Latency-Optimal File Assignment in Geo-Distributed Storage with Preferential Demands

TL;DR

This paper addresses optimal file placement in geo-distributed storage systems with non-uniform demands, aiming to minimize worst-case and average latency by modeling the problem as graph coloring and balanced assignment.

Contribution

It introduces a novel approach for optimal uncoded file placement considering non-uniform demands, extending prior work that assumed uniform demand probabilities.

Findings

Optimal uncoded placement scheme derived for non-uniform demands

Latency minimization achieved through vertex coloring and balanced assignment techniques

Framework applicable to systems with diverse demand distributions

Abstract

We consider the problem of data storage in a geographically distributed (or geo-distributed) network of servers (or nodes) where inter-node communication incurs certain round-trip delays. Every node serves a set of users who can request any file in the network. If the requested file is not available at the node, it communicates with other nodes to obtain the file, thus causing the user to experience latency in obtaining the file. The files can be placed uncoded, where each node stores exact copies of the files, or in coded fashion, where certain linear combination of files are placed at each node. We aim to obtain an optimal file placement on the nodes with respect to minimizing the worst-case latency at each node, as well as the system-average latency. The prior literature considered the case of equiprobable file demands at the nodes. In this paper, we investigate the generic case of non-uniform file-demand probabilities at each node. The scheme presented here is optimal within the family of uncoded schemes. It is obtained first by modeling the worst-case latency constraint as a vertex coloring problem, and then converting the system-average latency optimization to a problem of balanced-assignment.