SP SciComp 1999

October 6 - 8, 1999
Yorktown Heights, New York

www.spscicomp.org

The first annual IBM SP Scientific Computing Meeting, SP SciComp 1999, was hosted by the IBM Advanced Computing Technology Center (ACTC) at IBM Research. The meeting, open to 285 participants, was a user-oriented and planned 3-day workshop to share information on scientific computing techniques for SP users. The meeting was created on the recommendations of the attendees of the IBM SP Scientific Applications Development and Optimization Workshop held in March 1999 at the San Diego Supercomputer Center. The objective of that meeting was to help computational scientists and engineers develop applications that achieve maximum performance and scalability on the IBM SP systems.

The October meeting consisted of presentations by IBM staff and users from academia and government labs covered topics such as migrating codes from other systems, performance analysis and optimization for SP systems, programming tools and techniques, etc. Tutorials by IBM staff, open to 40 participants, were held on the two days prior to the meeting, October 4 - 5, in a computerized classroom.

Meeting Agenda

TUTORIALS
IBM Thomas J. Watson Research Center
Hawthorne, New York

Monday, October 4, 1999

MPI and OpenMP Internals for Programmers
Hawthorne 1 H1-F40

8:30 AM

5 Arrivals / Registration - Lobby, Hawthorne Building 1

9:00 AM

MPI Internals
  • POE, MPI Libraries, and Runtime Program Structure
  • Message protocols and their flows; the MPI protocol stack
  • User space communication
William Tuel
Richard Treumann
10:30 AM Break

10:45 AM

MPI Internals, Part II
  • Threads and the MPI Progress Rule
  • CCL algorithms
  • MPI-IO Architecture and Structural Details
  • Some environment variables and their significance
William Tuel
Richard Treumann

12:00 PM

Lunch

1:00 PM

OpenMP Internals
  • OpenMP implementation overview
  • Compiler transformations
  • Issues with Fortran90 conflicts
  • Thread management and scheduling policies
  • Performance issues and bottlenecks
Raul Silvera

3:00 PM

Break

3:15 PM

Mixed MPI/OpenMP Programming
  • Performance and scalability issues
  • 64-bit without 64-bit
  • Tips and tricks
Stefan Andersson
David Klepacki

4:45 PM

Questions / Wrapup

5:30 PM

Adjourn

Power3 Optimization
Hawthorne 2 J2-B34

8:30 AM

Arrivals / Registration - Lobby, Hawthorne Building 1

9:00 AM

Architecture of the Power3 John Levesque

9:45 AM

Cache and Register Blocking Fred Gustavson

10:30 AM

Break

10:45 AM

Cache and Register Blocking (Cont'd) Fred Gustavson

11:30 AM

Tools for Analyzing Performance Bob Walkup

12:00 PM

Lunch

1:00 PM

Atlas Code Generation Fred Gustavson

2:00 PM

Blocking for Non-Matrix Operations John Levesque

3:00 PM

Exercises

4:45 PM

Wrapup

5:30 PM

Adjourn

Tuesday, October 5, 1999

OpenMP / IBM Tools
Hawthorne 2 J2-B34

8:30 AM

Arrivals / Registration - Lobby, Hawthorne Building 1

9:00 AM

OpenMP John Levesque

9:45 AM

SMP Performance Programming Anshul Gupta

10:30 AM

Break

10:45 AM

SMP Performance Programming (Cont'd) Anshul Gupta

11:30 AM

Tools for Analyzing Performance Bob Walkup

12:00 PM

Lunch

1:00 PM

Programming in Pthreads Bob Walkup

2:00 PM

Runtime Considerations John Levesque

3:00 PM

Exercises

4:45 PM

Wrapup

5:30 PM

Adjourn

High-Performance I/O with GPFS
Hawthorne 1 H1-E/F53

8:30 AM

Arrivals / Registration - Lobby, Hawthorne Building 1

9:00 AM

Introduction to GPFS Carol Hartman

10:00 AM

Break

10:15 AM

Application Do's and Don'ts and their results Carol Hartman

11:30 AM

User questions / Exercises

12:00 PM

Lunch

1:00 PM

GPFS Installation - performance considerations Carol Hartman

2:00 PM

MPI-IO Jean-Pierre Prost

3:30 PM

Break

3:45 PM

I/O Optimization using user space prefetch and caching John Bauer

4:45 PM

Wrapup / Questions

5:30 PM

Adjourn

CONFERENCE
IBM Thomas J. Watson Research Center
Yorktown Heights, New York
October 6-8, 1999

