Workload-driven coordination between virtual machine allocation and task scheduling

摘要

The current task scheduling is separated from the virtual machine (VM) allocation, which, to some extent, wastes resources or degrades application performance. The scheduling algorithm influences the demand of VMs in terms of service-level agreement, while the number of VMs determines the performance of task scheduling. Workload plays an indispensable role in both dynamic VM allocation and task scheduling. To address this problem, we coordinate task scheduling and VM allocation based on workload characteristics. Workload is empirically time-varying and stochastic. We demonstrate that the acquired workload data set has Markov property which can be modeled as a Markov chain. Then, three workload characteristic operators are extracted: persistence, recurrence and entropy, which quantify the relative stability, burstiness, and unpredictability of the workload, respectively. Experiments indicate that the persistence and recurrence of workloads has a direct bearing on the average response time and resource utilization of the system. A nonlinear model between the load characteristic operators and the number of VMs is established. In order to test the performance of the collaborative framework, we design a scheduling algorithm based on genetic algorithm (GA), which takes the estimated number of VMs as input and the task completion time as the optimization target. Simulation experiments have been performed on the CloudSim platform, testifying that the estimated average absolute VMs error is only 2.6%. The GA-based task scheduling algorithm could improve resource utilization and reduce task completion time compared with the first come first serve and greedy algorithm. The proposed coordination mechanism in this paper has proved able to find the optimal match and reduce the resource cost by utilizing the interaction between VM allocation and task scheduling.