DOE Centers of Excellence Performance Portability Meeting

April 19–21, 2016 - Glendale, Arizona

 

Post-meeting report now available

Overview

The Department of Energy (DOE) Centers of Excellence (COEs) Performance Portability meeting is an opportunity for the five COEs to share ideas, progress, and challenges toward the goal of performance portability across DOE's large upcoming advanced architecture supercomputer procurements. The need for applications to run effectively on multiple vendor advanced architecture solutions (as well as on standard "cluster" technology) is pervasive across application teams within DOE and is a specified goal of the DOE's exascale plans for risk mitigation. The two primary goals of this meetings are to:

  • Inform application teams and tool developers of activities and methodologies being used across the COEs, and foster informal relationships that can help DOE participants benefit from activities beyond their own COE.
  • Identify major challenges toward the goal of performance portability, and work with the vendors and tool providers on determining implementations and solutions that will meet their own performance criteria without inadvertently impairing performance results elsewhere.

read more...

 

Talks

Tuesday, April 19, 2016

Overviews

Applications/Optimizations/Algorithms

Performance Portable Abstractions

Wednesday, April 20, 2016

Managing the Memory Hierarchy

Application Experience with Performance Portable Abstractions

Experience with OpenMP and Recommendations on Guilding Future Standards

Thursday, April 21, 2016

Tools for Performance Portability and Analysis

The Input/Output Bottleneck and Use of Burst Buffers

Use of Domain-Specific Languages for Performance Portability

Breakout Sessions

During the meeting, several breakout sessions were held to gather participant input on four topics areas outlined below. Each topic had two groups independently discuss a set of questions that the breakout moderators worked together to come up with before the meeting. Below are the summaries of the discussions presented during the meeting as out-briefs:

Agenda

Download the final agenda (updated 4/14/16).\

Call for Abstracts/Talks (CLOSED)

Meeting organizers put out an open call for speakers to give short talks on progress, ideas, and/or challenges in the following topical areas:

  • Algorithmic and application work aimed at addressing trends in advanced architectures (e.g. reduced data motion, increased concurrency, use of burst buffers, etc.)
  • Software tools, libraries, abstractions, and standards intended to help applications address performance portability
  • Early application experiences (both positive and negative) attempting to run portably across diverse platforms
  • Portable approaches application and library teams are taking to "manage the memory hierarchy", and optimize data placement and movement

Speakers were chosen from the submissions that were received by Feb 29, 2016, and will be reflected in the final agenda.

Speaker and Participant Guidance

The following "ground rules" have been established for participants and speakers:

  • Talks and discussions must refrain from discussing information held under non-disclosure agreements. Contact your steering committee representative (below) if you need specific guidance.
  • In the spirit of the meeting, talks and discussions should address general challenges to the goal of performance portability and approaches that might be applied to overcome those challenges, rather than identifying and comparing state-of-play at a particular point in time.
  • Talks and discussions should not compare performance across specific platforms. Talks and discussions can address performance improvements on a given platform due to programming approaches or can address performance achieved relative to a theoretical performance model.
  • The focus of talks and discussions should be on portable, non-vendor-specific solutions as seen from the application developer perspective (that is, abstractions that hide vendor-specific solutions are acceptable). It is expected that a particular focus of the meeting will be to address possible evolutions of current standards (for example, OpenMP and C++) to better support performance portability.
  • Projections to future machines should not be presented.
  • Talks and discussions must be unclassified and non-sensitive in nature.
  • Speakers and participants (both labs and vendors) should accept that DOE will have multiple target platforms as part of their national strategy and join the discussion in the spirit of cooperation. All COEs are working toward the goal of making these platforms the most useful and high performance they can be without the threat of "vendor lock-in."

Steering Committee

james.r.reinders [at] intel.com (James Reinders,) Intel/Trinity-Cori
mwglass [at] sandia.gov (Mike Glass), SNL/Trinity
rjhartmanbaker [at] lbl.gov (Rebecca Hartman-Baker), LBNL/Cori
levesque [at] cray.com (John Levesque), Cray/Trinity-Cori
hnam [at] lanl.gov (Hai Ah Nam), LANL/Trinity

 

neely4 [at] llnl.gov (Rob Neely )(chair), LLNL/Sierra
sextonjc [at] us.ibm.com (Jim Sexton), IBM/Sierra-Summit
straatsmatp [at] ornl.gov (Tjerk Straatsma), ORNL/Summit
zippy [at] alcf.anl.gov (Tim Williams), ANL/Aurora
czeller [at] nvidia.com (Cyril Zeller), NVIDIA/Sierra-Summit