from cs.IT updates on arXiv.org http://arxiv.org/abs/1202.1596

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 inequalities and

convex optimization. We propose two approximation algorithms and study the

asymptotic performance of the resulting allocations.