Achieving superlinear speedup on a heterogeneous, distributed system

Abstract

The CASA Gigabit Network Testbed, part of NSF and ARPA's Gigabit Project, is investigating whether a metacomputer consisting of widely distributed, heterogeneous supercomputers connected by a high-speed network is viable for large scientific applications. A particular challenge is to determine if such a metacomputer can produce superlinear speedup despite latency and communication overheads. One of the applications in the CASA testbed is a model we developed that couples a global atmosphere model to a world ocean model. Simulations using such coupled general circulation models for climate studies demand considerable computer resources. When distributing such a model, we need to consider the methods for masking latency with computation, the communications bandwidth requirements for different decomposition strategies, the optimal computer architecture for each major phase of the computation, and the effects of latency and communication costs for different decomposition strategies. Here we focus an the last two issues, and demonstrate that choosing the appropriate computer architectures and masking communication with computation can produce superlinear speedup.<>