Granularity experiments for the adaptive multiprocessor system (AMPS)

AMPS (Adaptive MultiProcessor System) is a shared-memory MIMD architecture intended for digital signal processing applications. We present our investigations on task granularity for AMPS. We have investigated task decomposition (granularity reduction) for matrix multiplication and fast Fourier transform (FFT) and task grouping (granularity enlargement) for Gram-Schmidt orthogonalization (GSO). The results indicate that significant performance improvement can be obtained for matrix multiplication and GSO by selecting the appropriate task granularity.