Experimental analysis and bug abstraction for distributed computation on ray framework

Abstract

This research aims to address challenges in distributed computing, focusing on the ray framework, which has potential for efficient parallel and distributed task execution. While methods such as model-checkers and fuzzing have been applied to detect bugs, both have limitations in handling the complexity of distributed computing, particularly in dealing with issues like state-space explosion and identifying rare bugs. This study proposes an alternative approach through experimental analysis and bug abstraction methods to discover, identify, and classify bugs in the ray framework. Experimental analysis involves isolating and re-testing bugs in a controlled environment to understand their characteristics, while bug abstraction analyzes the factors causing bugs to identify common patterns and characteristics. The results of this research successfully identified three main categories of bugs: crash, performance, and inaccurate status, and revealed bug characteristics that do not depend on actor instance multiplicity, actor type, specific event sequences, or particular configurations. This research makes a significant contribution to the development of more effective and efficient bug detection methods in distributed computing, particularly in the ray framework, and paves the way for further research to enhance the reliability of distributed systems.