Note: Most of the IBM presentations listed below are available from ACTC's Web site at http://www.research.ibm.com/actc

Wednesday, October 6, 1999

8:30 AM

Arrivals / Registration

9:00 AM

Welcome Jay Boisseau, San Diego Supercomputer Center, and John Levesque, IBM Advanced Computing Technology Center

9:05 AM

Deep Computing Institute William Pulleyblank, Director IBM Deep Computing Institute

10:00 AM

ACTC John Levesque

10:30 AM

Architecture of the Power3 Frank O'Connell, IBM

11:30 AM

Break

12:00 PM

OpenMP in C and Fortran Henry Zongaro, IBM

1:00 PM

MPI: Status and Directions Dick Treumann, IBM

1:30 PM

Lunch

2:30 PM

Compiler Update (C, C++, Fortran...) Henry Zongaro, IBM

3:30 PM

IBM Tools Judy Ingles, IBM

4:30 PM

Break

5:00 PM

LAPI: Status and Directions Rama Govindaraju, IBM

6:00 PM

Break / Reception

6:30 PM

Catered Dinner (All attendees are invited)

Thursday, October 7

8:30 AM

Arrivals / Registration

9:00 AM

Next Generation Software Research Directions Frederica Darema, NSF

10:00 AM

Current Switch Architecture Jamshed Mirza, IBM Research

10:45 AM

Future Roadmap: Hardware Jamshed Mirza, IBM Research

11:30 AM

Break

12:00 PM

Future Roadmap: Software Peg Williams, IBM

12:45 PM

Mixed-Mode Operation (Threads/MPI) David Klepacki, IBM ACTC

1:30 PM

Lunch

2:30 PM

Porting Applications to the SP John Levesque, Director, IBM ACTC

3:15 PM

Debugging with TotalView on the IBM SP Mary Kay Bunde, Etnus, Inc.

3:45 PM

Parallelization and Performance Issues of a Monte Carlo Photon Transport Code on Nighthawk and other machines Amitava Majumdar, San Diego Supercomputer Center

4:15 PM

Break

4:45 PM

I/O Issues in PGS's Application Suite Ray Paden, PGS Shared Processing Services

5:15 PM

What Makes ORNL's SP Different Richard Alexander, Oak Ridge National Laboratory

5:45 PM

Porting 64-bit MPI and Distributed Math/Numerical Libraries to the IBM SP Shuxia Zhang, University of Minnesota

6:15 PM

Adjourn

Friday, October 8

8:30 AM

Arrivals / Registration

9:00 AM

Performance Tuning and Evaluation of a Parallel Community Climate Model Pat Worley and John Drake, Oak Ridge National Laboratory

9:30 AM

Automatically Tuned Linear Algebra Software Jack Dongarra, ORNL / University of Tennessee

10:00 AM

Multi-Level Collective Communications Bronis R. de Supinski, Lawrence Livermore National Laboratory

10:30 AM

PAPI: Performance Data Standard and API Phil Mucci, University of Tennessee

11:00 AM

Break

11:30 AM

Multi-Level Parallel Implementation of the MIT General Circulation Model on the SP Chris Hill, Massachusetts Institute of Technology

12:00 PM

Parallel Performance of a 3D ALE ASCI Code Evi Dube and Scott Futral, Lawrence Livermore National Laboratory

12:30 PM

High Performance Biocomputing Philip LoCascio, Oak Ridge National Laboratory

1:00 PM

Intra-Message and Inter-Message Pipeline Effects: Contrasts and Comparisons of Communication Performance V. Getov, University of Westminster and
P. Gray and V. Sunderam, Emory University

1:30 PM

Lunch

2:30 PM

Discussion: Future Directions SP SciComp Committee

5:00 PM

Adjourn

SP SciComp 1999 Committee

 Jay Boisseau, San Diego Supercomputer Center
Evi Dube, Lawrence Livermore National Laboratory
Trey White, Oak Ridge National Laboratory
Shuxia Zhang, University of Minnesota
John Levesque, IBM ACTC
and Andy Pierce, IBM ACTC
Please direct questions to Jay Boisseau at boisseau@sdsc.edu.