Efficient Multidimensional Data Redistribution for Resizable Parallel Computations - Computer Science > Distributed, Parallel, and Cluster Computing

Efficient Multidimensional Data Redistribution for Resizable Parallel Computations - Computer Science > Distributed, Parallel, and Cluster Computing - Descarga este documento en PDF. Documentación en PDF para descargar gratis. Disponible también para leer online.

Abstract: Traditional parallel schedulers running on cluster supercomputers supportonly static scheduling, where the number of processors allocated to anapplication remains fixed throughout the execution of the job. This results inunder-utilization of idle system resources thereby decreasing overall systemthroughput. In our research, we have developed a prototype framework calledReSHAPE, which supports dynamic resizing of parallel MPI applications executingon distributed memory platforms. The resizing library in ReSHAPE includessupport for releasing and acquiring processors and efficiently redistributingapplication state to a new set of processors. In this paper, we derive analgorithm for redistributing two-dimensional block-cyclic arrays from $P$ to$Q$ processors, organized as 2-D processor grids. The algorithm ensures acontention-free communication schedule for data redistribution if $P r \leqQ r$ and $P c \leq Q c$. In other cases, the algorithm implements circular rowand column shifts on the communication schedule to minimize node contention.

Autor: Rajesh Sudarsan, Calvin J. Ribbens

Fuente: https://arxiv.org/