| Meeting Main Page | ||
| Schedule of Talks | Abstracts and Materials | The Access Grid |
| Download page for presentation source files for all presentations | ||
10:00 - 10:30 Scaling: Techniques and Pitfalls
David Skinner, User Services Group
10:30 - 11:15 Understanding and Using Profiling Tools on Seaborg
Richard Gerber, User Services Group
11:15 - 11:30 Break
11:30 - 12:00 IBM Compiler Optimization Arguments
Michael Stewart, User Services Group
12:00 - 12:45 Effective Memory Use
David Skinner, User Services Group
12:45 - 1:45 Lunch
1:45 - 2:30 Libraries and Their Performance
Thomas DeBoni & Frank Hale, User Services Group
2:30 - 3:15 A Comparison of Performance Analysis Tools on the SP
Jonathan Carter, User Services Group
3:15 - 3:30 Break
3:30 - 4:15 Hybrid MPI and OpenMP Programming on the SP
Yun Helen He, Scientific Computing Group
4:15 - 5:00 Survey of Visualization Tools and Capabilities at NERSC
Ken Schwartz, Vizualization Group
Scaling: Techniques and Pitfalls
David Skinner, User Services Group
Scaling issues for IBM SP scientific applications will be presented. The basics of the SP colony switch, MPI coding practices, and parallel I/O will be addressed along with common pitfalls in scaling up scientific codes on the IBM SP.
Understanding and Using Profiling Tools on Seaborg
Richard Gerber, User Services Group
I will describe the POWER 3 hardware counters and discuss some simple tools that can be used to profile a code: hpmcount, poe+ and the HPM library. I will describe the meaning of a number of the metrics reported by these tools. Finally, I'll show some simple optimization tips.
IBM Compiler Optimization Arguments
Michael Stewart, User Services Group
More than with most other vendor's compilers, it is necessary to select the proper compiler optimization arguments when compiling with IBM's Fortran, C, and C++ compilers to attain any level of acceptable performance. This talk will describe the most important optimization arguments and describe their effects on several test codes. This updated presentation contains new information and examples not in earlier versions.
Effective Memory Use
David Skinner, User Services Group
An overview of memory bandwidth issues impacting user codes on the POWER 3 architecture is presented. The basics of the memory architecture, bandwidth available to common programming idioms, misalignment, indirect addressing, and loop overhead will be treated.
Libraries and Performance
Thomas DeBoni and Frank Hale, User Services Group
The NERSC user environment is quite rich and offers a variety of alternative methods and library routines for performing equivalent computations. How can a user determine the relative performance of different functionally equivalent programming methods and library routines? This presentation will include an introduction to the use of the HPMToolkit for instrumenting program sections and comparing performance of different functionally equivalent user programming techniques and mathematical library functions.
A Comparison of Performance Analysis Tools on the SP
Jonathan Carter, User Services Group
This talk will introduce the IBM PE Benchmarker, Paraver, and VAMPIR tools. Each of these tools is designed to collect some subset of: hardware performance data, OpenMP scheduling data, and MPI message statistics. We will compare the use of these tools for analysing and tuning some simple programs.
Hybrid MPI and OpenMP Programming on the SP
Yun Helen He, NERSC Scientific Computing Group
Mixed MPI/openMP is the software trend for SMP architectures. But up until now here are still more negative experiences than positive ones. This talk will first cover some of the basic methodologies of using mixed MPI and openMP on the IBM SP. Issues and methods of tuning the performance of hybrid codes will then be discussed. Results from some real applications will be presented. Among these, one particular application shows that the hybrid MPI/OpenMP paradigm outperforms the pure MPI paradigm by a factor of more than 4.
Survey of Visualization Tools and Capabilities at NERSC
Ken Schwartz, Visualization Group
An Access Grid virtual meeting is an internet-mediated audio-video conference. Access Grid facilities are available at most DOE sites; see your local ESNet representative to find out the location and availability of an Access Grid Node at your site.
To verify the scheduling of this event, search for it on the "schedule" page of the Access Grid Web Site.
Info for AG Node Operators:
| Test Session: Wed., 28 May, 1:30 - 3:00 CDT, in the Big Horn Room. Please get all available materials from the download site before the test, and make sure you have installed working viewers on your nodes for MPEG, QuickTime, PDF, and PostScript files. | |
| Argonne Local Arrangements: Mike Minkoff; (630) 252-7234 | |
| Lead Operator: Mary Fritsch of ANL |
Coordinator:
Tom DeBoni of NERSC |
|
Virtual Venue:
the Big Horn Room |
Distributed PowerPoint Server: a216-display.mcs.anl.gov |
|
QuickBridge Servers: audio - 140.221.9.160/50028 video - 140.221.9.160/50030 |
Audio Bridge:
Local: (510) 486-5008 # 7377 Long Distance: (877) 252-5250 # 7377 VNC Server: to be established |
|
Presentation source files for all presentations can be
downloaded here. Watch your MUD for speaker/server changes. Media Timetable - file types for each talk - here. | |
Presentations will be presented to the remote sites via Distributed PowerPoint, which allows participating sites to view high quality local copies of the slides, under control of the presenter. The presentation source files should be downloaded by Node operators in advance of the event. To ensure consistency, please download final copies of each talk no earlier than 24 hours prior to the event.