Search

Server virtualization and consolidation techniques have been widely adapted in the modern large-scale computing systems to reduce energy consumption and increase resource utilization. In these systems, physical servers are turned on/off dynamically according to the workload variation, and the loading from computing tasks are balanced among active servers through virtual machine (VM) migration. However, the downside of this approach is the overhead of VM migration can cause several negative impacts to the system and users, including application performance degradation, service interruption, prolonged job execution time, extra network bandwidth consumption, and risk of failure, etc. The existing works in the literature attempt to reduce VM migration cost for persistent running web servers in a reactive manner. In contrast, we tackle the problem for parallel computing jobs of batch processing systems. Our approach can proactively avoid VM migrations with the co-design of between job scheduling and VM consolidation strategies, and minimize communication overhead of jobs by considering the traffic pattern between the tasks of a job. Our evaluations have used real parallel job workload trace and a synthetically generated workload to show that our approach can notably reduce the number of VM migrations by 35%–50% and communication cost by up to 25% compared to the traditional job scheduling approaches. [ABSTRACT FROM AUTHOR]