Architectures and message-passing algorithms for cluster computing: Designand performance

This payer considers the architecture of clusters and related message-passi ng (MP) software algorithms and their effect on performance (speedup and ef ficiency) of cluster computing (CC). We present new architectures for multi -segment Ethernet clusters and new MP algorithms which fit these architectu res. The multiple segments (e.g, commodity hubs) connect commodity processo r nodes so as to allow MP to be highly parallelized by avoiding network con tention and collisions in many applications where the all-gather and other collective operations are central. We analyze all-gather in some detail, an d present new network topologies and new MP algorithms to minimize latency. The new topologies are based on a design, called two-by-four nets (2 x 4 n ets), by Compbionics. An integrated MP software system, called Reduced Over head Cluster Communication (ROCC), which embodies the MP algorithms is also described. In brief, 2 x 4 nets are networks of "supernodes", called 2 x 4 's, each having 4 processors on 2 segments and segments usually being Ether net hubs, The supernodes are typically connected to form rings or tori of s upernodes. We present actual test results and supporting analyses to demons trate that 2 x 4 nets with the ROCC MP software are faster than many existi ng clusters and generally less costly. (C) 2000 Published by Elsevier Scien ce B.V. All rights reserved.