HIGH-PERFORMANCE COMPUTING

In Scientific Computing, Taking RISCs Can Yield Better Speed and Performance

Edward Friedman

[Ed: Links to web pages and/or e-mail addresses which have become inactive since the publication of this article have been enclosed in curly brackets { }. Replacement links have been provided where possible.]

Technical advances in materials research, circuit design, and fabrication have increased the density and speed of the components of RISC computers, most importantly the central processor. A consequence of these innovations is that the computing systems scale easily - that is, they can be made faster or larger, or used in multiples, without the need to redesign the basic components. These advances not only enhance existing efforts of users by making the work easier to accomplish and faster, but allow them to consider and solve more interesting and complex problems.

RISC systems have fewer and simpler instructions, whose execution times are directly proportional to the clock speed of the processor. The latest RISC processors can also issue and execute several instructions simultaneously, resulting in parallel processing within a single processor.

Two kinds of scalability are then possible, increased speed and larger problems. Having several RISC sytems working concurrently on a problem decreases the real time needed to accomplish the work. It also offers researchers the potential to increase the size and complexity of the scientific investigation.

RISC technology is not new, but only recently has it become widely available. This is in part due to the introduction of the Motorola Power PC chip set developed jointly by Apple, IBM, and Motorola, and by the rapid development and deployment of the Alpha processor by DEC (Digital Equipment Corp.). Silicon Graphics, Inc., has used RISC chips from its MIPS division for the central processors of its workstations and servers for nearly ten years, and is working on specialized RISC processors for high-end graphics. SGI also supplies RISC processors to majers of electronic games and networking devices, which need relatively large numbers of calculations per second for their effects and functions. Hewlett-Packard, a leader in engineering workstations, has continued to develop its PA systems, which have been used in both American and Japanese supercomputers. IBM has integrated Power PC chips into workstations and will likely be offering them in the successor to its SP-2 multiprocessor systems.

RISC systems range from the single-processor desktop units to very large multiprocessor systems configured into a variety of architectures.

Most of the major manufacturers of workstations and desktop systems have incorporated RISC technology into their systems. These include Apple, DEC, IBM, Intel, HP, SGI, and Sun. Larger RISC systems made up of clusters of individual workstations interconnected by high-speed networks and highly integrated multiprocessor shared-memory systems have been rapidly incorporated into research, industrial, and commercial enterprises. These developments probably hastened the demise or merger of several major computer manufacturers within the last few years: Cray Computer Corporation, Convex, Cray Research, Thinking Machines, and KSR all relied on proprietary rather than commodity components, leaving them at a competitive disadvantage.

Systems based on the Intel 80X86 CISC technology, currently the Pentium and Pentium Pro, have also improved performance. Many researchers are running variants of Unix (such as Linux or BSD/OS) on systems with one or more of these processors. The computational power of these systems compares very well to RISC-based systems, and their price-performance ratio is attractive.

Processor clock speeds of over 100 megahertz (MHz) are now common on both the RISC and CISC processors. Speeds of over 300 MHz are now being offered by DEC, and it is likely all the other vendors will soon follow. However, clock speed is only one characteristic of a computer system, and performance can be measured only by running an application on several systems and comparing the results. Performance measurements of so-called benchmarks are published both by the vendor and by several independent organizations, and are readily available from sources on the World-Wide Web.

The latest Linpack Benchmarks report can be obtained by going to the Netlib home page - http://www.netlib.org/. The Linpack Benchmark makes an attempt to measure floating-point CPU performance of a single linear-algebra problem which is of interest to scientists. It should be used only as a rough comparative measure of different computing systems, and must be used in conjunction with other measures to obtain an evaluation of overall system performance.

RISC Systems at the ACF

The ACF plans to replace most of its older shared servers with RISC systems from either DEC, IBM, or SGI. Sparc RISC-based systems from Sun Microsystems are in use at the Courant Institute and at the Center for Neural Sciences, and these systems are being upgraded with new RISC offerings from Sun.

In this rapidly evolving technology, any comparison of the characteristics of RISC processors is quickly out of date. For those interested in making their own comparisons, the box below lists the World-Wide Web sites for the major producers of RISC systems.

Major Makers of RISC Systems

Apple Computer	http://www.apple.com/
DEC	http://www.dec.com/
IBM	http://www.ibm.com/
Silicon Graphics	http://www.sgi.com/
MIPS	http://www.mips.com/
Motorola	http://www.motorola.com/
Hewlett-Packard	http://www.hp.com/
Sun Microsystems	http://www.sun.com/
Intel	http://www.intel.com/

Posted 20 May 1996. Revised 24 May 2004.