A performance analysis of system S, S4, and Esper via two level benchmarking

Abstract

Data stream processing systems have become popular due to their effectiveness in applications in large scale data stream processing scenarios. This paper compares and contrasts performance characteristics of three stream processing softwares System S, S4, and Esper. We study about which software aspects shape the characteristics of the workloads handled by these software. We use a micro benchmark and different real world stream applications on System S, S4, and Esper to construct 70 different application scenarios. We use job throughput, CPU, Memory consumption, and network utilization of each application scenario as performance metrics. We observed that S4's architectural aspect which instantiates a Processing Element (PE) for each keyed attribute is less efficient compared to the fixed number of PEs used by System S and Esper. Furthermore, all the Esper benchmarks produced more than 150% increased performance in single node compared to S4 benchmarks. S4 and Esper are more portable compared to System S and could be fine tuned for different application scenarios easily. In future we hope to widen our understanding of performance characteristics of these systems by investigating in to the code level profiling. © 2013 Springer-Verlag.