Hyper-Systolic Routing for SIMD Systems

Abstract

Massively parallel SIMD systems like the Italian supercomputer APE100/Quadrics[1] are very attractive for high performance computing applications as they offer a pathway to Teracomputing in the near future[2] at an affordable price. The 3-dimensional interprocessor network of the supercomputer APE100/Quadrics has been designed to carry out fast data movements in quantum chromodynamics calculations and, as a consequence, supports a next-neighbor connectivity exclusively. Unfortunately, a lack of more general routing facilities presents a hindrance to the implementation of many scientific problems on this powerful machine. In this work we propose a novel emulation of routing for the distributed permutation problem on parallel SIMD computers with next-neighbor connectivity on an n-dimensional toroidal network. The router is thought to provide a substitute for a dynamically reconfigurable switch network, providing a systematic routing approach for systems without hardware routing support. On SIMD machines with next-neighbor connectivity like Quadrics it can be a workaround to alleviate or even overcome the lack of routing. Our method needs a total number of interprocessor communincation operations of O(p 3

Extracted Key Phrases

6 Figures and Tables

Cite this paper

@inproceedings{Hoferichter1997HyperSystolicRF, title={Hyper-Systolic Routing for SIMD Systems}, author={A. Hoferichter and Thomas Lippert and Klaus Schilling and Paolo Palazzari}, booktitle={PARCO}, year={1997} }