Improving reliability and performance of redundant disk arrays by improving rebuild time and response time.

Abstract

Today's computer systems entrust the bulk of their data to disk drives. This data must be reliably stored and be quickly accessible. Dramatic improvements in microprocessor and main memory speeds, however, have far outpaced improvements in disk performance time. Redundant disk arrays have been proposed for addressing this bottleneck. Recently popularized by the RAID project, redundant disk arrays provide good performance by utilizing large numbers of disks while ensuring the data is reliably stored. Their performance, however, is lower than non-redundant disk arrays. In addition, as we move to an on-line society and store more data on disk, the need will exist for even higher data reliability. This dissertation proposes a new piggybacking rebuild algorithm that reduces the rebuild time (the time required to rebuild a failed disk in a disk array) and therefore improves the disk array's reliability. Rebuild time is reduced up to 32% for a mirrored array and up to 28% for a RAID5 array compared to the Redirection of Reads rebuild algorithm. This dissertation also shows that non-volatile storage provides a big win in terms of both rebuild time and performance. Using Track Piggybacking with non-volatile storage, a RAID5 array can rebuild a failed disk under an 83% heavier workload and still reduce the rebuild time by 45% compared to Redirection Reads without non-volatile storage. A mirrored array can rebuild a failed disk under a 54% heavier workload and still reduce the rebuild time by 46%. Using non-volatile storage also improves response times. Write request response times drop from 225 milliseconds to less than two milliseconds while read response times are lowered by 67%.