PRISM: PREDICTION-GUIDED RISK-AWARE DEEP REINFORCEMENT LEARNING FOR ENERGY-AWARE DYNAMIC RESOURCE SCHEDULING IN DISTRIBUTED CLOUD DATA CENTERS
Journal of Computer Science and Electrical Engineering · 05.12.2022 · 4 (1) · 53-61
DOI: 10.61784/jcsee4155 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Computer Science and Electrical Engineering, online erschienen am 05.12.2022. Verfasst von EnMiao Feng und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Geo-distributed cloud data centers spend a large share of their electricity on lightly loaded servers, while aggressive consolidation causes overloads, service-level agreement (SLA) violations and migration storms. This paper presents PRISM, a prediction-guided, risk-aware deep reinforcement learning (DRL) framework for energy-aware dynamic resource scheduling in distributed cloud data centers. PRISM couples three components. A lightweight multi-horizon quantile forecaster (MQF) predicts the mean, median and 90th percentile of the CPU demand of every virtual machine (VM), including newly arrived VMs without history. A risk-aware host model aggregates these quantiles into a chance-constrained capacity test that drives overload detection and placement feasibility at a tunable risk level. A residual Q-network, whose action value is an analytic energy-cost prior plus a learned correction, scores a compact set of candidate hosts for arrival dispatch, overload relief and consolidation across all data centers and is trained with Double DQN. Trace-driven experiments with 11,746 real PlanetLab VM traces on three heterogeneous data centers with SPECpower server models, time-of-use tariffs and different PUE values show that, on unseen test days and relative to the strongest baseline, PRISM reduces energy consumption by 12.5%, electricity cost by 5.2%, carbon emissions by 14.1%, the SLA violation time per active host by 44.6%, the energy–SLA-violation product by 90.1% and the number of migrations by 84.6%, while requiring 16.3 ms of computation per five-minute scheduling interval.
Zitieren
APA
Feng, E., Li, P., Mo, T., & Wang, G. (2022). PRISM: PREDICTION-GUIDED RISK-AWARE DEEP REINFORCEMENT LEARNING FOR ENERGY-AWARE DYNAMIC RESOURCE SCHEDULING IN DISTRIBUTED CLOUD DATA CENTERS. Journal of Computer Science and Electrical Engineering, 4(1), 53-61. https://doi.org/10.61784/jcsee4155
Vancouver
Feng E, Li P, Mo T, Wang G. PRISM: PREDICTION-GUIDED RISK-AWARE DEEP REINFORCEMENT LEARNING FOR ENERGY-AWARE DYNAMIC RESOURCE SCHEDULING IN DISTRIBUTED CLOUD DATA CENTERS. Journal of Computer Science and Electrical Engineering. 2022;4(1):53-61. doi:10.61784/jcsee4155
BibTeX
@article{Feng2022,
author = {Feng, EnMiao and Li, PengFei and Mo, TianJun and Wang, GaiKe},
title = {PRISM: PREDICTION-GUIDED RISK-AWARE DEEP REINFORCEMENT LEARNING FOR ENERGY-AWARE DYNAMIC RESOURCE SCHEDULING IN DISTRIBUTED CLOUD DATA CENTERS},
journal = {Journal of Computer Science and Electrical Engineering},
year = {2022},
volume = {4},
number = {1},
pages = {53--61},
doi = {10.61784/jcsee4155}
}
EUVIMED