Cloud Storage Workload Characterization: An Approach with Time-Series Analysis

Abstract

The cloud hosts diverse applications with different workload characteristics. Public cloud traces provide opportunities for analysis to gain insights informing autoscaling, forecasting among other operations. This paper presents the statistical analysis of a recent Alibaba cloud storage workload. The isolation & aggregation of all read/write time-series per recorded workload was done. Application of statistical methods yielded novel distributions from which forecasting solutions integrating time-varying variance captured workload burstiness. A 25% improvement in forecasting accuracy over current methods was achieved. The set of workload time-series has been made available online for further analysis by the research community.