SC17 Denver, CO

P72: New Developments for PAPI 5.6+

Authors: Anthony Danalis (University of Tennessee), Heike Jagode (University of Tennessee), Vince Weaver (University of Maine), Yan Liu (University of Maine), Jack Dongarra (University of Tennessee)

Abstract: The HPC community has relied on PAPI to track low-level hardware operations for over 15 years. In that time, the needs of software developers have changed immensely, and the PAPI team aims to meet these demands through a better understanding of deep and heterogeneous memory hierarchies and finer-grain power-management support.

This poster demonstrates how PAPI enables power-tuning to reduce overall energy consumption without, in many cases, a loss in performance. Furthermore, we discuss efforts to develop microbenchmarks intended to assist application developers who are interested in performance analysis by automatically categorizing and disambiguating performance counters. Finally, the poster illustrates efforts to update PAPI's internal sanity checks, designed to inspect that PAPI's predefined events are in fact measuring the values they claim to measure, and modernize the implementation of critical API functions, e.g., PAPI_read(), and the sampling interface so that more information can be captured and reported with lower overhead.

Award: Best Poster Finalist (BP): yes

Poster: pdf
Two-page extended abstract: pdf

Poster Index