This project focuses on methods and tools for improving the energy efficiency of computing in enterprise workspaces, data centers, and cloud infrastructures.
The growth rate of the energy use of computing infrastructures is significantly higher than that of total energy use and that of the production capacity. In several instances, building facilities have restricted the number of new PCs or servers that can be plugged in simply because the utilities do not have the means to provide additional power capacity. At the same time, there is a large potential for optimizing the design and operation of computing systems, at several levels including the hardware architecture, server engineering, and the design of system software and applications.
Our research focuses specifically on improving the energy use of system software, online services, and applications. We begin with tools to gain better insight into where the energy is being spent, drilling down deeper from the total computer power into individual hardware components as well as software containers such as virtual machines and software processes. We are also developing methods and techniques to optimize energy use within a computer, in the enterprise network, and in large scale virtualized cloud data centers.
Joulemeter is a software tool to estimate computational power consumption. It allows modeling the power use of software components such as virtual machines and processes that do not have a physical power supply to be metered. It also allows modeling the power use of individual hardware components within a computer such as the processor, screen, and storage disks that are hard to measure individually. The visibility gained is useful for improving battery life in mobile computing devices as well as reducing energy costs in large scale computing infrastructures. A downloadable version is available to play with some of the Joulemeter capabilities. More...
Virtualization is a powerful technique to reduce energy use through workload consolidation. However, the resource isolation provided by current virtualization technologies does not strictly partition certain shared hardware resources such as hardware caches and memory bandwidth. This leads to performance degradation upon consolidation. This prevents performance sensitive applications from reaping the benefits of consolidation. Also, this degradation reduces the energy savings since consolidated workloads run longer. In Cuanta, we are developing techniques to predict such degradation, and to enable resource and performance aware consolidation. Cuanta makes virtualization based energy savings accessible to a larger class of software applications and services by actively quantifying and controlling performance. More...
Virtualized Power Shifting (VPS)
The cloud presents an immense opportunity for resource efficiency through economies of scale, and statistical multiplexing among numerous hosted workloads. However, the large scale hosting platforms also require enormous amounts of power, leading to high power provisioning costs. Power budgeting is often employed to manage and reduce power provisioning costs and modern server hardware provides budgeting mechanisms. The VPS projects extends these power budgeting mechanisms to virtualized infrastructures where hardware budgeting is not directly applicable due to sharing of server hardware among multiple applications. More...
- Di Wang, Sriram Govindan, Anand Sivasubramaniam, Aman Kansal, Jie Liu, and Badriddine Khessib, Underprovisioning Backup Power Infrastructure for Datacenters, in 19th International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), ACM, 1 March 2014
- Di Wang, Chuangang Ren, Sriram Govindan, Anand Sivasubramaniam, Bhuvan Urgaonkar, Aman Kansal, and Kushagra Vaid, ACE: Abstracting, Characterizing and Exploiting Datacenter Power Demands, in IEEE International Symposium on Workload Characterization (IISWC), IEEE, 22 September 2013
- Sriram Sankar, Aman Kansal, and Jie Liu, Towards a Holistic Data Center Simulator, in ASCR Workshop on Modeling & Simulation of Exascale Systems & Applications (MODSIM), Other, 18 September 2013
- Alan Roytman, Aman Kansal, Sriram Govindan, Jie Liu, and Suman Nath, PACMan: Performance Aware Virtual Machine Consolidation, in 10th International Conference on Autonomic Computing (ICAC), USENIX, 26 June 2013
- Di Wang, Chuangang Ren, Sriram Govindan, Anand Sivasubramaniam, Bhuvan Urgaonkar, Aman Kansal, and Kushagra Vaid, ACE: Abstracting, Characterizing and Exploiting Peaks and Valleys in Datacenter Power Consumption, in ACM SIGMETRICS (Posters), ACM, 17 June 2013
- Aman Kansal, Bhuvan Urgaonkar, and Sriram Govindan, Using Dark Fiber to Replace Diesel Generators, in XIV Workshop on Hot Topics in Operating Systems (HotOS), USENIX, 13 May 2013
- Alan Roytman, Aman Kansal, Sriram Govindan, Jie Liu, and Suman Nath, Algorithm Design for Performance Aware VM Consolidation, no. MSR-TR-2013-28, 4 March 2013
- Aman Kansal, Building a More Efficient Data Center - from Servers to Software, Microsoft Research, 1 February 2013
- Christina Delimitrou, Sriram Sankar, Aman Kansal, and Christos Kozyrakis, ECHO: Recreating Network Traffic Maps for Datacenters with Tens of Thousands of Servers, in IEEE International Symposium on Workload Characterization , IEEE, 4 November 2012
- Arka Bhattacharya, Aman Kansal, David Culler, Sriram Sankar, and Sriram Govindan, The Need for Speed and Stability in Data Center Power Capping, in Third International Green Computing Conference (IGCC'12), 5 June 2012