Next: , Previous: , Up: MPI Data Distribution   [Contents][Index]


6.4.1 Basic and advanced distribution interfaces

As with the planner interface, the ‘fftw_mpi_local_size’ distribution interface is broken into basic and advanced (‘_many’) interfaces, where the latter allows you to specify the block size manually and also to request block sizes when computing multiple transforms simultaneously. These functions are documented more exhaustively by the FFTW MPI Reference, but we summarize the basic ideas here using a couple of two-dimensional examples.

For the 100 × 200 complex-DFT example, above, we would find the distribution by calling the following function in the basic interface:

ptrdiff_t fftw_mpi_local_size_2d(ptrdiff_t n0, ptrdiff_t n1, MPI_Comm comm,
                                 ptrdiff_t *local_n0, ptrdiff_t *local_0_start);

Given the total size of the data to be transformed (here, n0 = 100 and n1 = 200) and an MPI communicator (comm), this function provides three numbers.

First, it describes the shape of the local data: the current process should store a local_n0 by n1 slice of the overall dataset, in row-major order (n1 dimension contiguous), starting at index local_0_start. That is, if the total dataset is viewed as a n0 by n1 matrix, the current process should store the rows local_0_start to local_0_start+local_n0-1. Obviously, if you are running with only a single MPI process, that process will store the entire array: local_0_start will be zero and local_n0 will be n0. See Row-major Format.

Second, the return value is the total number of data elements (e.g., complex numbers for a complex DFT) that should be allocated for the input and output arrays on the current process (ideally with fftw_malloc or an ‘fftw_alloc’ function, to ensure optimal alignment). It might seem that this should always be equal to local_n0 * n1, but this is not the case. FFTW’s distributed FFT algorithms require data redistributions at intermediate stages of the transform, and in some circumstances this may require slightly larger local storage. This is discussed in more detail below, under Load balancing.

The advanced-interface ‘local_size’ function for multidimensional transforms returns the same three things (local_n0, local_0_start, and the total number of elements to allocate), but takes more inputs:

ptrdiff_t fftw_mpi_local_size_many(int rnk, const ptrdiff_t *n,
                                   ptrdiff_t howmany,
                                   ptrdiff_t block0,
                                   MPI_Comm comm,
                                   ptrdiff_t *local_n0,
                                   ptrdiff_t *local_0_start);

The two-dimensional case above corresponds to rnk = 2 and an array n of length 2 with n[0] = n0 and n[1] = n1. This routine is for any rnk > 1; one-dimensional transforms have their own interface because they work slightly differently, as discussed below.

First, the advanced interface allows you to perform multiple transforms at once, of interleaved data, as specified by the howmany parameter. (hoamany is 1 for a single transform.)

Second, here you can specify your desired block size in the n0 dimension, block0. To use FFTW’s default block size, pass FFTW_MPI_DEFAULT_BLOCK (0) for block0. Otherwise, on P processes, FFTW will return local_n0 equal to block0 on the first P / block0 processes (rounded down), return local_n0 equal to n0 - block0 * (P / block0) on the next process, and local_n0 equal to zero on any remaining processes. In general, we recommend using the default block size (which corresponds to n0 / P, rounded up).

For example, suppose you have P = 4 processes and n0 = 21. The default will be a block size of 6, which will give local_n0 = 6 on the first three processes and local_n0 = 3 on the last process. Instead, however, you could specify block0 = 5 if you wanted, which would give local_n0 = 5 on processes 0 to 2, local_n0 = 6 on process 3. (This choice, while it may look superficially more “balanced,” has the same critical path as FFTW’s default but requires more communications.)


Next: Load balancing, Previous: MPI Data Distribution, Up: MPI Data Distribution   [Contents][Index]