Details

Autor(en) / Beteiligte

• de Luna, Filipe Robalo

Titel

Oscar. a Noise Injection Framework for Testing Concurrent Software

Ort / Verlag

ProQuest Dissertations & Theses

Erscheinungsjahr

2022

Link zum Volltext

• ProQuest

Quelle

ProQuest Dissertations & Theses A&I

Beschreibungen/Notizen

• “Moore’s Law” is a well-known observable phenomenon in computer science that describes a visible yearly pattern in processor’s die increase. Even though it has held true for the last 57 years, thermal limitations on how much a processor’s core frequencies can be increased, have led to physical limitations to their performance scaling. The industry has since then shifted towards multicore architectures, which offer much better and scalable performance, while in turn forcing programmers to adopt the concurrent programming paradigm when designing new software, if they wish to make use of this added performance. The use of this paradigm comes with the unfortunate downside of the sudden appearance of a plethora of additional errors in their programs, stemming directly from their (poor) use of concurrency techniques.Furthermore, these concurrent programs themselves are notoriously hard to design and to verify their correctness, with researchers continuously developing new, more effective and effi- cient methods of doing so. Noise injection, the theme of this dissertation, is one such method. It relies on the “probe effect” — the observable shift in the behaviour of concurrent programs upon the introduction of noise into their routines. The abandonment of ConTest, a popular proprietary and closed-source noise injection framework, for testing concurrent software written using the Java programming language, has left a void in the availability of noise injection frameworks for this programming language.To mitigate this void, this dissertation proposes OSCAR — a novel open-source noise injection framework for the Java programming language, relying on static bytecode instrumentation for injecting noise. OSCAR will provide a free and well-documented noise injection tool for research, pedagogical and industry usage. Additionally, we propose a novel taxonomy for categorizing new and existing noise injection heuristics, together with a new method for generating and analysing concurrent software traces, based on string comparison metrics.After noising programs from the IBM Concurrent Benchmark with different heuristics, we observed that OSCAR is highly effective in increasing the coverage of the interleaving space, and that the different heuristics provide diverse trade-offs on the cost and benefit (time/coverage) of the noise injection process.