On the Utility of Machine Learning for Service Capacity Management of Enterprise Applications

Abstract

The IT industry drives the digital transformation and, at the same time, is affected itself by related trends of automation and computerization. This paper examines the applicability of machine-learning techniques to the process of service capacity management for Commercial-off-the-shelf enterprise applications. We use real monitoring data from more than 18.000 SAP application and database instances which are running on more than 16.000 different servers in order to train performance models for standard business functions. A learning algorithm which is based on Boosted trees achieves sufficient accuracy to predict mean response times for ten frequently used transactions. To evaluate utility, models are successfully applied as part of a scenario-based demonstration in the fields of server sizing, load testing, and server consolidation with the objective to identify cost-effective designs. Results strongly emphasize the need to integrate monitoring data from uniform business applications in order to allow for novel and cost-effective capacity management service offerings.