The effectiveness of combining in shared memory parallel computers in the presence of “hot spots”
Gyungho Lee, Clyde P. Kruskal, David J. Kuck · IEEE Computer Society Press eBooks · 1994
Concurrent requests to a shared variable by many processors on a shared memory machine obviously create contention, which seems to become serious enough to stall the whole machine as the machine size grows. To alleviate the contention, restricted combining, in which two requests can be combined into a single request at a 2 x 2 switch, was suggested for the NYU Ultracomputer. We study the effectiveness of combining by considering several schemes under the nonuniform traffic model used by Pfister and Norton. Although restricted combining has been claimed as an effective way of allowing concurrent requests, in practice we witness a serious contention that can cripple the performance if the machine size is ''large''. Restricted combining seems to be too severe a restriction. The number of requests that can be combined at a switch turns out to be an important factor in the effectiveness of combining. Even a ''slight'' increase in the number of requests combined at a switch can resolve the problem. 15 refs., 5 figs.