Grid-job scheduling with reservations and preemption

Computational grids make it possible to exploit grid resources across multiple clusters when grid jobs are deconstructed into tasks and allocated across clusters. Grid-job tasks are often scheduled in the form of workflows which require synchronization, and advance reservation makes it easy to guarantee predictable resource provisioning for these jobs. However, advance reservation for grid jobs creates roadblocks and fragmentation which adversely affects the system utilization and response times for local jobs. We provide a solution which incorporates relaxed reservations and uses a modified version of the standard grid-scheduling algorithm, HEFT, to obtain flexibility in placing reservations for workflow grid jobs. Furthermore, we deploy the relaxed reservation with modified HEFT as an extension of the preemption based job scheduling framework, SCOJO-PECT job scheduler. In SCOJO-PECT, relaxed reservations serve the additional purpose of permitting scheduler optimizations which shift the overall schedule forward. Furthermore, a propagation heuristics algorithm is used to alleviate the workflow job makespan extension caused by the slack of relaxed reservation. Our solution aims at decreasing the fragmentation caused by grid jobs, so that local jobs and system utilization are not compromised, and at the same time grid jobs also have reasonable response times.