Capacity Management as a Service for Enterprise Standard Software

Abstract

Abstract. Capacity management approaches optimize component utilization from a strong technical perspective. In fact, the quality of involved services is considered implicitly by linking it to resource capacity values. This practice hinders to evaluate design alternatives with respect to given service levels that are expressed in user-centric metrics such as the mean response time for a business transaction. We argue that utilized historical workload traces often contain a variety of performance-related information that allows for the integration of performance prediction techniques through machine learning. Since enterprise applications excessively make use of standard software that is shipped by large software vendors to a wide range of customers, standardized prediction models can be trained and provisioned as part of a capacity management service which we propose in this article. Therefore, we integrate knowledge discovery activities into well-known capacity planning steps, which we adapt to the special characteristics of enterprise applications. Using a realworld example, we demonstrate how prediction models that were trained on a large scale of monitoring data enable cost-efficient measurement-based prediction techniques to be used in early design and redesign phases of planned or running applications. Finally, based on the trained model, we demonstrate how to simulate and analyze future workload scenarios. Using a Pareto approach, we were able to identify cost-effective design alternatives for an enterprise application whose capacity is being managed.

Addressing IT Capacity Management Concerns Using Machine Learning Techniques

“…Therefore, we argue for a domainspecific knowledge base to integrate measurement data of different environments that utilize the same type of COTS software. As investigated in our preliminary studies, these data may be used to train standardized performance models for the application in various capacity management scenarios [48,51]. Hence, initial model training and evaluation surely involves expertise from the data science domain.…”

“…To improve utilization levels, services with orthogonal workloads are to be identified. Subsequent relocations, however, are subject to uncertainties since the effect on the service performance is rarely analyzed before solution deployment [51]. Here, machine-learning techniques which were trained on a large number of different hardware configurations may support decisions with respect to the expected quality of service performance.…”