Performance analysis of preemption-aware scheduling in multi-cluster grid environments

Abstract. Resource provisioning is one of the main challenges in re-source sharing environments such as multi-cluster Grids. In these envi-ronments each cluster serves requests from external (grid) users along with their own local users. The problem arises when there is not suffi-cient resources for local users (which have higher priority than grid users) to be served urgently. This problem could be solved by preempting re-sources from grid users and allocating them to the local users. However, resource preemption entails crucial side-effects such as decreasing re-source utilization and increasing grid users ’ response time. Considering preemption side-effects, the question is that how we can minimize the number of preemptions taking place in a resource sharing environment. In this paper, we propose a preemption-aware scheduling policy for a virtualized multi-cluster Grid in a way that the number of preemptions is minimized. The proposed policy is based on the stochastic analysis of routing in parallel non-observable queues which acts independent from the availability information of each cluster. Simulation results indicate that the proposed scheduling policy significantly decreases the number of virtual machine (VM) preemptions (up to 22.5%) which affects both resource utilization and average response time of grid requests.