Workshops Archive

PInsight: Enabling In-Situ Performance Analysis for Parallel Applications


Workshop: 8th Workshop on Programming and Performance Visualization Tools (ProTools)

Authors: Yonghong Yan (University of North Carolina, Charlotte) and Chunhua Liao (Lawrence Livermore National Laboratory)

Abstract: Performance analysis of parallel HPC applications is today a post-mortem workflow: measurement data is collected during a run, then moved and analyzed after the application terminates, and each tuning hypothesis costs another full execution. In this paper, we present PInsight (Performance Insight), a frame- work that leverages this iterative structure to enable in-situ performance analysis of applications spanning MPI, OpenMP, CUDA, and HIP in a single time-aligned trace. PInsight combines asynchronous tracing via LTTng with rate-controlled per-region sampling measurement, under a four-mode domain architecture with fine-grained filtering, tracing windows, node-level GPU activity policies, and runtime reconfiguration without application restart. Its analysis toolkit, both scripts and GUI, can consume the traces both post-hoc and in situ: window-end actions rotate the trace and launch analyses during execution with sub-second pause, turning a long-running application into a sequence of measure–analyze windows


Back to 8th Workshop on Programming and Performance Visualization Tools (ProTools) Archive Listing

Back to Full Workshop Archive Listing