Application porting and optimization on GPU-accelerated POWER architectures

chaired by Dirk Pleiter (Forschungszentrum Juelich & Universitaet Regensburg)
Monday, 13 November 2017 from to (US/Mountain)
Description

SC17 Tutorial: Application porting and optimization on GPU-accelerated POWER architectures

The POWER processor has re-emerged as a technology for supercomputer architectures. One major reason is the tight integration of processor and GPU accelerator through the new NVLink technology. Two major sites in the US, ORNL and LLNL, have already decided to have their pre-exascale systems being based on this new architecture. This tutorial will give an opportunity to obtain in-depth knowledge and experience with GPU-accelerated POWER nodes. It focuses on porting applications to a single node and covers the topics architecture, compilers, performance analysis and tuning, and multi-GPU programming. The tutorial will include an overview of the new NVLink-based node architectures, lectures on first-hand experience in porting to this architecture, and exercises using tools to focus on performance.

Lecturers:

  • Christoph Hagleitner (IBM Research)
  • Andreas Herten (Jülich Supercomputing Centre)
  • Jiri Kraus (NVIDIA)
  • Bronson Messer (Oak Ridge National Laboratory)
  • Dirk Pleiter (Jülich Supercomputing Centre)
  • Archana Ravindar (IBM)
Support Email: a.herten@fz-juelich.de
Go to day
  • Monday, 13 November 2017
    • 08:30 - 09:15 Minsky Platform (1)
      • 08:30 IBM POWER8 processor, NVIDIA P100 GPU and IBM Minsky node hardware architecture (lecture) 45'
        Material: Slides pdf file
    • 09:15 - 10:00 POWER8 Processor (1)
    • 10:00 - 10:30 Coffee Break
    • 10:30 - 12:00 POWER8 Processor (2)
      • 10:30 Strategies to Improve Application Performance on POWER8 (lecture) 30'
        Material: Slides pdf file
      • 11:00 POWER8 performance optimization (hands-on) 1h0'
        Material: Slides pdf file
    • 12:00 - 13:30 Lunch Break
    • 13:30 - 16:00 Pascal GPU
      • 13:30 P100 GPU programming and optimization (lecture) 30'
        Material: Slides pdf file
      • 14:00 Multi-GPU Programming (lecture) 30'
        Material: Slides pdf file
      • 14:30 P100 performance optimization and multi-GPU Programming (hands-on) 1h30'
        Material: Slides pdf file
      • 15:00 Coffee Break 30'
    • 16:00 - 17:00 Minsky Platform (2)
      • 16:00 Best practices for porting scientific applications (lecture) 1h0'
        Material: Slides pdf file