Date of this Version
Executions of modern parallel programs often yield complex communications among compute nodes of large-scale clusters of workstations or supercomputers. Analyzing communication patterns is becoming increasingly critical to performance optimiza- tion. As the scale and complexity of parallel applications drastically increases, visu- alization has become a feasible means to conduct analysis of massive communication patterns. However, most visualization tools fall short in showing comprehensive dy- namic communication graph and addressing the scalability issue. Our solution for analyzing dynamic communication patterns is based on an analytics framework cou- pled with a new visualization technique, named CommGram , that provides a flexible solution to the scalability issue. We can explore large communication data at different levels of detail, and detect potential communication bottlenecks of massive parallel programs. The conclusion of our studies is based on large-scale scientific ap- plications that include end-to-end simulation pipelines and AMR-based simulations.
Adviser: Hongfeng Yu