Greedy window stochastic optimization algorithm for reducing data center energy consumption
Résumé
This paper presents a stochastic model for Dynamic Power Management (DPM) in data centers-a strategy that involves dynamically activating and deactivating servers to balance energy efficiency with the maintenance of high Quality of Service (QoS). Unlike traditional methods that rely on predefined job arrival distributions, our approach leverages histograms to dynamically characterize job arrival patterns derived from real-world traces, empirical data, or live traffic measurements. We model the data center as a queuing system and introduce an efficient Greedy Window algorithm that is fast, memory-efficient, and capable of adapting to real-time fluctuations in job arrivals. This algorithm computes a suboptimal policy that closely approximates the optimal strategy attainable through stochastic combinatorial optimization techniques such as the Markov Decision Process (MDP). However, MDP-based approaches face challenges related to state space explosion and substantial computational demands in both time and memory. We evaluate our method across a range of data center configurations using real Google traffic traces and compare its performance against several optimization strategies.
Domaines
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |