Abstract:

To reduce the data center’s (DC) operational costs and provide grid services, large computational jobs need to be scheduled or transferred to the intended DCs with lower energy prices. Usually, the transfer or scheduling of large computation jobs are managed by job scheduling software, such as a portable batch system. The paper proposes a novel rolling window mixed integer linear programming-based (MILP) approach that can be leveraged by the job scheduling system to determine a job’s location and start time to minimize cost and provide grid services. To address higher number of jobs in a stipulated time frame, the constraints arising from the job’s resource requirements, execution time, location of execution, and their dependencies on the previous state are formulated linearly. The optimization module considers the grid service request and newly arrived jobs along with jobs in the queue, which ensures the feasibility and helps the operator with dynamic decision-making of pausing or transferring jobs. The effectiveness of the proposed approach is shown on the subset of data from the ALIBABA GPU cluster.