RISC SERVERS

By John H. MayerSpecial to GCNA little over a decade ago, designers of reduced-instruction-set computing, or RISC, processors stormed into the systems arena with a radical idea. Build a processor with a simpler, more elegant architecture that would execute just one instruction per clock cycle instead of many, and you could end up with a smaller, faster CPU capable of delivering turbocharged processing.The market jumped on the idea, and the terms RISC and high-performance computing quickly became synonymous. Systems built around processors based on traditional complex-instruction-set computing (CISC) concepts, such as Intel's X86 and Motorola's 680X0 families, were quickly relegated to low-end desktop PC applications.But things don't stay the same for long in information technology. RISC servers still own the high end of the market and usually win all industry performance awards. But the RISC radicals who proposed their revolutionary idea just a few years ago are now looking over their own shoulders.Distinctions between RISC and CISC systems have evaporated as designers from each camp incorporated concepts from the other into succeeding designs. Next-generation processors, such as the Merced from Intel, promise performance on par with high-end RISC chips while retaining the application support and price-performance advantages of Microsoft Windows operating systems.The highly anticipated Merced, due next year, not only will run Microsoft's first 64-bit OS but also has won commitments from most of the leading suppliers of RISC Unix systems. Hewlett-Packard Co., a co-developer of Merced, and Compaq Computer Corp. are developing systems using the new processor.IBM Corp., maker of the PowerPC, will work with other vendors on a new open version of Unix for the Merced platform. Even Sun Microsystems Inc. is enabling its SunSoft Solaris OS to run on Merced. And Intel Corp.'s plans to introduce 64-bit processors for enterprise applications'the McKinley in 2001 and the Madison in 2002'are expected to keep driving processing capability skyward.Complicating the picture has been the rise of Microsoft Windows NT in large enterprisewide applications. Once considered unsophisticated and unreliable for large, mission-critical applications, the latest versions of NT have slowly eliminated those reservations.Backed by its massive breadth of applications, NT has quickly picked up momentum in institutions. And at the low end, many users who haven't already moved to NT plan to make the change. For more and more IT managers, the prevailing belief is that you cannot go wrong by betting on servers with Intel chips running NT.Numbers from market research analysts back up that view. International Data Corp. of Framingham, Mass., reported this year that sales of Unix systems declined for the second year in a row, and researchers at Dataquest Inc. of San Jose, Calif., said they expect Unix server sales to remain flat through 2002.Vendors of RISC servers are well aware of the shifting market. Hewlett-Packard is playing a major role in the development of the Merced processor and will focus on Intel processor systems. Earlier this year SGI announced its intention to shift away from Unix systems and toward workstations and servers with Intel processors. And most other leading RISC server vendors, with the exception of Sun, are offering or planning to offer systems based on Intel chips.Despite market trends, there are still plenty of good reasons to opt for a RISC server.One is server consolidation. Organizations looking to reduce support costs and simplify IT management find that RISC servers' ability to support more users and larger amounts of storage per system is a formidable advantage, particularly as IT managers seek to consolidate the tasks performed by multiple smaller servers into one. IT managers are quickly discovering that fewer servers mean fewer service calls, and that allows them to cut support staff or reallocate coveted resources.Raw performance is the best reason to go with a RISC system. In general, RISC servers continue to outperform those using Intel processors, according to industry-standard benchmarks such as the Transaction Processing Council's Benchmark C. And RISC system vendors say their products will continue to outperform Merced and its successors over the next few years.But the difference depends largely on applications. The performance advantage RISC systems have over non-RISC systems can be up to 50 percent in some applications but negligible in others, analysts said.Compute-intensive server applications such as data warehousing, decision support or large enterprise resource planning tasks are particularly able to take advantage of the tremendous processing power RISC servers offer.For example, Sun Microsystems, the leading vendor of Unix servers, recently announced that its top-of-the-line Enterprise 10000 servers set an industry record for the Standard Bank Customer Accounts benchmark from SAP America Inc. of Wayne, Pa. The benchmark measures processing of batch and online postings, as well as the si-multaneous display of account balances and statements. In applications such as these, a system's ability to handle a tremendously high volume of transactions at high speed from a central location is critical.RISC servers' successes with high-end applications highlight one of the most important advantages they offer over servers with Intel chips'scalability. Most RISC system vendors offer symmetric multiprocessing systems that can be configured with 20 or more processors.Rather than run an application on one processor, SMP systems break up the task and distribute the workload across two, four, eight or more. By sharing memory across those processors, the systems deliver billions of instructions per second of processing power while running an industry-standard OS such as Unix or NT.Moreover, users can scale their systems simply by adding more processors, a key consideration for IT departments running data warehouse or decision support systems that quickly grow to multiterabyte size.The real test for an SMP system, however, is its ability to display performance linearity as it expands from two processors to four, eight, 12 or 16. Eventually, as the user adds more processors, all SMP systems hit a wall. Data transfer between processors reaches a point that it becomes a bottleneck and performance no longer improves.Historically, RISC servers running Unix have exhibited excellent linearity as they scale up to 20 processors and beyond. That characteristic contrasts sharply with systems running NT, which'until the most recent version of the OS came along'offered little performance improvement once a system went beyond four processors.Despite the attention paid to the capabilities of individual RISC processors, experts recommend that IT managers in the market for a server take a close look at a server's system architecture and its applicability to specific applications.SMP system developers typically rely heavily on high-speed system buses, specialized input/output architectures, and extensive memory caches to keep data moving across a multiprocessor system. Many RISC servers use a traditional system design built around a very high-performance bus to accommodate the massive CPU-to-memory traffic generated by shared-memory SMP architectures. Others use a crossbar switch or some other specialized interconnect to help speed data throughput.Generally, the larger the capacity of the system bus, the less likely a server will encounter system bottlenecks as the processors in the system pass increasingly larger amounts of data back and forth.Another important consideration in selecting a RISC server is its data storage and I/O bandwidth capabilities. Typically, mass storage is the most expensive part of a server. So it is critical that you buy a configuration that meets your needs but does not exceed them to avoid wasting critical resources. Although some low-end RISC servers support only a limited amount of disk storage internal to the system, most accommodate multiterabytes of disk arrays external to the system.Nevertheless, it is important that the overall system architecture is capable of supporting that mass-storage configuration. Applications for high-end data warehousing and decision support, for example, typically bring together massive amounts of data and require substantial disk or tape storage.In those applications the ability to move data off a disk and into the processors can easily overshadow raw processor speed. Usually, these types of applications require an extremely high number of simultaneous disk accesses and the I/O bandwidth to support them. Specifications such as maximum I/O bandwidth and the ability to support large amounts of disk storage may play a more important role in your system selection than the type or number of processors supported.The ability to handle large amounts of RAM is also critical. Systems with large memory configurations let users place entire applications of large portions of databases into physical memory and minimize disk I/O. Once limited to relatively small memory configurations, Unix systems today can support many gigabytes of RAM.Vendors and independent analysts offer a variety of benchmarks to help you measure the performance of one server against another, but the best place to start is with a few key questions: What will the system be used for? How many will use it? Is the application going to be disk- or calculation-intensive? How fast will the application requirements grow? The best system for you will not be the one with the fastest processor or largest disk capacity but the system with the right configuration for your needs.XXXSPLITXXX-