Storage and repair bandwidth tradeoff for heterogeneous cluster distributed storage systems

Abstract

The storage and repair bandwidth tradeoff is an important issue in distributed storage systems (DSSs) where large scale data are stored in multiple nodes with erasure coding to ensure reliability. There are lots of studies on the DSS model with multiple clusters where the repair bandwidths from intra-cluster and cross-cluster nodes are differentiated to improve repair efficiency based on the realistic network topological structures. At the same time, separate nodes are also prevalent due to the variety of practical networks, but the work on the cluster DSS model with multiple separate nodes is insufficient, which is a main motivation of this paper. We formulate the tradeoff bound between storage repair bandwidth for a heterogeneous DSS model consisting of clusters and separate nodes by analyzing the min-cuts of heterogeneous information flow graphs corresponding to the orders of failed nodes. Additionally, the tradeoff bounds are investigated in multiple aspects when the repair bandwidth constraints and the amount of separate nodes vary, respectively. Moreover, a class of regenerating codes are illustrated to achieve the tradeoff in the heterogeneous cases.

Regenerating codes on graphs

“…Constructions with nonuniform download are considered when the transmission cost from different helpers to the failed node is not the same, giving rise to heterogeneous regenerating codes. Different versions of such systems include models with two subsets of nodes assigned different communication costs [1], systems with different link capacities [35], rack-aware storage [5], [6], [12], [13], [32], [33], general models of clustered systems [25], [31], systems with varying amounts of data downloaded from the helper nodes [34], [16], and general systems with topology-aware repair [36].…”