A model for tracing and debugging large-scale task-parallel programs with MPE

Application frameworks and domain-specific languages (DSLs) (both here called high-level tools) aid developers when developing programs for next-generation, highly concurrent systems. An increase in the use of high-level tools, however, creates a problem: the prevention and detection of defects in the high-level program. Traditional debuggers, designed for operating on highly popular, line-oriented languages (C, C++, Java), will operate at too low a level to detect defects with the use of the high-level tool. With regard to scale, while lineoriented tools are expected to remain viable on foreseeable systems, the ability of the human user to effectively use these tools at larger scale is questionable. Logging is a typical approach to monitoring and detecting problems in program execution. At present, logging faces similar challenges to traditional debuggers: utility and scale. In an application built in a high-level tool, logging must also be efficient and high-level. Low-level logging, say, at the messaging level, will be too complex for users of high-level tools. Thus, the tool builders must integrate effective, highlevel logging functionality. Here, we consider the use of the Message Passing Environment (MPE) [1] logging applied to Swift [2] programs. This case study presents a model for effectively detecting programming defects based on the MPE logs in a manner appropriate for Swift programs (not MPI programs in general).