Improving the performance of batch schedulers using online job runtime classification

Job scheduling in high-performance computing platforms is a hard problem that involves uncertainties on both the job arrival process and their execution times. Users typically provide only loose upper bounds for times, which are not so useful heuristics based processing Previous studies focused applying regression techniques to obtain better time estimates, worked reasonably well improved metrics. However, these approaches require long period of training data. In this work, we propose simpler approach by classifying jobs as small or large prioritizing over ones. Indeed, most impacted queuing delays, but they represent light load incur burden other jobs. The classifier operates online learns using data collected previous weeks, facilitating its deployment enabling fast adaptation changes workload characteristics. We evaluate our four policies seven HPC platform traces. show that: first, incorporating such classification reduces average bounded slowdown all scenarios, second, considered improvements comparable ideal hypothetical situation where scheduler would know advance exact running