Video: Profiling at Scale

"Developers of modern HPC applications face a challenge when scaling out their hybrid (MPI/OpenMP) applications. As cluster sizes continue to grow, the amount of analysis data collected can easily become overwhelming when going from 10s to 1000s of ranks and it's tough to identify which are the key metrics to track. There is a need for a lightweight tool that aggregates the performance data in a simple and intuitive way, provides advice on next optimizations steps, and hones in on performance issues. We'll discuss a brand new tool that helps quickly gather and analyze statistics up to 100,000 ranks. We'll give examples of the type of pertinent information collected at high core counts, including memory and counter usage, MPI and OpenMP imbalance analysis, and total communication vs. computation time. We'll work through analyzing an application and effective ways to manage the data."

The post Video: Profiling at Scale appeared first on insideHPC.