Abstract: Performance analysis of modular or multi-layered applications can often feel like peering into a black box through a small hole. The Performance API (PAPI), which has served as the de facto standard for measuring hardware performance events for the past two decades, now supports a mechanism by which the software layers can expose internal information to the outside world, as software-defined events (SDEs). Such events can enable performance analysts to form a more complete picture of an application’s performance. In this talk, we present the basic ideas behind SDEs; we explore using SDEs to export useful information from inside diverse software layers, such as linear algebra libraries, task runtimes, and chemistry applications; we discuss the impact of SDEs on performance and software interdependence; and we touch on technical aspects regarding the inclusion of SDEs in a third-party application.
Bio: Anthony Danalis is a Research Assistant Professor with the Innovative Computing Laboratory, at the University of Tennessee. His interests span several topics within the broad domain of performance, ranging from performance measurement and evaluation, to compiler optimization, to novel programming paradigms. In previous work he has been involved in projects such as the CPU benchmark suite BlackjackBench, the GPU Benchmark suite SHOC, and the task scheduling runtime PaRSEC. His current research focuses on various extensions of the Performance Application Programming Interface (PAPI), which aim to (a) improve understanding of hardware counters, and (b) extend the notion of performance events to include not only hardware but also software-based information—all through one consistent interface.
Last Updated: August 25, 2020 - 9:36 am