Asocial: Adaptive Task Re-Allocation in Distributed Computing Systems with Node Failures
Weiwei Zheng, Yanli Shen, Xiao Taoshun · 2020
Task allocation-the problem of efficiently allocating a set of M tasks to a set of N nodes-is a fundamental issue in distributed computing systems (DCSs). The problem is particularly challenging in the presence of fail-prone nodes. In this work, we design an adaptive task (re-)allocation mechanism, named Asocial. The Asocial allows us to optimize the utilization of system resources (including computing resources and network resources) while increasing service-level fault tolerance as well as failure resilience. In addition, we propose a cooperative game model based task (re-)allocation approach for cooperation among Physical Nodes (PNs). We consider both the performance states and reliability levels of candidate PNs when deploying tasks. Specifically, we exploit failure prediction techniques to evaluate PNs' reliability levels. As a result, we can utilize the resources efficiently and thus improve the service reliability (i.e., the probability of serving all the tasks before their delivery time). We show by means of numerical evaluations that the proposed Asocial can significantly improve service reliability, system availability as well as resource utilization. In particular, by using failure prediction result (F-measure is around 0.8), the application completion rate (ACR), the task completion rate (TCR) and the computing resource utilization (CRU) reach 85.20%, 85.67% and 78.34%, respectively. Compared with 67.30%, 70.07% and 65.79% of the initial allocation scheme, the performance achieves a significant improvement (26.60%, 22.26% and 19.08%, respectively).