A study of hardware performance counters selection for cross architectural GPU power modeling

By: Contributor(s): Material type: ArticleArticleDescription: 1 archivo (1,2 MB)Subject(s): Online resources: Summary: In the exascale race where huge corporations are spending billions of dollars on designing highly efficient heterogeneous supercomputers, the real need to reduce power envelopes forces current technologies to face crucial challenges as well as it demands the scientific community to evaluate and optimize the performance-power ratio. While energy consumption continues to climb up, the viability of these massive systems becomes a growing concern. In this context, the relevance of specific power-related research works turns into a priority. So we here develop an exhaustive step-by-step process for selecting a comprehensive set of hardware performance counters to serve as an input in an eventual GPU cross-architectural power consumption model. Our experiments show a high power-performance correlation between shared GPU events. Also, we present a set of events that delivers exclusive performance information in order to predict accurately GPU power fluctuations.
Star ratings
    Average rating: 0.0 (0 votes)
Holdings
Item type Home library Collection Call number URL Status Date due Barcode
Capítulo de libro Capítulo de libro Biblioteca de la Facultad de Informática Biblioteca digital A1302 (Browse shelf(Opens below)) Link to resource No corresponde

Formato de archivo PDF. -- Este documento es producción intelectual de la Facultad de Informática - UNLP (Colección BIPA/Biblioteca)

In the exascale race where huge corporations are spending billions of dollars on designing highly efficient heterogeneous supercomputers, the real need to reduce power envelopes forces current technologies to face crucial challenges as well as it demands the scientific community to evaluate and optimize the performance-power ratio. While energy consumption continues to climb up, the viability of these massive systems becomes a growing concern. In this context, the relevance of specific power-related research works turns into a priority. So we here develop an exhaustive step-by-step process for selecting a comprehensive set of hardware performance counters to serve as an input in an eventual GPU cross-architectural power consumption model. Our experiments show a high power-performance correlation between shared GPU events. Also, we present a set of events that delivers exclusive performance information in order to predict accurately GPU power fluctuations.

Congreso Argentino de Ciencias de la Computación (25to : 2019 : Río Cuarto, Córdoba)