Understanding Potential Performance Issues Using Resource-Based alongside Time Models

Authors: Nan ding (Lawrence Berkeley National Laboratory), Victor W. Lee (Intel Corporation), Wei Xue (Tsinghua University), Weimin Zheng (Tsinghua University)

Abstract: Numerous challenges and opportunities are introduced by the complexity and enormous code legacy of HPC applications, the diversity of HPC architectures, and the nonlinearity of interactions between applications and HPC systems. To address these issues, we propose the Resource-based Alongside Time (RAT) modeling method to help to understand the application run-time performance efficiently. First, we use hardware counter-assisted profiling to identify the key kernels and non-scalable kernels in the application. Second, we show how to apply the resource-based profiling into performance models to understand the potential performance issues and predict performance in the regimes of interest to developers and performance analysts. Third, we propose an easy-to-use performance modeling tool for scientists and performance analytics. Our evaluations demonstrate that by only performing a few small-scale profilings, RAT is able to keep the average model error rate around 15% with average performance overheads of 3% in multiple scenarios.

Best Poster Finalist (BP): no

Poster: pdf
Poster summary: PDF


Back to Poster Archive Listing