Perfdiff: a framework for performance difference analysis in a virtual machine environment

• Authors: Xiaotong Zhuang;Suhyun Kim;Mauri io Serrano;Jong-Deok Choi
• Venue: Proceedings of the 6th annual IEEE/ACM international symposium on Code generation and optimization
• Year: 2008

Although applications running on virtual machines, such as Java, can achieve platform independence, performance evaluation and analysis becomes difficult due to extra intermediate layers and the dynamic nature of virtual execution environment. We present a framework for analyzing performance across multiple runs of a program, possibly in dramatically different execution environments. Our framework is based upon our prior lightweight instrumentation technique for building a calling context tree (CCT) of methods at runtime. We first represent each run of a program by a CCT, annotating its edges and nodes with various performance attributes such as call counts or elapsed times. We then identify components of the CCTs that are topologically identical but with significant performance-attribute differences. Next, the topological differences of two CCTs are identified, while ignoring the performance attributes. Finally, we identify the differences in both topology and performance attributes that can be fed back to the software developers or performance analyzers for further scrutiny. We have applied our methodology to a number of well-known Java benchmarks and a large J2EE application, using call counters as the performance attribute. Our results indicate that this approach can efficiently and effectively relate differences to a small percentage of nodes on the CCT. We present an iterative framework for program analysis, where topological changes are performed to identify differences in CCTs. For most of the test programs, applying a few topological changes such as deletion, addition, and renaming of nodes - are needed to make any two CCTs from the same program identical, whereas less than 2% of performance-attribute changes are needed to achieve a 90% overlap of any two CCTs in performance attributes, after the two CCTs are topologically matched. We have applied our framework to identify subtle configuration differences for complex server applications.