Conference: 2023 14TH IEEE International Conference on Cloud Computing Technology and Science (CloudCom)
ABSTRACT
Serverless computing has reshaped the cloud com- puting landscape by offering benefits such as auto-scalability, streamlined operational management, and granular billing. As its adoption grows, challenges related to performance optimization in hybrid architectures combining private servers and public cloud clusters have emerged. Central to these challenges are achieving optimal response latency and balancing performance and cost. To address these challenges, this paper introduces an adaptive routing service specifically designed for hybrid environments, proficient in leveraging real-time function metrics. Our proposed service pivots on three integral components: a monitor that captures performance metrics and raises alarms for predefined anomalies; a forecaster that predicts function latency across clusters, which includes wait and execution times, and produces request distributions for each cluster to equalize the overall function latency; and a router then processes incoming requests, taking cues from the forecaster’s predictions. Notably, based on user-defined objectives, the forecaster can be directed to either minimize latency or optimize execution costs. Comprehen- sive evaluations on AWS and Azure using Apache OpenWhisk clusters showcase our approach’s effectiveness, yielding a 9% improvement in average latency, a 45% decrease in standard deviation latency and a 17% cost reduction compared to con- ventional 50-50 routing. The advantages of elevated monitoring frequency are also illuminated, emphasizing quicker convergence times.
Authors
- Georgios Fatouros
- George Kousiouris
- Georgios Makridis
- John Soldatos
- Michael Filippakis
- Dimosthenis Kyriazis
You may follow the PHYSICS project activities on Twitter and LinkedIn.
