Adaptive Job Scheduling Via Predictive Job Resource Allocation

摘要

Standard job scheduling uses static job sizes which lacks flexibility regarding changing load in the system and fragmentation handling. Adaptive resource allocation is known to provide the flexibility needed to obtain better response times under such conditions. We present a scheduling approach (SCOJO-P) which decides resource allocation, I.e. the number of processors, at job start time and then keeps the allocation fixed throughout the execution (I.e. molds the jobs). SCOJO-P uses a heuristic to predict the average load on the system over the runtime of a job and then uses that information to determine the number of processors to allocate to the job. When determining how many processors to allocate to a job, our algorithm attempts to balance the interests of the job with the interests of jobs that are currently waiting in the system and jobs that are expected to arrive in the near future. We compare our approach with traditional fixed-size scheduling and with the Cirne-Berman approach which decides job sizes at job submission time by simulating the scheduling of the jobs currently running or waiting. Our results show that SCOJO-P improves mean response times by approximately 70% vs. traditional fixed-size scheduling while the Cirne-Berman approach only improves it 30% (which means SCOJO-P improves mean response time by 59% vs. Cirne-Berman).