Optimal sleeping: models and experiments for energy-delay tradeoff

Abstract

Multi-core architectures have supplanted single core schemes, in part because the maximum clock speed of a single core is limited by its energy consumption. They provide the additional, less exploited benefit of allowing a finer trade-ff between energy consumption and delay by turning off subsets of cores. We investigate how this tradeoff varies with the number of cores, and whether heterogeneity brings additional benefits to outweigh its increased complexity. We study optimal sleep policies in two settings: switching of homogeneous cores on a fast timescale, which models multiple cores in a CPU, and switching of heterogeneous cores on a slow timescale, which models different generation servers in a data centre. In the homogeneous-core case, we show the optimal policy is monotone hysteretic, and that the performance is less sensitive to load estimation errors. By implementing this policy on a real experimental testbed, we show that theory makes a good prediction of the resource pooling benefits, and extend our study to realistic bursty job arrivals and heavy-tailed job sizes. In the heterogeneous-core case, we provide a low complexity algorithm to find a combination of servers that have approximately the minimum power draw while providing a specified minimum processing speed.

Optimal energy-efficient policies for data centers through sensitivity-based optimization

“…In general, such a study is more interesting, difficult and challenging due to the fact that a complicated queueing model with synchronously multiple control objectives (e.g., reducing energy consumption, reducing system response time and guaranteeing quality of service) needs to be synthetically established in a Markov decision process. For a data center with multiple identical servers, Kamitsos et al [23][24][25] constructed a discrete-time Markov decision process by uniformization and proved that the optimal sleep energyefficient policy is simply hysteretic. Hence, this problem has a double threshold structure by means of the optimal hysteretic policy given in Hipp and Holzbaur [19] and Lu and Serfozo [32].…”

Scalable Load Balancing Scheme for Distributed Controllers in Software Defined Data Centers

“…Its shown that the optimal sleeping policy for the above MDP is a monotone hysteretic policy characterized by two queue length thresholds enabling to tune the modem from on to off or vice versa. Authors in [7] [8] [9] show that in a data center infrastructure based on multiple identical servers the optimal sleep energy efficient policy may be hysteretic and governed a double threshold.…”

Optimal Asynchronous Dynamic Policies in Energy-Efficient Data Centers

“…To date, little work has been done on applications of Markov decision processes (MDPs) to find the optimal dynamic control policies of energy-efficient data centers. Readers may refer to recent publications for details, among which Kamitsos et al [23] constructed a discrete-time MDP and proved that the optimal sleep energy-efficient policy is simply hysteretic, so that it has a double threshold structure. Note that such an optimal hysteretic policy follows Hipp and Holzbaur [24] and Lu and Serfozo [25].…”