Minimization of Execution Time and Energy Consumption of Parallel Applications on Heterogeneous Hybrid Parallel Computers Through Optimal Combination of Workload Distribution and Operating Voltage and Frequency
IEEE Transactions on Computers · 2026-09 · 75 (9) · 3410-3424
DOI: 10.1109/tc.2026.3711149 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Computers, 2026-09. Verfasst von Hesam Nejati Sharifaldin und 3 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 35 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Taighde Éireann – Research Ireland.
Abstract (Originalsprache)
The paper addresses the problem of optimization for performance and energy of parallel applications on heterogeneous hybrid parallel computers. State-of-the-art solutions optimize only either the configuration of the platform or the configuration of the application but not both. In this work, we propose a holistic solution of this problem. We mathematically formulate the problem as a bi-objective minimization problem for execution time and energy consumption of a given workload on $p$ p heterogeneous processors, where each processor is modeled by a pair of functions of two variables, processor configuration and workload size, returning the execution time and dynamic energy consumption, correspondingly. The problem has two decision variables: a vector of $p$ p processor configurations (representing the platform configuration) and a vector of $p$ p workload sizes (workload distribution representing the application configuration). We then propose exact algorithms that solve this optimization problem. We apply this solution to a heterogeneous hybrid server comprising a multicore CPU and GPUs, where the processor configuration is represented by a set of (voltage, frequency) pairs for independently powered processor components. The crucial problem in the practical application of the proposed algorithms is the high cost of construction of the input execution time and dynamic energy functions for all platform configurations. To address this problem, we propose and validate a circuit-level Dennard scaling power model and use it to derive scaling models of the execution time and dynamic energy as functions of operating voltages and frequencies of processors consisting of one and two independently powered components. These models are employed in our scaling method, which only requires one and two pairs of experimental time and energy profiles to build the complete set of profiles for the processor, which consists of one and two independently powered components, respectively. Finally, we demonstrate the efficiency and accuracy of our solution in optimizing 2D fast Fourier transform and mixed-precision GEMM hybrid parallel applications on a heterogeneous platform comprising an Intel Icelake multicore CPU and two Nvidia A40 GPUs.
Förderung
- Taighde Éireann – Research Ireland
Zitieren
APA
Sharifaldin, H. N., Ghoshdastidar, S., Manumachu, R. R., & Lastovetsky, A. (2026). Minimization of Execution Time and Energy Consumption of Parallel Applications on Heterogeneous Hybrid Parallel Computers Through Optimal Combination of Workload Distribution and Operating Voltage and Frequency. IEEE Transactions on Computers, 75(9), 3410-3424. https://doi.org/10.1109/tc.2026.3711149
Vancouver
Sharifaldin HN, Ghoshdastidar S, Manumachu RR, Lastovetsky A. Minimization of Execution Time and Energy Consumption of Parallel Applications on Heterogeneous Hybrid Parallel Computers Through Optimal Combination of Workload Distribution and Operating Voltage and Frequency. IEEE Transactions on Computers. 2026;75(9):3410-3424. doi:10.1109/tc.2026.3711149
BibTeX
@article{Sharifaldin2026,
author = {Sharifaldin, Hesam Nejati and Ghoshdastidar, Sourodip and Manumachu, Ravi Reddy and Lastovetsky, Alexey},
title = {Minimization of Execution Time and Energy Consumption of Parallel Applications on Heterogeneous Hybrid Parallel Computers Through Optimal Combination of Workload Distribution and Operating Voltage and Frequency},
journal = {IEEE Transactions on Computers},
year = {2026},
volume = {75},
number = {9},
pages = {3410--3424},
doi = {10.1109/tc.2026.3711149}
}
EUVIMED