PERFORMANCE MODELS FOR THE PROCESSOR FARM PARADIGM

In this paper, we describe the design, implementation, and modeling of a runtime kernel to support the processor farm paradigm on multicompu ters. We present a general topology-independent framework for obtainin g performance models to predict the performance of the start-up, stead y-state, and wind-down phases of a processor farm. An algorithm is des cribed, which for any interconnection network determines a tree-struct ured subnetwork that optimizes farm performance. The analysis techniqu e is applied to the important case of k-ary tree topologies. The model s are compared with the measured performance on a variety of topologie s using both constant and varied task sizes.