ALCF Developer Session November 2017

Argonne, Illinois
Wednesday, November 29, 2017
The ALCF Many-Core Developer Sessions are videoconferences aimed at increasing the dialogue between ALCF users and developers of many-core systems and software.  Attendees are encouraged to bring any questions they may have related to ALCF's Theta system and the Intel Xeon Phi technology in general.


ALCF Developer Session November 2017
Wednesday, November 29, 2017 11:00 AM - 12:00 PM (US Central)

ALCF Videoconference
Argonne, Illinois 60539
United States


The TAU Performance System

Sameer Shende, University of Oregon

Abstract

The complex nature of HPC platforms and their application development environment - combining multiple languages, programming paradigms, hardware, and compilers - make effective performance engineering a challenging task. To meet the needs of computational scientists in tuning their codes, we present the TAU Performance System. TAU is a powerful profiling and tracing toolkit that covers multiple aspects of performance instrumentation, measurement, and analysis. After describing and demonstrating how performance data is collected using TAU’s automated instrumentation, we will present ways to analyze the performance data collected and how to drill down to find performance bottlenecks. TAU can internally use PAPI and LIKWID tools to generate valuable data from hardware performance counters on Intel® Xeon Phi™ (KNL) cpus in Cray systems such as Theta at ALCF that can help developers gain insight into the inner workings of the processor and memory. Performance data generated includes memory bandwidth to the HBM, data cache misses, and instructions executed. We will demonstrate the use of tau_exec, a lightweight profiling tool that can launch un-instrumented binaries to generate profiles and traces on the Theta system and collect data from the MPI, OpenMP, pthread runtimes and use event-based sampling (EBS) to pinpoint performance bottlenecks at the statement level in hybrid multi-threaded codes.  TAU may also be used for instrumenting applications that use the Kokkos library. The webinar will include a demo of tau_exec and paraprof, TAU’s 3D profile browser, running on ALCF systems.

About the Speaker

Sameer Shende is the Director of the Performance Research Lab at the University of Oregon.  He is also the President and Director of ParaTools, Inc, which specializes in performance evaluation, algorithms, and software development for parallel and distributed computing.

About the Series

The ALCF Many-Core Developer Sessions are videoconferences aimed at increasing the dialogue between ALCF users and developers of many-core systems and software. Attendees are encouraged to bring any questions they may have related to ALCF's Theta system and the Intel Xeon Phi technology in general.

 

Contact Information