Adaptive distributed computing through competition

Abstract

A framework for supporting adaptive execution of parallel applications on networks of workstations is presented. The framework is comprised of two levels of competition. At the first level, the tasks of each application are partitioned into grains. The grains race one another until all their tasks are finished. The turn-around time of an application can be shortened by sharing the tasks of the heavily loaded grains with the neighboring grains. At the second level, a prototype system called Comedians has been developed, which enables competition among applications for workstations by mechanisms of auction and bidding. The objectives of the Comedians system are to maximize the speedup of individual parallel applications and, at the same time, to allocate workstations efficiently and fairly to the applications. Unlike related work, the paper suggests an integrated solution to both the issues of adaptive parallelism and parallel application scheduling in a multi-user environment. The experimental results demonstrate the effectiveness of the support for adaptive parallelism and the dynamics of competition among parallel applications.