Towards the Prediction of the Performance and Energy Efficiency of Distributed Data Management Systems

Abstract

The ability to accurately simulate and predict the metrics (e.g. performance and energy consumption) of data management systems offers several benefits. It can save investments in both time and hardware. A prominent example is the resource planning. Given a specific use case, a datacenter operator is able to find the most performant or most energy efficient configuration without performing benchmarks or aquiring the necessary hardware. Another possibility would be to study the effects of architectural changes without having them implemented. In this paper, Queued Petri Nets were used to predict and to study the performance and energy consumption of a distributed data management system like Cassandra. The prediction accuracy was evaluated and compared to actual experimental results. On average, the predicted and experimental results differ only by 8 percent for the performance and 16 percent for the energy efficiency, respectively. In addition to this, the experimental results of the used Cassandra cluster revealed a super-linear behavior for the performance and a sub-linear one for the energy consumption.