Locality-aware fountain codes for massive distributed storage systems

Abstract

Low repair locality of a distributed storage code has been shown to reduce strain on storage node input-output (I/O) resources during node repair operations after a failure. In this paper, we consider the use of Fountain codes for distributed storage systems and aim to understand the relationship between repair locality and code parameters for a systematic Fountain code. While the information-theoretic trade-off between repair locality and storage overhead has been understood and characterized, the challenge of choosing a locality value that satisfies multiple storage system design metrics is yet to be resolved. We approach this problem by deriving an expression for the probability distribution of repair locality in terms of the rateless code degree distribution coefficients and suggest that factoring this relationship into the code design process enables the design of rateless codes better adjusted to the needs of a massive distributed storage system.