SC19 Proceedings

The International Conference for High Performance Computing, Networking, Storage, and Analysis

Pinpointing Performance Inefficiencies via Lightweight Variance Profiling

Authors: Pengfei Su (College of William & Mary), Shuyin Jiao (College of William & Mary), Milind Chabbi (Scalable Machines Research), Xu Liu (College of William & Mary)

Abstract: Execution variance among different invocation instances of the same procedure is often an indicator of performance losses. On the one hand, instrumentation-based tools can insert calipers around procedures and identify execution variance; however, they can introduce high overheads. On the other hand, sampling-based tools insert no instrumentation and have low overheads; however, they cannot synchronize samples with procedure entry and exit.

In this paper, we propose FVSampler, a lightweight, sampling-based variance profiler. FVSampler employs hardware performance monitoring units in conjunction with hardware debug registers to sample and monitor whole procedure instances (invocation till return) and collect hardware metrics in each sampled procedure instance. FVSampler, typically, incurs only 6% runtime overhead and negligible memory overhead making it suitable for HPC-scale production codes. We evaluate FVSampler with several parallel applications and demonstrate its effectiveness in pinpointing execution variance. Guided by FVSampler, we tune data structures and algorithms to obtain significant speedups.

Presentation: file

Back to Technical Papers Archive Listing