Intel/Hewlett-Packard Project
Introduction
In September, 2001, the University of Kentucky received a 2001 HP/Intel
Itanium-based System Award. This consisted of two HP i2000 workstations. The workstations have been used in four projects so far:
- Port and optimize a partial differential equation (PDE) solver.
- Investigate how effective the Itanium is for a particular life
sciences project.
- Investigate how effective the Itanium is for some ocean
modeling codes.
- Globus 2.0 port for HP-UX.
One
of the i2000's is running Red Hat Linux 7.1 and the Intel compilers.
The other is running HP-UX and HP's and GNU's compilers.
PDE
Solvers
We tuned Dimepack, which is a state of the art constant
coefficient scalar PDE equation solver. The tuning will allow us to
easily port the code to any further IA-64 processors that are in the
pipeline (at least the ones that I have heard about under non-disclosure
from Dr. Greg Astfalk).
We also are writing a new solver for one or more
processors that is cache aware and can handle coupled PDEs and variable
coefficients. The two
dimensional code is working and the three dimensional solver is nearly so.
It has also been tested on an HP 4610 4-way server that HP donated in
the late fall, 2001.
We will be first reporting on the work at the
International Conference on Computational Sciences 2000 in Amsterdam in late
April, 2002. One or more papers
will be written and submitted for publication in archival journals by June,
2002. A draft of the three
dimensional paper, which will appear in the conference proceedings, is
available on the web. The
numerical experiments section has to be carefully completed and reviewed
before a journal article is submitted.
We gave very preliminary results at a half day
tutorial at SC2001 (Denver, November, 2001).
We expect to give final results for both the Itanium and McKinley
processors at a full day tutorial at SC2002 (Baltimore, November, 2002).
Brain Scanning
Using a single chip MAReNIR scanning
system, we are left trying to determine the 15-20 largest moments in a
system of order at least two million. The
results tell us if the blood in a brain is crossing an unacceptable toxicity
level, the location of an aneurysm, or the degree a patient is developing
Alzheimer’s disease. (It can
also pick out a vehicle a kilometer away in a pine forest, a problem of
extreme difficulty for the army now.)
We only need good enough estimates, not the exact
moments. We can estimate them
using much smaller, representative matrices on the order of a few thousand.
We have tested a sample code for calculating the moments on both
Linux and HP-UX on the i2000’s and on a HP Superdome processor (HP-PA 8700
chip) running HP-UX. The Linux timings bode very well for the
McKinley.
Ocean Modeling
The code works well on Linux. The code refuses to link on HP-UX due to a compiling problem.
It is unclear at the moment if the problem is the Fortran-95 compiler
or the code itself using some feature of Fortran-90 that is incorrect for
Fortran-95. We are sufficiently
happy with the Linux results that we have not pursued the problem on the HP-UX side.
A paper with the results will be submitted in May, 2002
to an archival journal. The
code normally runs on 48-512 processors.
On one processor, with two algorithmic variants that are active
research, we have the timings (in seconds) for 200 time steps. The timings bode very, very well for McKinley
based workstations.
Globus 2.0 Port for HP-UX
We have used the HP-UX i2000 workstation as part of a
project to port Globus 2.0 beta to HP-UX.
This is a very large set of packages, coming in a large number of
bundles. All of the
bundles have been made and are still being tested and debugged.
This is nontrivial since HP-UX has been completely ignored by the
Globus group. HP
unfortunately does not yet provide quite all of the library support
required. We expect to have the first complete port of Globus 2.0.
The official release of version 2.0 occurred very late on Friday,
April 12. We started over with
that and are making fast progress back to where we were with 2.0 beta
earlier in the day.
The Linux i2000 made all of Globus 2.0 beta in about 36
hours. It has been tested
successfully.
Working Papers
My prime paper writing period is May through August and
a short period in December to early January.
- C. C. Douglas, G. Haase, and J. Hu, The State of Cache Aware
Multigrid on Unstructured Grids, expected submission date: May,
2002. See DRAFT.
- C. C. Douglas, J. Hu, J. Ray, D. T. Thorne, and R. S. Tuminaro, Fast
Adaptively Refined Computational Elements in 3D, in Proceedings of
the International Conference on Computational Science, Amsterdam, 2002,
Springer-Verlag, Berlin, 2002, 10 pages. See PAPER.
- C. C. Douglas, J. Hu, J. Ray, D. T. Thorne, and R. S. Tuminaro, Fast
Adaptively Refined Computational Elements, expected submission date:
Summer, 2002. Paper will greatly expand on the one in 2 and go to
an archival journal.
- C. C. Douglas and R. Lodder, MAReNIR Moments Calculation, Report,
2002. See REPORT.
- C. C. Douglas, G. Haase, and M. Iskandarani, Ocean
Modeling on Complex Domains, expected submission date: May, 2002.
- C. C. Douglas, V. Chopra, and C. Li, Globus 2.0 beta Port to HP-UX,
Report, 2002. See WEB
PAGE.
- C. C. Douglas, V. Chopra, and C. Li, Globus Version 2.0 Port to
HP-UX, Report, 2002. See WEB
PAGE.