Using MPI Communication Patterns to Guide Source Code Transformations
ICCS '08 Proceedings of the 8th international conference on Computational Science, Part III
Dragon: a static and dynamic tool for OpenMP
WOMPAT'04 Proceedings of the 5th international conference on OpenMP Applications and Tools: shared Memory Parallel Programming with OpenMP
Trace-based performance analysis for the petascale simulation code FLASH
International Journal of High Performance Computing Applications
| Hi-index | 0.00 |
A high-level understanding of how an application executes and which performance characteristics it exhibits is essential in many areas of high performance computing, such as application optimization, hardware development, and system procurement. Tools are needed to help users in uncovering the application characteristics, but current approaches are unsuitable to help develop a structured understanding of program execution akin to flow charts. Profiling tools are efficient in terms of overheads but their way of recording performance data discards temporal information. Tracing preserves all the temporal information but distilling the essential high level structures, such as initialization and iteration phases can be challenging and cumbersome. We present a technique that extends an existing profiling tool to capture event flow graphs of MPI applications. Event flow graphs try to strike a balance between the abundance of data contained in full traces and the concise information profiling tools can deliver with low overheads. We describe our technique for efficiently gathering an event flow graph for each process of an MPI application and for combining these graphs into a single application-level flow graph. We explore ways to reduce the complexity of the graphs by collapsing nodes in a step-by-step fashion and present techniques to explore flow graphs interactively.