Towards workload-aware self-management: Predicting significant workload shifts

Abstract

The workloads of enterprise DBS often show periodic patterns, e.g. because there are mainly OLTP transactions during day-time and analysis operations at night. However, current DBS self-management functions do not consider these periodic patterns in their analysis. Instead, they either adapt the DBS configuration to an overall “average” workload, or they reactively try to adapt the DBS configuration after every periodic change as if the workload had never been observed before. In this paper we propose a periodicity detection approach, which allows the prediction of workload changes for DBS self-management functions. For this purpose, we first describe how recurring DBS workloads, i.e. workloads that are similar to workloads that have been observed in the past, can be identified. We then propose two different approaches for detecting periodic patterns in the history of recurring DBS workloads. Finally we show how this knowledge on periodic patterns can be used to predict workload changes, and how it can be adapted to changes in the periodic patterns over time.