Towards New Metrics for Appraising Performance and Energy Efficiency of Parallel Scientific Programs

Multiple and conflicting non-functional goals of program execution, such as speedup or energy efficiency, lead to an increasing effort for assessing the program properties. The time-consuming evaluation of program execution is further increased by multiple degrees of freedom influencing the execution behavior, such as the number of threads or the operational frequency of the processors. Furthermore, the efficiency properties strongly depend on the programs or algorithms and may vary when changing the implementation. In this article, we propose new metrics, e.g. the relative power increase factor, which are able to capture the overall quality of program execution such that it is possible to appraise multiple goals depending on different influential factors in an easy-to-use way. The new metrics are evaluated with respect to their usefulness and expressiveness. The investigations are done for the PARSEC benchmark suite on Intel DVFS processors.