Introduction
|
The primary motivation for this work comes from the recent growth in the
energy usage of desktop computers. Recent
studies estimate that PCs and monitors consume about 100TWh/year, out of
which 65TWh/year is consumed by enterprise PCs. More statistics can be found
in LBNL studies.
We have conducted our own measurements as part of this project, and found that many short idle periods exist (Figure 1), which are not exploited using usual power management mechanisms. User activity and background processes make it difficult for power management mechanisms to kick in. Though proxying can help, developing application layer proxies is cumbersome, and may cause some user disruption. |
Figure 1
|
Approach: Exploit Virtualization
|
Our idea is to run the desktop in a virtual machine, and migrate it to a central server, when the VM (running the desktop) is idle. The original host is switched off to save energy. After the VM is migrated to the server, its memory is reduced to allow high consolidation of VMs. Figure 2 shows the architecture of our system. Key challenges to realize this are
|
Figure 2
|
Deployment
|
Efforts are underway at Microsoft Research to bring LiteGreen to enterprises. A small-scale deployment at MSR India show that LiteGreen can save up top 74% of energy compared to 32% with existing Windows and manual power management. A video showing the LiteGreen usage is available [Low res (8.1MB WMV), High res (28MB WMV)]. See this page for more details. |
Figure 3
|
People
- Tathagata Das, Microsoft Research India
- Pradeep Padala (Part of the work is done during my summer'09 internship at MSR India)
- Venkat Padmanabhan, Microsoft Research India
- Ram Ramjee, Microsoft Research India
- Kang G. Shin, U. Michigan
Publications
- Tathagata Das, Pradeep Padala, Venkat Padmanabhan, Ram Ramjee, and Kang G. Shin. LiteGreen: Saving Energy in Networked Desktops Using Virtualization. In the proceedings of USENIX Annual Technical Conference (USENIX ATC '10), Jun 2010. [PDF]. Best paper award!
Figure 1
Figure 2
Figure 3