SC21 Proceedings

The International Conference for High Performance Computing, Networking, Storage, and Analysis

Chaining Multiple Tools and Libraries Using Gotcha

Authors: Yiheng Xu (University of Maryland); Kathryn Mohror, Hariharan Devarajan, and Cameron Stanavige (Lawrence Livermore National Laboratory); and Abhinav Bhatele (University of Maryland)

Abstract: In HPC, it is common to use tools to understand the performance of applications. Users often want to run applications linked with multiple performance tools to complement each other, as tools may record different information. Many tools operate by intercepting application function calls with "wrapper functions" that perform performance measurement tasks. This approach, however, only works for using one tool at a time. Using multiple tools that intercept a common set of functions in the same run can lead to erroneous results.

In this work, we enable multiple tools to measure a single application execution together utilizing a library called Gotcha. With our approach, we can use multiple tools in the same run to reduce overheads of running the application repeatedly and to produce more temporally aligned profiles from each of the tools. We present our work in making two I/O performance tools, Recorder and Darshan, Gotcha enabled.

Best Poster Finalist (BP): no

Poster: PDF
Poster summary: PDF

Back to Poster Archive Listing