Ensuring Data Durability with Increasingly Interdependent Content

Abstract

Data entanglement is a novel approach to generate and propagate redundancy across multiple disk nodes in a fault-tolerant data store. In this paper, we analyse and evaluate helical entanglement codes (HEC), an XOR-based erasure coding algorithm that constructs long sequences of entangled data using incoming data and stored parities. The robust topology guarantees low complexity and a greater resilience to failures than previous codes mentioned in the literature, however, the code pattern requires a minimum fixed amount of storage overhead. A unique characteristic of HEC is that fault tolerance depends on the number of distinct helical strands (p), a parameter that could be changed on the fly and does not add significantly more storage. A p-HEC setting can tolerate arbitrary 5+p failures. Decoding has a low reconstruction cost and good locality. Besides, a deep repair mechanism exploits the available global parities. We perform experiments to compare the repairability of HEC with other codes and present analytical results of its reliability.