Power consumption is a major problem for data centers of all sizes which impacts the density of servers and the total cost of ownership. This work developed and measured a solution based on write off-loading which enables prolonged periods of disk spin-down. Our results show that this provides power savings of 45—60%. This work was presented at the Usenix FAST conference in February 2008.
Power consumption is already driving changes in data center configuration and management. Some components already support some power management features, for example server CPUs support dynamic clock and voltage scaling that enables power requirements to be reduced significantly during idle periods. Storage subsystems do not have power management and are consume a significant amount of power in the data center. Modern enterprise grade disks require approximately 10W when idle. As storage requirements generally increase in data centers, the number of disks in data centers is increasing.
Based on 1-week long traces of core servers in our data center, we have found that there are significant periods of idle time during which disks can be spun down, and even longer “write-only” periods during which all I/O operations are writes. Based on this we have developed a technique called “write off-loading” which allows disks to stay spun down during these write-only periods, by temporarily off-loading the write requests to other volumes in the data center.
- Dushyanth Narayanan, Austin Donnelly, and Antony Rowstron, Write Off-Loading: Practical Power Management for Enterprise Storage, in Proceedings of 6th USENIX Conference on File and Storage Technologies (FAST ’08), USENIX, San Jose, CA, February 2008.
The traces used in this paper can be downloaded from here and have also been submitted to the SNIA trace repository. Please read the README.txt file in the download directory for details about the trace format, tracing methodology etc.