Parallel Performance Evaluation Using TAU

 

To meet the needs of computational scientists to evaluate the performance of their parallel, scientific applications, we present the TAU performance system and its interfaces with other tools such as  PAPI, Perfetto, OTF2, and Vampir. This two-day workshop will cover performance evaluation of applications on tri-lab OCF platforms, and consulting sessions. This workshop will focus on performance data collection, analysis, and performance optimization. After describing and demonstrating how performance data (both profile and trace data) can be collected in a straightforward manner using TAU’s (Tuning and Analysis Utilities) automated instrumentation, the bulk of the workshop will cover how to analyze the performance data collected and drill down to find performance bottlenecks and determine their causes.

 

The workshop will include some sample codes that illustrate the different instrumentation and measurement choices available to the users. Topics will cover generating performance profiles and traces with memory utilization and headroom, I/O, and hardware performance counters data using PAPI. The workshop will demonstrate scalable tracing using OTF2 and visualization using the Vampir trace analysis tool. Performance data analysis using ParaProf and PerfExplorer will be demonstrated using the performance data management framework (TAUdb) that includes TAU’s performance database.

 

The workshop will also feature cross experiment analysis including comparing the effects of multi-core architectures on code performance. The demonstrations will include using TAU on programming paradigms such as ROCm, Intel oneAPI (DPC++/SYCL), OpenCL, OpenACC, CUDA,.

 

The workshop will also cover using TAU in the Extreme-Scale Scientific Software Stack (E4S) [https://e4s.io] using containers and AWS. E4S is a curated Spack based collection of HPC and Ai/ML tools and includes PETSc, Trilinos, TAU, HPCToolkit, HDF5, as well as TensorFlow, PyTorch and other Generative AI toolkits available on commercial  cloud platforms.

 

We will attempt to collect and analyze performance data for additional user codes during the hands-on portion of the workshop. Users and developers are welcome to contact the instructor ahead of time to begin collecting data to discuss at the workshop.

The second day of the workshop is dedicated to one-on-one consultation sessions with Dr. Shende for further, more in-depth, instruction and help in addressing performance bottlenecks in your codes. Please see below for information on scheduling an appointment.