Research Of A Moldable Job Scheduling Model Based On Revenue Per Processor

Job scheduler is one of the most important components in clusters, which determines the performance of jobs(e.g., average turnaround time). Moldable jobs have attracted more and more concerns, because they can adjust the number of allocated processors according to the real-time state of the system, which can potentially reduce the waiting time or runtime of jobs(and thus turnaround time). Moldable job scheduling comprises two kinds of schemes: resource allocation scheme and job selection scheme. Job selection scheme has already been studied extensively, while existing resource allocation schemes do not perform well.To solve the above problem, the moldable job scheduling approach proposes an indicator named Revenue per Processor(RP), and design a resource allocation strategy based on this indicator, which integrates the Highest Revenue First resource allocation algorithm. RP is calculated by Downey model, which expresses the revenue on shortening runtime with every allocated processor. The allocation algorithm always allocates the processor to the job with highest RP, in order to reduce jobs' runtime and thus average turnaround time. After the allocation of resource, the allocation result is handled to the job selection scheme to instruct jobs' scheduling. The job selection scheme integrates FCFS and backfilling algorithms, which can be selected by the administrators of the cluster according to their requirements. Because the RP-based resource allocation approach allocates resources to the jobs with highest RP, the scheduler is adaptive to the system's load and hence improve its performance.In our experiments, we explore how parameters in our resource allocation algorithm affect the performance of scheduling, and give some suggestions on how to set them in practical scheduling. We also compare our method with other state-of-the-art method. Results show that the scheduling approach combining HRF with backfilling can always achieve the shortest average turnaround time among all approaches(up to 71% improvement).