Contributors:
Abstract: In an effort to guide optimizations and detect performance regressions, developers of large HPC codes must regularly collect and analyze application performance profiles across different hardware platforms and in a variety of program configurations.
However, traditional performance profiling tools mostly focus on ad-hoc analysis of individual program runs.
Ubiquitous performance analysis is a new approach to automate and simplify the collection, management, and analysis of large numbers of application performance profiles.
In this regime, performance profiling of large HPC codes transitions from a sporadic process that often requires the help of experts into a routine activity in which the entire development team can participate.
We discuss the design and implementation of an open source ubiquitous performance analysis software stack with three major components: the Caliper instrumentation library with a new API to control performance profiling programmatically; Adiak, a library for automatic program metadata capture; and SPOT, a web-based visualization interface for comparing large sets of runs.
A case study shows how ubiquitous performance analysis has helped the developers of the Marbl simulation code for over a year with analyzing performance and understanding regressions.
