Multi-dimensional server consolidation for commercial off-the-shelf enterprise applications using shared performance counters

Abstract

Internal and external IT service providers increasingly use commercial-off-the-shelf software to support business processes. As these applications are continuously monitored, emerging log data follows a standardized format and, in its internal logic, is comparable across organizational boundaries. Resulting automation potential leads to cost savings and quality improvements in the context of capacity management. These objectives are examined by the thesis at hand using the capacity management method PPSS (Performance prediction supported service placement), which is designed to address the automation potential for server consolidation scenarios. For this purpose, a placement problem is formulated that meets the special requirements of enterprise applications. The objective of the problem is to save operations costs by minimizing the required capacity. Solution quality is further a ected by the compliance with a variety of constraints. Four heuristics, two metaheuristics, and two hybrid algorithms are evaluated in 12,384 field experiments with respect to their solution quality. As the number of constraints increases, genetic algorithms tend to identify solutions of the highest relative quality. In all scenarios, more than 20% of the original server capacity can be saved on average while complying with the given constraints. Too aggressive capacity reduction, however, increases the risk of violating performance-related service level agreements and entails the payment of penalty costs. The prediction of transactional response times that are expected from a solution candidate enables to estimate the amount of such penalty costs. At the same time, solution credibility is increased. For this purpose, PPSS integrates the use of black-box approaches, which are based on machine learning and keep personnel costs low. In addition, the tested techniques benefit from the cross-organizational integration of log data due to large volume and variety of the observations on which the learning process is based. Using the example of a widespread standard transaction, both Random forests and Boosted trees prove to be suitable methods for predicting the mean response times of dialog steps. Boosted trees show mean absolute percent errors between 19% and 30% across additional test cases on frequently used business transactions. A case study demonstrates the utility of the method. Here, alternative solution candidates of the placement problem are analyzed with regard to their total costs. These consist of operations costs and penalty costs. Using the predicted response times and an exemplary service level agreement, the amount of expected penalty costs can be estimated for alternative load scenarios. If future load probabilities are known, a single solution can be recommended to minimize the total costs. PPSS is technically enabled by a performance knowledge base consisting of three layers which cover presentation, analysis, and data. The knowledge base is implemented to evaluate the research artifact in a real environment. Selected process steps are supported on the presentation level by a graphical user interface. This interface is offered to a group of test users from two different data centers as part of a pilot operations phase. The user feedback proves the utility and indicates cost savings of the method when compared to existing approaches.