Software performance engineering by models and measurement

点击翻译按钮获取中文摘要

Software performance engineering (SPE) has inadequate tools, resulting in serious long-standing and continuing problems in delivering software with adequate performance (response time or throughput capacity). Examples abound: Microsoft's disastrous Vista release is one. The literature on problems in web service system development documents many kinds of performance problems. Software performance is a difficult scientific problem with large economic value, Recently the proposer and colleagues proposed integrating - the prevalent approach using measurements, with - the considerable (but largely academic) recent progress in performance modeling. To achieve this vision, this proposal includes: 1. semi-automatic construction and maintenance of performance models (Pmodels), 2. exploitation of Pmodels to evaluate designs, to create efficient performance tests, for capacity planning, and to configure and operate efficient deployed systems. The proposal includes extraction of layered performance models by transformation of UML specifications of software, with standard performance annotations; estimation and tracking of the parameters of performance models, simplification of models to improve scalability of the analysis to very large systems in clouds; optimal deployment and configuration of large applications in clouds (in the sense of optimal cost or energy consumption, subject to performance constraints computed by a model), and improvements to LQNS, the tool used to solve layered performance models.