Reliability and Performance of Mirrored Disk Organizations

Abstract

This is a followup to the 1994 tutorial by Berkeley RAID researchers whose 1988 RAID paper foresaw a revolutionary change in storage industry based on advances in magnetic disk technology, i.e., replacement of large capacity expensive disks with arrays of small capacity inexpensive disks. NAND flash SSDs which use less power, incur very low latency, provide high bandwidth, and are more reliable than HDDs are expected to replace HDDs as their prices drop. Replication in the form of mirrored disks and erasure coding via parity and Reed-Solomon codes are two methods to achieve higher reliability through redundancy in disk arrays. RAID(4+k), k=1,2,... arrays utilizing k check strips makes them k-disk-failure-tolerant with maximum distance separable coding with minimum redundancy. Clustered RAID, local recovery codes, partial MDS, and multilevel RAID are proposals to improve RAID reliability and performance. We discuss RAID5 performance and reliability analysis in conjunction with HDDs w/o and with latent sector errors -LSEs, which can be dealt with by intradisk redundancy and disk scrubbing, the latter enhanced with machine learning algorithms. Undetected disk errors causing silent data corruption are propagated by rebuild. We utilize the M/G/1 queueing model for RAID5 performance evaluation, present approximations for fork/join response time in degraded mode analysis, and the vacationing server model for rebuild analysis. Methods and tools for reliability evaluation with Markov chain modeling and simulation are discussed. Queueing and reliability analysis are based on probability theory and stochastic processes so that the two topics can be studied together. Their application is presented here in the context of RAID arrays in a tutorial manner.

Mirrored and hybrid disk arrays and their reliability

“…This mean response time holds if all requests are writes, but simulation results have shown that with interfering read requests the write response time is higher and closer to R max 2 , when the overall ρ remains the same. Performance analyses of RAID1 in normal and degraded mode are given in [16,103]. The performance of LSI RAID with RMW versus Reconstruct Write (RCW) method [94] is given in [107].…”

Analysis of Fork/Join and Related Queueing Systems

“…When a disk fails in basic mirroring, the rate of read requests to the surviving disk is doubled. A solution to this problem is to distribute the data of each disk over n > 1 disks so that if a disk fails, the load increase at surviving disks is 1/n [ Thomasian and Xu 2008;Thomasian and Tang 2012]. Clustered RAID5 to attain a similar effect is described later.…”

RAID level selection for heterogeneous disk arrays

“…When one of two disks fails in BM the read load on the surviving disk is doubled. RAID1 configurations with a smaller load increase than BM are described in [23,26]. The Interleaved Declustering (ID) layout, which was implemented in the Teradata DBC/1012 database machine, distributes the contents of each disk evenly on the remaining n − 1 disks of a cluster of n > 2 disks.…”

“…The Interleaved Declustering (ID) layout, which was implemented in the Teradata DBC/1012 database machine, distributes the contents of each disk evenly on the remaining n − 1 disks of a cluster of n > 2 disks. If a disk fails the increase of read load per disk is n/(n − 1) < 2 [23,26]. ID will yield a more balanced disk load in HDA following a disk failure, but only BM is considered in this study for the sake of brevity.…”