Data-driven software performance engineering : models and estimation algorithms

• Main Author: Wang, Weikun
• Other Authors: Casale, Giuliano
• Published: Imperial College London 2016
• Subjects: 004
• Online Access: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.754671

The accurate performance measurement of computer applications is critical for service providers. For these providers, to ensure that the performance constraints signed by users can be respected, software performance models are used to provide quantitative predictions and evaluation of the applications such that timely adjustment of the architecture and fine tuning of the configurations can be achieved. For effective use of performance models in performance engineering, the fundamental problem is to assign realistic parameters to the models. In addition, the complexity of real-world application leads to dynamic behavior, caused by parallel computations with multiple CPUs or caching and shared data structures, which is challenging to capture for performance models to capture. Characterizing these changing effects, also known as load-dependent or queue-dependent application behaviors, is necessary for accurate prediction and evaluation of application performance. To enhance the model tractability and applicability, in this thesis we develop efficient algorithms to estimate the parameters of closed queueing networks, especially the resource demand of requests. To efficiently estimate these parameters, we introduce two classes of algorithms based on Markov Chain Monte Carlo (MCMC) algorithms and Maximum Likelihood Estimation (MLE) techniques. In particular, the MLE based approach can be generalized to load-dependent queueing networks, enhancing the applicability of the models. Moreover, we set out to resolve the problem of efficiently evaluating load-dependent and queue-dependent closed queueing network models. The complication of load-dependent or queue-dependent behavior makes the models challenging to analyze, a fact that discourages practitioners from characterizing workload dependencies. To solve this problem, we develop an algorithm for evaluating the performance of queue-dependent product-form closed queueing network. Our approach is based on approximate mean value analysis (AMVA) and is shown to be efficient, robust and easy to apply, thus enhancing the tractability of queue-dependent models.