Spectrum Logo
A non-profit publication of the Office of the University Relations of Virginia Tech,
including The Conductor, a special section of the Spectrum printed 4 times a year

Tech SP turns three this year

Spectrum Volume 20 Issue 28 - April 16, 1998

The Virginia Tech Computing Center SP provides a unique resource to researchers running numerically intensive applications. The SP has been well received because it provides the means to run large jobs that were either difficult or impossible to run in the past. It has special data areas designed to hold large data files which eliminates some of the disk constraints experienced before the arrival of the SP.
In 1995 Virginia Tech received a grant from IBM to support university research by the acquisition of a 10-node RISC System/6000 Scalable POWERparallel System (SP). In 1996 another grant from IBM enabled the additional of four nodes as well as additional disk drives.
As well as the XL FORTRAN and C++ compilers, software on the SP includes the IMSL, ESSL and Parallel ESSL subroutine libraries. Serial and parallel analysis tools are also available to enhance performance of your code. Application software packages on the SP include SAS, Gaussian, Matlab and Mathematica.
The SP is many powerful processors in one, all interconnected by high-speed data links. Appropriately written software is able to use four or eight processors at once (in parallel) to achieve computing speeds significantly faster than other central IBM systems. The SP continues to represent a significant step forward in the university's attempts to meet burgeoning demand for the processing of large data sets (using packages like SAS), for faster execution of the numerically intensive algorithms used in computational research, and for long-running jobs.
The nodes of the SP are POWER2 architecture RS/6000 processors with their own memory and disk storage. These nodes are connected by a low-latency, high-bandwidth switching network called the High-Performance Switch. Virginia Tech has two wide nodes, each roughly equivalent to an RS/6000 model 590, and 12 thin nodes, each roughly equivalent to an RS/6000 model 390. All of the nodes have directly attached SCSI-2 fast and wide disk storage. Each wide node has 512 megabytes of memory and up to 163 gigabytes of disk space. Eight of the thin nodes have 128 megabytes of memory and 4.4 gigabytes of disk space and four thin nodes have 256 megabytes of memory and 4.4 gigabytes of disk space.
In addition to the 14 nodes on the SP, an IBM RS/6000 Symmetric Multi-Processor Machine (SMP) is available for running numerically intensive applications. This machine is capable of running multiple long-running jobs because it has six processors and 512MB of real memory. Users submit jobs to the SMP machine by logging in to the SP and using the same procedures they use to submit jobs to the SP nodes. Most jobs will use 10-30 percent more cpu time on the SMP system than an SP node. However, since the SMP system is scheduled to allow each job to have a dedicated CPU, jobs usually complete more quickly on the SMP than on an SP node shared with another job.
Applications presently running under AIX on other machines should be able to run serially on the SP with few, if any, changes. To take advantage of the parallel capabilities of this machine, it is necessary to use an application package which has already been written for multi-node processing or modify existing code to include calls to the appropriate routines in one of the parallel libraries. Opportunities for parallel execution are automatically detected by the High Performance Fortran compiler which is available for use on the SP.
Some examples of situations for which the SP would be suitable include long-running jobs that may run for multiple days; large (>100MB) disk requirements; memory requirement up to 512MB; numerically intensive programs currently running forever on another machine where sections of the code could be run concurrently; the need to use software that was designed to run concurrently on multiple processors.
Faculty and staff members and graduate students may obtain accounts on the SP by submitting an abstract describing the research that is the basis of the computing project to 4help@vt.edu by e-mail. The inclusion of related journal publications and preprints is welcomed. Since the machine was acquired with the help of a research grant from IBM, the abstracts will help affirm support for research computing and define upgrade requirements.
Submitted abstracts are made available on-line via hypertext links from the SP homepage (URL: http://spider01.cc.vt.edu/). Along with the abstract, include the following information which will be used to create a userid on the SP: your name, faculty/staff/student ID number, e-mail address; department, campus mailing address, phone number, user name (maximum eight characters), initial log in password (minimum six characters), desired UNIX (csh or ksh) shell.
Note: For those without a preferred shell, the Computing Center recommends use of the Korn shell.
Refer to the Virginia Tech web Page (URL:http://spider01.cc.vt.edu/) for additional details and SP-specific usage information. This page has links to Cornell's SP WEB Page where there is information on parallel programming and other SP topics.
Please send applications and abstracts to 4help@vt.edu.