Latency Performance Model of Direct and K-Access Reads in Distributed Storage Systems

Abstract

Access latency is a crucial performance metric in distributed storage systems as it greatly impacts user experience, especially for hot data visitors. Existing papers argue that coding can reduce access latency compared with the replication method, and many dispatch schemes, such as those implementing redundant requests, dynamically changing code rates, and so on, are proposed and proved to work well in improving latency performance under certain conditions. However, some practical issues, such as direct reads, and the additional cost of redundant requests, are ignored and a general and practical model of latency performance is still lacking. Considering a more comprehensive set of practical issues in distributed storage systems, we propose a performance model that can be easily used to compare latency performance of different codes and replication methods under different conditions. We also use our performance model to evaluate many schemes and show their different impacts on the latency performance of different types of reads. To the best of our knowledge, we are the first to study the latency performance of direct reads under different codes or schemes, and the first to propose a model of latency performance including both direct and k-access reads.