Deadline constrained scheduling in hybrid clouds with Gaussian processes

Abstract

In hybrid clouds, deciding which workloads to outsource and at what time is far from trivial. The objective of this decision is to maximize the utilization of the internal data center and to minimize outsourcing. Neither all tasks' runtime nor their issue time are known in advance. However, a majority of tasks are always issued automatically during the day, e.g. common batch jobs. This work presents experimental results on different optimization strategies for cost-optimal dynamic scheduling in hybrid cloud environments. We estimate task execution times as random variables over day time from past observations using Heteroscedastic Gaussian Processes (HGP). HGP are suitable in particular for the presented scheduling problem because they not only provide an estimation of a task's mean runtime (as given by standard regression methods), but also the certainty of this estimation. We show that HGP provide an intuitive framework to model a variety of different distributions. The overall results are similar to optimization results with the unknown generating distribution.