Performance sensitivity of Operating System parameters in Microservice environments

Cloud services are undergoing a major shift from monolithic applications to graphs of hundreds of loosely-coupled microservices. Microservice architecture presents both opportunities and challenges while optimizing for system characterization and resource utilization. In this study, we characterize the performance sensitivity of Operating System parameters for two academic microservice benchmarks with respect to their thread scheduling, disk writes, and the TCP buffer. We demonstrate that through simple affinitization of kernel threads to cores of a Non-Uniform Memory Access processor and disk buffer configurations, there is a 26% and 19% decrease in Average Latency accompanied by a 27% decrease in 99% Tail latency for the benchmarks under consideration. While performance improvements can be gained, we also show that alternate nonoptimal configurations of the system parameters degrade the performance substantially.