Exploiting unused storage resources to enhance systems' energy efficiency, performance, and fault-tolerance.

Abstract

The invention of better fabrication materials and processes in solid-state devices has led to unprecedented technological breakthroughs in computer hardware. Today's system software, however, often cannot take full advantage of the hardware's rapidly improving capabilities, thus resulting in idling resources, e.g., unoccupied memory and disk space. To make hardware operate more efficiently and to reduce the amount of idle resources, this thesis proposes several techniques that can harness such resources to the benefit of users. Although there are many different types of hardware resources, this thesis focuses on the reclamation of idle resources in the storage hierarchy. First, we implemented a pure software technique to reduce the power dissipation of main memory. By aggregating unmapped and unused memory pages and powering down unused memory ranks, a significant amount of energy can be saved with little or no performance degradation. Next, we explored several architectural-level solutions that can more aggressively reduce energy, but at the expense of performance. We also developed techniques to exploit unused disk capacity to improve disks' performance and energy-efficiency. These techniques are realized in our implementation of the Free Space File System (FS2). Unlike traditional file systems where extra disk space is not used, FS2 actively makes use of it to hold replicas of temporally-related data blocks that were poorly placed by the underlying file system. Using contiguous regions of free disk space to place related data blocks closer to one another enables in disk heads to work more efficiently. Free disk space may also be used to enhance the fault-tolerance of disks. The placement of replicas is shown to be critical to both the fault-tolerance and performance of file systems that make use of replicas. However, without a thorough understanding of how disks fail and how data become corrupted when failures occur, good data placement strategies are difficult to devise. We studied a large number of failed disks and analyzed their failure characteristics. This characterization study will help design more fault-tolerant file systems that can take advantage of today's large capacity hard drives.