Is the programming environment ready for hybrid supercomputers?

 Alistair Hart, Harvey Richardson (Cray UK Ltd.)

Contact: parco-hybridms [at] cray [dot] com (parco-hybridms at cray.com)

Hybrid supercomputers are increasingly prevalent in the Top500 list, including large GPU-accelerated machines in Switzerland and the US. Future high end systems will be much more parallel with heterogeneous processing elements (e.g. GPU or MIC accelerators), more processors per node, more threads per processor, longer vector lengths and more complex memory hierarchies. A good programming environment (PE) is therefore essential to allow application developers to maximise programmability, ease porting, correctness-checking and tuning efforts and maintain performance-portability across a wide range of systems.
 
In this mini-symposium we will discuss the state-of-the-art in PEs for hybrid supercomputers from the perspectives of application specialists, tools developers and vendors, with a view to constructing a list of recommendations for future PE components and integration.
 
Specific areas of interest are:

  • Use of high level programming models, especially directive-based ones like OpenACC and OpenMP4accel
  • "Message passing" techniques (MPI or PGAS) in environments with diverse memory spaces
  • Development, performance analysis and debugging tools

 
A provisional list of invited presentations includes presentations
from:

  • Intel
  • NVIDIA
  • CSCS, Switzerland
  • Cray Inc.

A panel session will allow all participants to contribute to a gap analysis and recommendations for hybrid programming environments.

 

Program:

Thurdsay 3rd September

1000-1010 Introduction
1010-1100 Stefan Andersson, John Levesque and Aaron Vose (Cray Inc.) - A Strategy for Devleoping a Performance Portable Highly Scalable Application
1130-1200 Alan Gray & Kevin Stratford - Performance Portability across Traditional and Hybrid Supercomputers
1200-1230 Aidan B. G. Chalk, Pedro Gonnet, Matthieu Schalle - Towards Task-Based Parallelism on Heterogeneous Architectures using Quick Sched


1400-1450 Timothy Lanfear (NVIDIA) - The Tesla Accelerated Computing Platform
1450-1510 Rayman Reyes - SYCL: Single-source C++ accelerator programming
1510-1530 Christos Margiolas - Hexe: A Heterogeneous Execution Engine for LLVM

Friday 4th September

1000-1100 Will Sawyer (CSCS) - Hybrid Programing Case Studies at CSCS
1130-1220 Andrew Mallinson (Intel) - Preparing for Knights Landing
1220-1245 Discussion Session
1245-1300 Wrapup


 
Call for Participation
 
Authors are invited to submit abstracts for the symposium which will be peer reviewed. Abstracts should be between 1-2 pages, in English and PDF format. A short abstract (up to 250 words) will be required for the conference agenda soon after acceptance if not supplied along with the submission. Authors with accepted submissions should register for the ParCo conference and will have the option to have a paper published in
the ParCo proceedings.
 
Submissions should be sent to parco-hybridms [at] cray [dot] com (parco-hybridms at cray.com)
 
Important Dates:

Submissions are now closed and the mini-symposium agenda is now being finalized.
 

Last updated: 17 Aug 2015 at 12:04