Improving Disk Array Reliability Through Faster Repairs

Introduction

Magnetic disk capacities have grown over the last decades by a factor of at least ten thousand. While the minicomputer disk drives of the late eighties could only store 600 MB of data, 8TB or 10TB disk drives are common today. The same is not true for disk transfer rates: they are just one hundred times higher than those of the late eighties. As a result, copying the entire contents of a disk will take now considerably more time than thirty or forty years ago.

This development has now a critical impact on the reliability of disk arrays. Magnetic disks are the least reliable component of most computer systems, with failure rates that can sometimes reach 25 percent per year. As a result, all medium-size to large-size disk arrays include some redundancy and provisions for the quick replacement of failed units. Mirroring, RAID level 5 and RAID level 6 are the best known exemplars of these redundant disk array organizations. In all cases, the level of protection that these organizations offer depends on both the prompt detection of disk failures and the time it will take to reconstitute the contents of failed disks on spare devices. In particular, any increase in this reconstruction time will have a negative impact on the reliability of the array.

A well-known solution to this problem is declustering. Declustering partitions each physical disk into k physically contiguous parts, which we will call disklets, and combines these disklets into distinct reliability stripes. The scheme requires that each of these parity stripe have no more than one disklet located on any physical disk.

The main advantage of declustering disk array organizations is their faster recovery. Once a disk has failed, we can recover the contents of its k disklets in parallel through their own parity stripes. This will allow us to reconstruct the lost data k times faster as long as no two of the k parity stripes has to access disklets located on the same physical disk.

To the best of our knowledge, none of the recent papers analyzing the fault-tolerance of disklet-based storage systems have ever tried to estimate their five-year reliability, preferring instead to evaluate the probability of a data loss in the presence of a fixed number of failures.

We present here a preliminary study of the effect of declustering on the five-year reliability of disk arrays. We chose for our investigation an array consisting of ten RAID level 6 reliability stripes with ten disks each for a total of one hundred disks. Such organization is large enough to demonstrate the benefits of declustering while remaining small enough to let us neglect the impact of non-fatal quadruple disk failures on the array reliability.