Allocations for heterogenous distributed storage

Abstract—We study the problem of storing a data object in a set of data nodes that fail independently with given probabilities. Our problem is a natural generalization of a homogenous storage allocation problem where all the nodes had the same reliability and is naturally motivated for peer-to-peer and cloud storage systems with different types of nodes. Assuming optimal erasure coding (MDS), the goal is to find a storage allocation (i.e, how much to store in each node) to maximize the probability of successful recovery. This problem turns out to be a challenging combinatorial optimization problem. In this work we introduce an approximation framework based on large deviation inequal-ities and convex optimization. We propose two approximation algorithms and study the asymptotic performance of the resulting allocations. SUBMITTED TO ISIT 2012. I