Performance analysis of parallel computations
Bin Qin · 1988
This study investigates the modeling and analysis of performance of parallel computations. Time cost is used in the study as the performance metrics. It is assumed parallel computations under investigation reside in a computer system in which there is a limited number of processors, all the processors have the same speed and they communicate with each other through a shared memory. It has been found in the study the performance of parallel computations depends on the input, the algorithm, the data structure, the processor speed, the number of processors, the processing power allocation, the communication, the execution overhead and the execution environment. In the study, the performance is defined as a function of the first seven factors listed above. The computation structure model is modified to describe the impact of the number of processors, the processing power allocation and the communication on performance. Activation signals are used in the computation structure model are allowed to carry processors and processing power allocation policy. Once a node in a parallel computation receives all the required signals, it uses the processors and allocation policy carried by the signals to perform the specified operations. The performance of the node then depends on the number of processors the node receives which in turn, depends on the allocation policy. Locking technique is used to model communication. Analytic techniques based on the modified model are developed to analyze performance. A software tool, TCAS (Time Cost Analysis System), is designed and implemented to aid users in determining the performance of their parallel computations.