Submitted by scrawford on
Title | PAPI's New Software-Defined Events for In-Depth Performance Analysis |
Publication Type | Presentation |
Year of Publication | 2018 |
Authors | Jagode, H., A. Danalis, and J. Dongarra |
Date Published | 2018-09 |
Event | CCDSC 2018: Workshop on Clusters, Clouds, and Data for Scientific Computing |
Location | Lyon, France |
Abstract | One of the most recent developments of the Performance API (PAPI) is the addition of Software-Defined Events (SDE). PAPI has successfully served the role of the abstraction and unification layer for hardware performance counters for over a decade. This talk presents our effort to extend this role to encompass performance critical information that does not originate in hardware, but rather in critical software layers, such as libraries and runtime systems. Our overall objective is to enable monitoring of both types of performance events, hardware- and software-related events, in a uniform way, through one consistent PAPI interface. Performance analysts will be able to form a complete picture of the entire application performance without learning new instrumentation primitives. In this talk, we outline PAPI's new SDE API and showcase the usefulness of SDE through its employment in software layers as diverse as the math library MAGMA, the dataflow runtime PaRSEC, and the state-of-the-art chemistry application NWChem. We outline the process of instrumenting these software packages and highlight the performance information that can be acquired with SDEs. |