Load Index Metrics for an Optimized Management of Web Services: a Systematic Evaluation

Abstract

The lack of precision to predict service performance through load indices may lead to wrong decisions regarding the use of web services, compromising the service performance and raising the platform cost unnecessarily. This paper presents experimental studies to qualify the behaviour of load indices in the web service context. The experiments consider three services that generate controlled and significant demands on the server, four levels of workload for each service and six distinct execution scenarios. The evaluation considers three disassociated perspectives: capability for representing recent workloads, capability for predicting a near-future performance and finally, stability. Eight different load indices were analysed and one of them, the JMX Average Time, is proposed in this paper. A systematic approach is applied to evaluate the different load indices, considering a multiple linear regression model based on the stepwise-AIC method. The results show that the load indices studied represent the workload to some extent; however, differently from what is expected, most of them do not present coherent correlation to the service performance and can result in problems with stability. JMX Average Time is an exception, showing a stable behaviour, tightly-coupled to the service runtime for all executions. Load indices are used to predict the service runtime and show that their mistaken use can lead to decisions that will impact negatively on both service performance and execution cost. (AU)