TailCmp - A Tail Latency Evaluation Solution of Public Cloud and Labeled von Neumann Architecture based Cloud Prototype

Xiangyu Kong, Xuesong Gao, Shibao Pan, Yizhi Zhou, Yanan Yang, Laiping Zhao, Heng Qi

Abstract: Tail latency is the key performance metric for cloud computing systems (CCSs). When latency-critical (LC) applications and best-effort batch (BE) jobs are co-located on CCSs, the unmanaged contention for computing resources usually leads to significant fluctuations in tail latency, which have a severe negative impact to the end user experience. Therefore, some promising computing systems and effective scheduling solutions are proposed to guarantee better quality of service (QoS), e.g., La-beled von Neumann Architecture (LvNA) based cloud prototype system. However, relatively few work focuses on the tail latency observation and evaluation of public CCSs and newest prototype systems. Therefore, we introduce an evaluation framework named TailCmp which includes nine representative workloads ranging over different tail latency requirements and application domains. With the TailCmp, we can collect and analyze four evaluation metrics including tail latency entropy. In this paper, we conduct a large number of experiments with TailCmp on three CCSs (two public CCSs and one LvNA-based prototype system). The results show that the co-location on existing public CCSs can seriously affect the QoS of LC workloads, while the labeling mechanism in Lv Na - based prototype can improve the performance of LC workloads in co-location.

External IDs:dblp:conf/ispa/KongGPZYZQ22