The performance of asynchronous algorithms on hypercubes

Abstract

Many asynchronous algorithms have been developed for parallel computers. Most implementations of asynchronous algorithms, however, have been for shared memory machines. In this paper, we study the implementation and performance of some common asynchronous algorithms on the NCUBE/ten, a 1024 node hypercube. In addition, we summarize existing theoretical work and discuss some classes of algorithms that can be made asynchronous and some that cannot.