David Skinner

"Integrated Performance Monitoring from Application to Workload"
LBNL


Abstract:

Profiling parallel codes on DOE Office of Science and NSF HPC machines provides value to scientists as well as managers of HPC resources. The focus of our work is on profiling in the wild of production workloads via the Integrated Performance Monitoring (IPM) framework which stresses ease-of-use and low overhead. Over the last two years the large volume of profiles collected has lead to challenges and opportunities in managing and analyzing large numbers of profiles. We present new research in workload profiling and moving from profiles to concise application models.