Cluster-based systems consume significant amounts of power and energy. Research on power and energy conservation for these systems can ease their deployment and installation, reduce costs, and protect the environment. Even when it is not possible to reduce the maximum power requirements of a cluster, reducing the common-case power and energy consumption reduces its operational and electricity costs.

With these advantages in mind, we are currently designing the first power-aware cluster-based file system. The key idea behind the file system is to statically bias the demand for resources towards certain cluster nodes. This bias is created mainly by allocating files to different disks according to their popularity. For file systems that exhibit a significant disparity in file popularity (such as those of WWW servers, for instance, which are known to exhibit a Zipf distribution of file accesses), this allocation strategy causes certain disks to be active more frequently than others, allowing our system to put the latter disks in low-power modes frequently and for long periods of time.

Compared to a cluster-based file system that is oblivious to power and energy considerations, such as Frangipani, our file system can provide significant power and energy savings, but the performance of accesses to unpopular files may be degraded. Compared to a file system in which disks are put in low-power modes based on inactivity thresholds, our system can provide higher performance for popular file accesses, lower performance for unpopular file accesses, and additional power and energy savings. To confirm these intuitions, we have collected a few preliminary results by simulating our system for a real WWW server trace. Our results show that file allocation based on popularity can indeed provide significant power and energy savings, while achieving overall performance that is as good as that of a power-oblivious file system.