Scalability and Performance Analysis of OpenMP Codes Using the Periscope Toolkit
keywords: Memory accesses analysis, OpenMP, performance analysis, program transformations, speedup, supercomputers
In this paper, we present two new approaches while rendering necessary extensions to Periscope to perform scalability and performance analysis on OpenMP codes. Periscope is an online-based performance analysis toolkit which consists of a user defined number of analysis agents that automatically search for the performance properties while the application is running. In order to detect the scalability and performance bottlenecks of OpenMP codes using Periscope, a few newly defined performance properties and meta properties are formalized. We manifest our implementation by evaluating NAS OpenMP benchmarks. As shown in our results, our approach identifies the code regions which do not scale well and other performance problems, e.g. load imbalance in NAS parallel benchmarks.
mathematics subject classification 2000: 68M14, 68M20
reference: Vol. 33, 2014, No. 4, pp. 921–942