Introduction to GASPI - Scalability in HPC

Date:

Tuesday, May 8, 2018 09:00 - Wednesday, May 9, 2018 16:30
Location: LRZ Building, University campus Garching, near Munich, Kursraum 2 H.U.010
Contents:

Overview

GASPI (Global Address Space Programming Interface) offers a small, yet powerful API for scalable HPC applications following a single program multiple data (SPMD) approach. Single-sided communication primitives are supplemented by a light weight synchronization mechanism (notifications). As such, GASPI aims at asynchronous data flows with remote completion, rather than bulk-synchronous message exchanges. Perfect overlap of communication by computation paired with minimal synchronization yields strong scalability.The strong scalability of GASPI has been shown by several use cases, such as seismic or CFD applications. See http://www.gaspi.de and http://www.gpi-site.com/ for more details.

The workshop aims to give an introduction to the GASPI programming model by dealing with following key aspects:
Day 1: What does GASPI offer?
Day 2: How can I use it and which typical use-cases do we have?

Agenda

Tuesday:

The GASPI API

  • Introduction to GASPI
  • Execution model
  • Segments
  • One-sided communication
  • Collectives, Passive communication, Fault tolerance


Wednesday:

The GASPI programming model

  • Interoperability with MPI
  • Efficient parallel programming:
    - from bulk synchronous to fully asynchronous
    - execution
  • Overlap communication/computation
  • Data dependency driven execution

On both days the acquired knowledge will be applied in hands-on sessions using the CoolMUC-2 system.
Prerequisites Good working knowledge of at least one of standard HPC languages, such as C or C++. Basic knowledge in parallel programming is useful.
Language:

English
Teachers:

Daniel Grünewald, Martin Kühn (Fraunhofer Institute for Industrial Mathematics ITWM, Kaiserslautern)
Registration:

Via the LRZ registration form.
Contact: Dr. Andre Kurzmann