PerfPal: Making Performance Analysis of High-Fidelity Computer Codes Simple and Accessible to DoD Users
Tiwari, Ananta (EP Analytics)
Co-Authors:
Carrington, Laura
Category:
Intersection of Digital Engineering and High Performance Computing/High End Computing
The Defense Threat Reduction Agency (DTRA) uses High Fidelity Computer Codes to investigate weapon effects phenomenology and techniques for countering Weapons of Mass Destruction. End-to-end, High-Fidelity (HF) simulations will require calculations including multiple phenomena occurring in vastly different time scales (microseconds to hours). As DTRA increasingly relies on computational modeling and simulation techniques, efficient use of existing computational resources and planning and optimizing for next-generation architectures become critical. Under the auspices of the Small Business Innovation Research (SBIR) program, DTRA has sponsored EP Analytics, a woman-owned small business headquartered in San Diego, CA, to conduct the commercialization of a performance analysis framework called PerfPal. The main objective of the framework is to remove the barriers to obtaining actionable performance insights for application scientists who may not possess performance engineering expertise. PerfPal accomplishes this objective by deploying tools to collect HF codes' inter-node and intra-node performance characteristics. The framework then analyzes such characterizations to derive intuitive performance reports and viewgraphs. Unlike other performance analysis frameworks, PerfPal eschews the need to recompile the application code to enable performance data collection. PerfPal utilizes binary instrumentation-based techniques to modify the production binaries to collect execution behavior.
The talk will present several case studies involving large DTRA HF codes to show PerfPal's capabilities in pinpointing critical performance bottlenecks. PerfPal-driven optimizations have delivered more than two-fold performance benefits, halving the turnaround time for critical simulation results.