JSBench: Automatically Constructing Realistic JavaScript Benchmarks

Опубликовано 17 августа 2016, 21:55

While artificial benchmark suites such as SunSpider have become popular in recent years, recent work has shown that their behavior is unrepresentative of real JavaScript applications found in the wild. The goal of JSBench is to rectify this problem by using real-world JavaScript programs as benchmarks. As real web pages are highly interactive, this task involves reproducing a sequence of user-initiated events in a consistent way across subtly-incompatible browsers and JavaScript engines. This is accomplished by a source-level instrumentation to produce a record-replay system. This method has several advantages: ΓÇó Has sufficient fidelity to accurately recreate much of the dynamic characteristics of the original program. ΓÇó Is sufficiently flexible that a recording produced in one environment can be replayed in a different environment. ΓÇó Creates benchmarks which require no more expertise to use than the currently-existing industry-accepted benchmarks. This talk will cover the concepts, implementation, complications and results of this project.