Software performance engineering using virtual time program execution

• Main Author: Baltas, Nikolaos
• Other Authors: Field, Anthony ; Kelly, Paul
• Published: Imperial College London 2013
• Subjects: 004
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.587357

In this thesis we introduce a novel approach to software performance engineering that is based on the execution of code in virtual time. Virtual time execution models the timing-behaviour of unmodified applications by scaling observed method times or replacing them with results acquired from performance model simulation. This facilitates the investigation of "what-if" performance predictions of applications comprising an arbitrary combination of real code and performance models. The ability to analyse code and models in a single framework enables performance testing throughout the software lifecycle, without the need to to extract performance models from code. This is accomplished by forcing thread scheduling decisions to take into account the hypothetical time-scaling or model-based performance specifications of each method. The virtual time execution of I/O operations or multicore targets is also investigated. We explore these ideas using a Virtual EXecution (VEX) framework, which provides performance predictions for multi-threaded applications. The language-independent VEX core is driven by an instrumentation layer that notifies it of thread state changes and method profiling events; it is then up to VEX to control the progress of application threads in virtual time on top of the operating system scheduler. We also describe a Java Instrumentation Environment (JINE), demonstrating the challenges involved in virtual time execution at the JVM level. We evaluate the VEX/JINE tools by executing client-side Java benchmarks in virtual time and identifying the causes of deviations from observed real times. Our results show that VEX and JINE transparently provide predictions for the response time of unmodified applications with typically good accuracy (within 5-10%) and low simulation overheads (25-50% additional time). We conclude this thesis with a case study that shows how models and code can be integrated, thus illustrating our vision on how virtual time execution can support performance testing throughout the software lifecycle.