Performance Evaluation of Low Latency Communication Alternatives in a Containerized Cloud Environment

Abstract

Minimized and stable communication latency opens up the possibility for various new cloud computing domains, ranging from industrial control to stateless network functions. In this paper we evaluate the related performance aspects of two kernel bypassing technologies, RDMA and DPDK, and plain kernel sockets in a containerized environment. We show that RDMA and DPDK can provide similar latency characteristics for short messages, while RDMA outperforms DPDK as the message size grows. We demonstrate that if CPU usage is a concern, plain UDP sockets are considerable alternatives to the bypassing solutions, however, if single digit microsecond latency is a must, a dedicated CPU is necessary even for RDMA. Finally, we show that a software based message reliability can provide the same, or in some cases even better latency characteristics than the hardware based solution of RDMA, and that noisy neighbors in a cloud environment may significantly impact the performance, even with dedicated CPU cores.