Scalable Message Tracing and Profiling Tool for Parallel Applications

To efficiently use leadership-class supercomputers with thousands of processors, such as the Department of Energy's BlueGene/L or Seaborg, developers of parallel applications need to understand the communication behavior in their applications. Therefore, tools are required to measure communication in parallel applications under realistic conditions. Existing tools either do not scale to these large numbers of processors or create an prohibitive amount of data. This project will develop a scalable tool to trace and visualize messages sent between processors. Particular emphasis will be on the scalability and applicability of long running applications on large numbers of processors. Phase I will extend an existing profiling and message tracking tool with capabilities to dynamically steer the collection of measurements. A scalable communication infrastructure will be implemented to transfer these measurements from the processors to the disk. Finally, methods for visualizing these measurements will be investigated. Commercial Applications And Other Benefits as described by the Applicant: The technology should increase the efficiency with which the Department of Energy's supercomputing infrastructure is used. It should be of interest to all developers of parallel applications. While the main focus is on largest parallel computers, the tool also would be useful for smaller scale clusters, which are becoming increasingly popular in private industry.