Code-Centric Communication Graphs for Shared-Memory Multithreaded Programs

With more and more complex pieces of software using explicit multithreading, tools to extract the structure of such software are increasingly important. We present a novel tool that builds graphs describing how threads in sharedmemory parallel programs communicate. Compared to prior work, our communication graphs are code-centric: nodes represent units of code (e.g., functions) and edges represent inter-thread shared-memory communication via these units of code. Our approach is dynamic, using actual executions to build graphs, and exploits binary-code instrumentation to work for large, real-world applications. The graphs are useful for understanding software structure and computing program properties, such as the effect of nondeterministic thread-scheduling on the communication pattern.