Finding the bottleneck: Performance in large-scale systems

Shubbi Taneja + Spencer Greene
Worcester Polytechnic Institute

Stevenson 1300
12:00 PM

Modern scientific software is increasingly complex - mixing multiple programming languages, running across CPUs and GPUs, and weaving in machine learning alongside traditional simulations. When something runs slowly, figuring out where the bottleneck is can be surprisingly difficult. ProfTool is an open-source tool that makes this easier: you simply tag the functions you care about, and ProfTool automatically tracks how long each one takes. It works with C, C++, Python, and GPU code, and produces a visual timeline you can explore in any web browser. In this talk, we will walk through how ProfTool works, the engineering improvements we made - including a compact recording mode that cuts overhead by 3x - and a real-world case study profiling GROMACS, a widely used molecular dynamics application, on a supercomputer.