System Energy Efficiency Lab

Announcement: Postdoc/research assistantship positions available in SeeLab at UCSD

The focus of research is on datacenter energy efficiency with specific emphasis on redesigning the scheduling and resource management policies for virtual machines (e.g. using Xen). This project is a part of The Multiscale Systems Center (MuSyC center) and involves collaboration with Google, Microsoft, Intel and others. PhD or graduate students with background and interest in operating systems, computer architecture and/or applications of statistical modeling and learning techniques are encouraged contact Prof. Tajana Rosing (tajana at ucsd.edu).

Motivation and Goals

The cost of energy consumption in modern data centers has reached and even surpassed the cost of the physical data center itself. This necessitates research for dynamically reducing the amount of energy used for computing, cooling and maintaining a data center. The primary motivation for this work is to develop a data center power management scheme that delivers energy efficiency while minimizing the impact on performance.

Research overview

Our approach is based on developing policies for power management techniques like dynamic power management (DPM) and dynamic voltage frequency scaling (DVFS) based on online learning for a computing system. Such an approach reduces these problems to one of dynamic workload characterization, where the policies adapt to changes in the workloads. Our experiments with CPU and hard disks confirm the efficiency and adaptability of our online learning based policies. We further propose extensions to adapt this approach in a virtualized environment encompassing multiple virtual and physical machines, which is fairly common in modern data centers. The idea is to perform characterization of virtual machines at the hypervisor level in order to drive both the power management policies and energy aware scheduling. The energy aware scheduler will schedule virtual machines both within and across physical machines for higher energy efficiency.

Online learning algorithm

Our online learning approach is based on an adaptation of the online allocation algorithm. The algorithm allows us to formulate both dynamic power management (DPM) and dynamic voltage frequency scaling (DVFS) problems as one of workload characterization and selection. The selection is done among a set of experts, which refers to a set of DPM policies and voltage-frequency settings, leveraging the fact that different experts outperform each other under different workloads and device leakage characteristics. The online learning algorithm adapts to changes in the characteristics and guarantees fast convergence to the best performing expert.

Workload characterization

The suitability of a given DPM policy or a voltage-frequency setting for a workload is determined by certain properties or characteristics of that workload. On general purpose systems where these properties are not known in advance, it is necessary to estimate them dynamically for effective power management. We refer to this process as dynamic workload characterization. Thus, in our approach we dynamically characterize the workloads and feed this information to the online learning algorithm to determine the best "expert" for the workload based on its current characteristics.

Publications

Journals

Dhiman, G. and Rosing, T.S., "System Level Power Management Using Online Learning" In IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , Vol. 28 , Nr. 5 , May (2009) , p. 676-689. [pdf]

Conferences

Dhiman, G., Marchetti, G. and Rosing, T.S., “vGreen: A System for Energy Efficient Computing in Virtualized Environments” In Proceedings of the 14th IEEE/ACM International Symposium on Low Power Electronics and Design, 2009, ISLPED ’09.
[Acceptance Rate (Regular Papers) = 52/210 = 24.7%] [pdf]
(Best Paper Nominee)

Dhiman, G., Pusukuri, K.K. and Rosing, T. S., “Analysis of Dynamic Voltage Scaling for System Level Energy Management” In Proceedings of the 2008 Workshop on Power Aware Computing and Systems, HotPower'08. [pdf]

Dhiman, G. and Rosing, T. S., “Dynamic voltage frequency scaling for multi-tasking systems using online learning” In Proceedings of the 2007 International Symposium on Low Power Electronics and Design, ISLPED '07.
[Acceptance Rate (Regular Papers) = 56/192 = 29%] [pdf]

Dhiman, G. and Rosing, T. S., “Dynamic power management using machine learning” In Proceedings of the 2006 IEEE/ACM international Conference on Computer-Aided Design, ICCAD '06. [Acceptance Rate = 130/541 = 24%] [pdf]
(Best Paper Nominee)

Resources

Energy efficiency in virtualized environments - Poster

Related Work

People

Tajana Simunic Rosing, UCSD CSE Professor

Gaurav Dhiman, UCSD PhD Candidate

Raid Ayoub,, UCSD PhD Candidate

Giacomo Marchetti, University of Bologna MS Student

Sponsored by:

NSF-GreenLight project, CNS, Sun Microsystems, UC Micro, Cisco, GSRC/DARPA