GLU Implementation Architectures for Heterogeneous Systems

While a heterogeneous system ooers substantial aggregate processing power to exploit parallelism in GLU programs, its processors and links exhibit uctuating loads and can fail arbitrarily. On such a system, eeective exploitation of coarse-grain parallelism requires the runtime strategy for load distribution to adapt to system dynamics. In this paper, we describe two eduction-based architectures for implementing GLU on heterogeneous systems that use diierent approaches to cope with dynamically uctuating loads, processor failures, as well as processor additions. The rst implementation architecture, which is based on multiple distributed generators, deals with uctuating loads as well as processor changes using an adaptive value-demand distribution function. The second implementation architecture, which is also based on multiple distributed generators, uses dynamic inter-generator term-demand migration to adapt to uctuating loads as well as processor changes.