Prediction models for performance, power, and energy efficiency of software executed on heterogeneous hardware

Dénes Bán, R. Ferenc, István Siket, Ákos Kiss, T. Gyimóthy

Research output: Contribution to journalArticle

Abstract

Heterogeneous computer environments are becoming commonplace so it is increasingly important to understand how and where we could execute a given algorithm the most efficiently. In this paper we propose a methodology that uses both static source code metrics, and dynamic execution time, power, and energy measurements to build gain ratio prediction models. These models are trained on special benchmarks that have both sequential and parallel implementations and can be executed on various computing elements, e.g., on CPUs, GPUs, or FPGAs. After they are built, however, they can be applied to a new system using only the system’s static source code metrics which are much more easily computable than any dynamic measurement. We found that while estimating a continuous gain ratio is a much harder problem, we could predict the gain category (e.g., “slight improvement” or “large deterioration”) of porting to a specific configuration significantly more accurately than a random choice, using static information alone. We also conclude based on our benchmarks that parallelized implementations are less maintainable, thereby supporting the need for automatic transformations.

Original languageEnglish
Pages (from-to)4001-4025
Number of pages25
JournalJournal of Supercomputing
Volume75
Issue number8
DOIs
Publication statusPublished - Aug 1 2019

Fingerprint

Energy Efficiency
Prediction Model
Computer hardware
Energy efficiency
Hardware
Electric power measurement
Software
Benchmark
Program processors
Deterioration
Field programmable gate arrays (FPGA)
Metric
Parallel Implementation
Execution Time
Field Programmable Gate Array
Predict
Configuration
Methodology
Computing
Energy

Keywords

  • Configuration selection
  • Green computing
  • Heterogeneous architecture
  • Performance optimization
  • Power-aware execution

ASJC Scopus subject areas

  • Software
  • Theoretical Computer Science
  • Information Systems
  • Hardware and Architecture

Cite this

Prediction models for performance, power, and energy efficiency of software executed on heterogeneous hardware. / Bán, Dénes; Ferenc, R.; Siket, István; Kiss, Ákos; Gyimóthy, T.

In: Journal of Supercomputing, Vol. 75, No. 8, 01.08.2019, p. 4001-4025.

Research output: Contribution to journalArticle

@article{3816c30ca7e7454db3879b1b687c0e97,
title = "Prediction models for performance, power, and energy efficiency of software executed on heterogeneous hardware",
abstract = "Heterogeneous computer environments are becoming commonplace so it is increasingly important to understand how and where we could execute a given algorithm the most efficiently. In this paper we propose a methodology that uses both static source code metrics, and dynamic execution time, power, and energy measurements to build gain ratio prediction models. These models are trained on special benchmarks that have both sequential and parallel implementations and can be executed on various computing elements, e.g., on CPUs, GPUs, or FPGAs. After they are built, however, they can be applied to a new system using only the system’s static source code metrics which are much more easily computable than any dynamic measurement. We found that while estimating a continuous gain ratio is a much harder problem, we could predict the gain category (e.g., “slight improvement” or “large deterioration”) of porting to a specific configuration significantly more accurately than a random choice, using static information alone. We also conclude based on our benchmarks that parallelized implementations are less maintainable, thereby supporting the need for automatic transformations.",
keywords = "Configuration selection, Green computing, Heterogeneous architecture, Performance optimization, Power-aware execution",
author = "D{\'e}nes B{\'a}n and R. Ferenc and Istv{\'a}n Siket and {\'A}kos Kiss and T. Gyim{\'o}thy",
year = "2019",
month = "8",
day = "1",
doi = "10.1007/s11227-018-2252-6",
language = "English",
volume = "75",
pages = "4001--4025",
journal = "Journal of Supercomputing",
issn = "0920-8542",
publisher = "Springer Netherlands",
number = "8",

}

TY - JOUR

T1 - Prediction models for performance, power, and energy efficiency of software executed on heterogeneous hardware

AU - Bán, Dénes

AU - Ferenc, R.

AU - Siket, István

AU - Kiss, Ákos

AU - Gyimóthy, T.

PY - 2019/8/1

Y1 - 2019/8/1

N2 - Heterogeneous computer environments are becoming commonplace so it is increasingly important to understand how and where we could execute a given algorithm the most efficiently. In this paper we propose a methodology that uses both static source code metrics, and dynamic execution time, power, and energy measurements to build gain ratio prediction models. These models are trained on special benchmarks that have both sequential and parallel implementations and can be executed on various computing elements, e.g., on CPUs, GPUs, or FPGAs. After they are built, however, they can be applied to a new system using only the system’s static source code metrics which are much more easily computable than any dynamic measurement. We found that while estimating a continuous gain ratio is a much harder problem, we could predict the gain category (e.g., “slight improvement” or “large deterioration”) of porting to a specific configuration significantly more accurately than a random choice, using static information alone. We also conclude based on our benchmarks that parallelized implementations are less maintainable, thereby supporting the need for automatic transformations.

AB - Heterogeneous computer environments are becoming commonplace so it is increasingly important to understand how and where we could execute a given algorithm the most efficiently. In this paper we propose a methodology that uses both static source code metrics, and dynamic execution time, power, and energy measurements to build gain ratio prediction models. These models are trained on special benchmarks that have both sequential and parallel implementations and can be executed on various computing elements, e.g., on CPUs, GPUs, or FPGAs. After they are built, however, they can be applied to a new system using only the system’s static source code metrics which are much more easily computable than any dynamic measurement. We found that while estimating a continuous gain ratio is a much harder problem, we could predict the gain category (e.g., “slight improvement” or “large deterioration”) of porting to a specific configuration significantly more accurately than a random choice, using static information alone. We also conclude based on our benchmarks that parallelized implementations are less maintainable, thereby supporting the need for automatic transformations.

KW - Configuration selection

KW - Green computing

KW - Heterogeneous architecture

KW - Performance optimization

KW - Power-aware execution

UR - http://www.scopus.com/inward/record.url?scp=85073008373&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85073008373&partnerID=8YFLogxK

U2 - 10.1007/s11227-018-2252-6

DO - 10.1007/s11227-018-2252-6

M3 - Article

AN - SCOPUS:85073008373

VL - 75

SP - 4001

EP - 4025

JO - Journal of Supercomputing

JF - Journal of Supercomputing

SN - 0920-8542

IS - 8

ER -