June/July2007 News Letter

--------------------------------------------------------------------

--------------------------------------------------------------------

--------------------------------------------------------------------

The National Science Foundation awarded two HPCC users their most prestigious award for young

researchers. Texas Tech researcher Lenore Dai, Assistant Professor of Chemical Engineering, received

the CAREER award for her proposal, “CAREER: Heterogeneous and Competitive Self-assembly at Liquid-

Liquid Interfaces.”

Fellow HPCC user Jorge Morales, Assistant Professor of Chemistry, was also awarded the esteemed

CAREER award by the NSF. Morales’ proposal, “Building a Direct Dynamics with Coherent States” involves

inventing and developing new theories to computationally describe chemical reactions.

HPCC would like to congratulate Lenore Dai and Jorge Morales.

--------------------------------------------------------------------

--------------------------------------------------------------------

HPCC supplies several different forms of the Lapack and Blas mathematical libraries to link with your

programs. The Lapack library includes serial matrix and eigenvalue solvers. Lapack requires a Blas

library, and the Blas library may also be used independently. The Blas library includes basic matrix and

vector functions such as matrix multiply. These functions will be much faster if executed in the performance

Blas library than in your own code, or in other high-level code such as Numerical Recipes or the reference

Fortran Blas.

Lapack offloads most of its work to Blas, so its performance is mostly determined by the Blas. We have

compiled Lapack for the Intel and Gnu compilers in IA32 (m32) and X86_64 (m64). Most HPCC users will

want the m64 version. These libraries are located in /share/apps/lib/lapack/. It is open software.

The Goto Blas Library (Kazushige Goto, U. of Texas) is usually the fastest available. It is free for

educational use. The libraries are independent of compiler, but there are m32 and m64 versions, and also

versions for 32-bit (4-byte) matrix indexing, and 64-bit (8-byte) indexing. There are also different versions

for the single core Xeon version (“xeon”) in “old” hrothgar and minigar and for the multicore Xeons

(version “core”) in “new” hrothgar and antaeus.

HPCC also has the Intel MKL 7.2 implementation of Lapack and Blas. MKL is a reasonably good match for

any Intel processor. AMD also has a free library ACML for its processors, not installed at HPCC.

Both Intel and Goto Blas can use OpenMP multithreading inside nodes.

Here are some example runs for an order 3000 matrix solve using the double-precision Lapack solver dgesv

on a single quad-core node. Each run requires either export OMP_NUM_THREADS=1 or export

OMP_NUM_THREADS=8, and MKL, being dynamically linked, requires export

LD_LIBRARY_PATH=/opt/intel/mkl72cluster/lib/em64t. The dgesv function is included in Goto Blas, so a

call of the Lapack libraries is then not required. The default ifort compiler is used, so m64 (or Intel

em64t) libraries are required throughout. Default (i4) integers are used.

The best Goto Blas version is more than 5 times faster than the reference Blas with one thread and 30 times

faster with 8 threads. Also, the 41.1 Gflops from a single dual-quad core is about twice the performance of

our entire SGI Origin, now retired.

Please contact David Chaffin (David.Chaffin@ttu.edu) for help or more information.

--------------------------------------------------------------------

--------------------------------------------------------------------

TTU's participation in the Texas Internet Grid for Research and Education (TIGRE) received a boost in May

when the project selected a state-wide metascheduler for use in coordinating job submissions. The Grid

Resource Management System (GRMS) was selected after comparison with several other grid-oriented

multi-site metascheduling choices. Work is still being conducted with GridWay, an alternate choice, to

provide backup and a basis for feature comparison. According to our TIGRE project scientists Ravi Vadapalli

and Alan Sill, this central metascheduling system will provide improved access to computation and storage

systems throughout the state at institutions participating in the TIGRE project.

Considerable progress has been demonstrated now in all three of the targeted application areas for TIGRE,

consisting of Biosciences and Medicine, Air Quality Modeling and Energy Exploration. These targeted

application areas were chosen by the project steering committee, consisting of principal investigators from

Texas Tech, Texas A&M, Rice University, the University of Houston and the Texas Advanced Computing

Center at UT - Austin. The combined computational resources contributed so far for testing of TIGRE have

already been supplemented by the addition of several resources contributed by early participants in the

project at UT Health Sciences Center - San Antonio. TIGRE project scientists have also given talks at

several important conferences and meetings, including the International Symposium on Grid Computing in

Taipei, Taiwan (via video), the High Performance Computing Across Texas (HiPCAT) meeting at UT - El

Paso, and meetings of several other grid-related groups, as well as presentations and posters at the TTY

System Cancer Research Group Symposium and the upcoming Teragrid conference.

The existence of a project-wide metascheduling system will improve the potential and capabilities for use of

TIGRE across a wide variety of other application areas as well. The TIGRE software selections discussed

below are also being adopted within the SURAgrid project (http://sura.org) and are being looked at closely

by several other state-wide and regional grid efforts.

http://gridge.org/grms

http://tigreportal.hipcat.net

--------------------------------------------------------------------

--------------------------------------------------------------------

In other news related to TIGRE, two senior members of the Globus software team (http://globus.org) visited

TTU in late April to discuss projects of mutual interest to both teams. TIGRE has attracted a high degree of

national and international attention from software teams through its decision to focus on implementation of a

web services oriented architecture through a variety of simplifying software choices. TIGRE client and

server software stacks available through the TIGRE portal (http://tigreportal.hipcat.net) based on the Virtual

Data Toolkit (http://vdt.cs.wisc.edu) and other related components can be installed on

a variety of scientific computing platforms, and the GRMS job submission client discussed above can be run

on essentially any modern platform, including Windows, Mac OS X and Linux.

The two Globus team members, Dan Fraser (manager of Globus development) and Rachana

Ananthakrishnan (senior security developer) discussed a wide variety of topics ranging from application

development to security architectures over the course of their two- day visit. They also met by previous

arrangement with researchers Shameem Siddiqui from the Petroleum Engineering, Sukanta Basu of

Atmospheric Sciences and Mike Sobolewski and Noe Lopez-Benitez in the Computer Science department at

TTU. The methods used by the TIGRE project to engage with researchers in these and other related areas

were discussed and complimented by the Globus team. Globus software forms the basis of most grid

projects deployed throughout the world, so this was a high compliment to the members of our TIGRE team.

--------------------------------------------------------------------

--------------------------------------------------------------------

If you have questions about HPCC at TTU or need assistance, there are several ways to receive help. Our

office is located in Administration Building, room 233 and our office phone number is (806) 742-4350.

Additionally, you may contact us via email at hpcc@ttu.edu.

To request an account with HPCC, please complete the electronic form on our website. If you have

problems which our website does not enable you to solve, questions about system outages, or if you just

have an issue and you don't know who to contact, send an email to hpcc@ttu.edu.

Please email comments, questions, and subscribe/unsubscribe requests, to hpcc@ttu.edu.

--------------------------------------------------------------------