Characterisation based bottleneck analysis of parallel systems

Bottleneck analysis plays an important role in the early design of parallel computers and programs. In this paper a methodology for bottleneck analysis based on an instruction level characterisation technique is presented. The methodology is based on the assumption that a bottleneck is cauused by the slowest component of a computing system. These components are: memory (internal, external), processor (ALU, FPU), communication and I/O. Three metrics were used to identify bottlenecks in the system components. These are the B-ration, the communications-computation ratio and the memory-processing ratio. These ratios are dimensionless with values greater than unity indicates the presence of a bottleneck. The methodology is illustrated and validated using a communication intensive linear solver algorithm (Gauss-Jordan elimination) which was implemented on a mesh connected distributed memory parallel computer (128 T800 Parystec SuperCluster)