A Scalable Approach to Distributed Scheduling Scheme on Cluster Architecture of Real-Time Server

摘要

For cluster servers demanded to sustain high workload,system scalability imposes a critical challenge to cluster design. Most of scheduling algorithms proposed for clusters are push-based,i.e. the resource scheduling and load distribution functions are implemented at the cluster frontend, which may lead to a performance bottleneck when system scales up to meet growing needs of computing power. In this paper,we propose a new Pull-Based Distributed Scheme (PBDS) to completely spread the scheduling functionality to cluster backend nodes,which ultimately reduces the workload at the cluster front-end. Experiments on Linux cluster at varied load levels demonstrate that the PBDS scheme outperforms conventional push-based schemes by cutting the average response time by 27% to 32% while maintaining overall throughput level. Furthermore,scalability test shows the PBDS scheme presents a consistent better performance when system is expanded from 10 nodes to 30 nodes,showing a promising scalable solution to large-scale cluster systems.