chris@19: functions are collective: that is, which functions must Chris@19: always be called in the same order from every Chris@19: process in a given communicator. (For example, MPI_Barrier is Chris@19: the canonical example of a collective function in the MPI standard.) Chris@19: Chris@19: Chris@19:

The functions in FFTW that are always collective are: every Chris@19: function beginning with ‘fftw_mpi_plan’, as well as Chris@19: fftw_mpi_broadcast_wisdom and fftw_mpi_gather_wisdom. Chris@19: Also, the following functions from the ordinary FFTW interface are Chris@19: collective when they are applied to a plan created by an Chris@19: ‘fftw_mpi_plan’ function: fftw_execute, Chris@19: fftw_destroy_plan, and fftw_flops. Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: