Fast detection of communication patterns in distributed executions
Thomas Kunz, Michiel F. H. Seuren · 1997
Understanding distributed applications is a tedious and difficult task. Visualizations based on process-time diagrams are often used to obtain a better understanding of the execution of the application. The visualization tool we use is Poet, an event tracer developed at the University of Waterloo. However, these diagrams are often very complex and do not provide the user with the desired overview of the application. In our experience, such tools display repeated occurrences of non-trivial communication patterns, appearing throughout the trace data and cluttering the display space. This paper describes an event abstraction facility which tries to simplify the execution visualization shown by Poet by efficiently detecting and abstracting such patterns. A user can define patterns, subject to only very few constraints, and store them in a hierarchical pattern library. We also provide the user with the possibility to annotate the source code as a help in the abstraction process. We detect these communication patterns by employing an enhanced e cient multiple string matching algorithm. The results indicate that the matching process is indeed very fast. A user can experiment with multiple patterns at potentially different levels in the hierarchy, checking for their occurrence in the trace le, while trying to gain some understanding in a short period of time.