cannam@167: Next: Calling FFTW from Modern Fortran, Previous: Multi-threaded FFTW, Up: Top [Contents][Index]
cannam@167:In this chapter we document the parallel FFTW routines for parallel cannam@167: systems supporting the MPI message-passing interface. Unlike the cannam@167: shared-memory threads described in the previous chapter, MPI allows cannam@167: you to use distributed-memory parallelism, where each CPU has cannam@167: its own separate memory, and which can scale up to clusters of many cannam@167: thousands of processors. This capability comes at a price, however: cannam@167: each process only stores a portion of the data to be cannam@167: transformed, which means that the data structures and cannam@167: programming-interface are quite different from the serial or threads cannam@167: versions of FFTW. cannam@167: cannam@167:
cannam@167: cannam@167:Distributed-memory parallelism is especially useful when you are cannam@167: transforming arrays so large that they do not fit into the memory of a cannam@167: single processor. The storage per-process required by FFTW’s MPI cannam@167: routines is proportional to the total array size divided by the number cannam@167: of processes. Conversely, distributed-memory parallelism can easily cannam@167: pose an unacceptably high communications overhead for small problems; cannam@167: the threshold problem size for which parallelism becomes advantageous cannam@167: will depend on the precise problem you are interested in, your cannam@167: hardware, and your MPI implementation. cannam@167:
cannam@167:A note on terminology: in MPI, you divide the data among a set of
cannam@167: “processes” which each run in their own memory address space.
cannam@167: Generally, each process runs on a different physical processor, but
cannam@167: this is not required. A set of processes in MPI is described by an
cannam@167: opaque data structure called a “communicator,” the most common of
cannam@167: which is the predefined communicator MPI_COMM_WORLD which
cannam@167: refers to all processes. For more information on these and
cannam@167: other concepts common to all MPI programs, we refer the reader to the
cannam@167: documentation at the MPI home
cannam@167: page.
cannam@167:
cannam@167:
cannam@167:
We assume in this chapter that the reader is familiar with the usage cannam@167: of the serial (uniprocessor) FFTW, and focus only on the concepts new cannam@167: to the MPI interface. cannam@167:
cannam@167:| • FFTW MPI Installation: | cannam@167: | |
| • Linking and Initializing MPI FFTW: | cannam@167: | |
| • 2d MPI example: | cannam@167: | |
| • MPI Data Distribution: | cannam@167: | |
| • Multi-dimensional MPI DFTs of Real Data: | cannam@167: | |
| • Other Multi-dimensional Real-data MPI Transforms: | cannam@167: | |
| • FFTW MPI Transposes: | cannam@167: | |
| • FFTW MPI Wisdom: | cannam@167: | |
| • Avoiding MPI Deadlocks: | cannam@167: | |
| • FFTW MPI Performance Tips: | cannam@167: | |
| • Combining MPI and Threads: | cannam@167: | |
| • FFTW MPI Reference: | cannam@167: | |
| • FFTW MPI Fortran Interface: | cannam@167: |
cannam@167: Next: Calling FFTW from Modern Fortran, Previous: Multi-threaded FFTW, Up: Top [Contents][Index]
cannam@167: