Optimizing Apportionment of Redundancies in Hierarchical RAID

TL;DR

This paper investigates how to optimally distribute redundancy in hierarchical RAID systems to maximize data reliability, revealing that intra-SN redundancy contributes more to reliability than inter-SN redundancy.

Contribution

It introduces an approximate reliability analysis combined with Monte-Carlo simulation to optimize redundancy apportionment in hierarchical RAID systems.

Findings

Higher intra-SN redundancy improves Mean-Time-to-Data-Loss.

Optimal redundancy distribution favors intra-SN over inter-SN levels.

Results challenge previous assumptions from IBM studies.

Abstract

Large disk arrays are organized into storage nodes -- SNs or bricks with their own cashed RAID controller for multiple disks. Erasure coding at SN level is attained via parity or Reed-Solomon codes. Hierarchical RAID -- HRAID -- provides an additional level of coding across SNs, e.g., check strips P, Q at intra-SN level and R at the inter-SN level. Failed disks and SNs are not replaced and rebuild is accomplished by restriping, e.g., overwriting P and Q for disk failures and R for an SN failure. For a given total redundancy level we use an approximate reliability analysis method and Monte-Carlo simulation to explore the better apportionment of check blocks for intra- vs inter-SN redundancy. Our study indicates that a higher MTTDL -- Mean-Time-to-Data-Loss -- is attained by associating higher reliability at intra-SN level rather than inter-SN level, which is contrary to that of an IBM study.