Fine Grain Performance Evaluation Based on a Geometric Model of Execution

摘要

Performance evaluation can help improve system performance, predict the degree of scalability, and detect faults to improve efficiency and reliability. General performance measurement techniques such as benchmarking are useful for reporting overall system performance, but lack of details necessary to pin-point system strengths and bottlenecks. A strategy for fine grain performance evaluation is proposed, where performance is estimated by using performance points provided by a geometric model of execution. Estimation of centric points from the distribution of the performance points is obtained by the proposed Clusterig Screening Method. As an example the average performance of the Sun SuperSpark station is estimated and experimental results are compared with those generated by an Intersection Point Method. Problems of a geometric model of performance evaluation are discussed. The experimental results of the strategy are presented and analyzed.