CSR-PDOS: Hardening Distributed Data Stores for Disaster Recovery

University Of Pittsburgh, Pittsburgh PA

Investigators

Abstract

In this project the investigators focus on improving the reliability of replicated data stores using an efficient and unconventional parity-based replication scheme the investigators call SSPiRAL. SSPiRAL (Survivable Storage using Parity in Redundant Array Layouts) is a data redundancy scheme that is easily tailored to the underlying infrastructure. SSPiRAL can be tailored to available resources and desired levels of fault-tolerance, capable of surviving multiple node failures with almost no degradation in performance. SSPiRAL is a near-optimal algorithm in terms of storage utilization, and yet it is distinguished by its adaptability to the available resource constraints, and its ability to sustain high performance in spite of the loss of component storage nodes. SSPiRAL has applications at multiple scales: improving the reliability of small storage arrays; improving the performance of large storage arrays; enabling more efficient grid storage; improving the reliability and maintainability of distributed storage; and enabling a novel approach to the construction of collaborative archival and backup stores. A prototype SSPiRAL layout is the focus in this project. Individual storage nodes for the prototype SSPiRAL system are built using commodity computers. The project aims to develop and refine the system for distributing data to the individual storage nodes, and to demonstrate the advantages in reliability and manageability offered by the SSPiRAL scheme.