Multiphase complete exchange: A theoretical analysis

Complete Exchange requires each of N processors to send a unique message to each of the remaining N-1 processors. For a circuit switched hypercube with N = 2(sub d) processors, the Direct and Standard algorithms for Complete Exchange are optimal for very large and very small message sizes, respectively. For intermediate sizes, a hybrid Multiphase algorithm is better. This carries out Direct exchanges on a set of subcubes whose dimensions are a partition of the integer d. The best such algorithm for a given message size m could hitherto only be found by enumerating all partitions of d. The Multiphase algorithm is analyzed assuming a high performance communication network. It is proved that only algorithms corresponding to equipartitions of d (partitions in which the maximum and minimum elements differ by at most 1) can possibly be optimal. The run times of these algorithms plotted against m form a hull of optimality. It is proved that, although there is an exponential number of partitions, (1) the number of faces on this hull is Theta(square root of d), (2) the hull can be found in theta(square root of d) time, and (3) once it has been found, the optimal algorithm for any given m can be found in Theta(log d) time. These results provide a very fast technique for minimizing communication overhead in many important applications, such as matrix transpose, Fast Fourier transform, and ADI.