New Choices for Workstations and Servers

Despite the rapid advance of processor performance, there are times when one CPU just isn’t enough. That’s when businesses — and increasingly individual users — are doubling or even quadrupling their computing power by opting for multiprocessor systems.

For years, big companies have relied on multiprocessing for server applications (imagine a train pulled by two locomotives), and hardcore PC enthusiasts have built “duallies” to speed up games like Quake III Arena (imagine a dragster with two engines). But now a new crop of multi-CPU systems from both big and small vendors is reaching the market, along with mainstream, multiprocessing-compatible operating systems like Windows 2000 and XP Professional and Mac OS X, to make multiprocessing affordable and practical for high-end PC applications.

The Secret To SMP

Any computer that has more than one CPU — typically two, four, six, eight, 16, or 32 of them — is a multiprocessor system. The super desktop PCs known as workstations, used frequently for computer-aided design (CAD), 3D rendering, or complex financial modeling, are typically dual-processor machines — capable of what’s called two-way multiprocessing. The higher multiples — four-way and eight-way multiprocessing and so on — are reserved for network servers. The simultaneous use of several CPUs is also called parallel processing, although that term is increasingly used to describe massively parallel processing or grid computing — the division of a task among multiple computers connected by a network (as in the SETI@home screensaver).

Multiprocessing comes in two flavors. Asymmetric systems reserve separate processors for specific tasks — one to handle disk operations, say, while another crunches numbers. No major operating systems support asymmetric multiprocessing; it’s used only in exotic applications. Today’s workstations and servers rely on symmetric multiprocessing (SMP), in which any idle processor can be assigned any task, and more CPUs can be added to improve performance with bigger workloads.

Multiprocessing is sometimes confused with multitasking, in which an operating system like Windows loads and runs two or more tasks or programs in memory at once (right now, for instance, you probably have your e-mail program and word processor open as well as your Web browser).

Technically, on your uniprocessor PC, the OS allocates CPU cycles to one task, then another with such lightning speed that they seem to be running simultaneously. On an SMP system, they really are — or a single task can run faster (though not fully twice as fast; system overhead keeps a second processor from delivering a 100-percent performance boost).

SMP Software

Indeed, while it would seem obvious that adding more processors would make a program run faster, it isn’t that simple — operating systems and applications must be written to take advantage of multiple processors. Such software is called multithreaded, because it can divide tasks into multiple threads and execute one without interfering with another or getting overall operation scheduling out of sync. Because writing multithreaded applications is harder and takes longer than writing single-threaded ones, not all software vendors put in the extra effort.

Users of Windows 95, 98, and Me — and Windows XP Home Edition — will get zero performance gain when using an SMP system; those operating systems can’t “see” or recognize an extra processor, even when running applications that can. By contrast, Windows XP Professional, Windows 2000 Professional, and Windows NT Workstation all support two CPUs; the server editions of Win NT, 2000, and XP support as many as eight.

Outside the Microsoft universe, Unix and its relatives including BSD, Solaris, and Mac OS X all support multiprocessing, and the latest Linux distributions support up to 32 CPUs. However, getting these systems to recognize additional processors may take extra work — if you’re running Linux, for example, make sure you’ve installed the SMP kernel.

Although having a Web or application server that runs faster is great for businesses, individual users can take advantage of a growing number of SMP-enabled, multithreaded programs. SMP really excels in processor-intensive tasks such as image and video editing, and these are the most common applications that support it. Among the many multimedia applications that profit from a second processor are Adobe Photoshop and Premiere, Autodesk’s AutoCAD 2000i, Avid Technology’s SoftImage, Sonic Foundry’s Sound Forge, Discreet’s 3DS Max (formerly 3D Studio Max), and NewTek’s LightWave.

Multiple Processors, Multiple Systems

Just as with software, hardware works best when specially designed for SMP. While PC hobbyists have a long history of using many popular processors in pairs, CPU vendors typically segregate their product lines, offering chips certified for use in SMP systems (so-called “glueless” multiprocessing support, with cooperative circuitry built into the CPU instead of requiring special bridges and chipsets) as well as uniprocessing desktop and laptop CPUs.

In Intel’s case, this means the Xeon DP (dual processor) and MP (four or more processors) and older Pentium III Xeon, as opposed to standard Pentium 4 and Pentium III, processors. The Xeon is built on the same 0.13-micron process as the current Pentium 4, but at least for now uses the older 400MHz instead of the latest P4s’ 533MHz system bus.

In AMD’s case, it’s the Athlon MP rather than Athlon XP that supports two-way multiprocessing, thanks to what the company calls “Smart MP” technology that combines dual point-to-point, high-speed system buses; so-called “snoop” buses for high-speed communication between CPUs; and optimized cache management to manage data and memory traffic in multiprocessing environments.

With August’s release of the Athlon MP 2200+, AMD moved the MP to the 0.13-micron “Thoroughbred” core of recent Athlon XP chips, yielding higher performance while using 12 percent less power than previous Athlon MPs. In 2003, new entries in the Opteron line of processors will mark AMD’s debut in four- and eight-way SMP.

Will Apple Lead By Playing Catch-Up?

Today’s most prominent push for desktop SMP comes from Apple Computer, which has both a big audience among graphics and multimedia pros and a perceived-performance PR problem — Macs’ PowerPC processors have much lower clock speeds than today’s megahertz- and gigahertz-galore Windows PC processors.

To even the odds (and differentiate its professional Power Mac from consumer iMac desktops), Apple has introduced dual-processor Power Mac G4 systems, starting at $1,699 with two 867MHz PowerPC G4 chips and 256MB of system memory and rising to $3,299 for a dual 1.25GHz machine with 512MB of memory. According to Apple, the latter delivers up to 18 gigaflops of processing power — five times that of the original, single 500MHz G4 model released two years ago.

The twin-engined Power Macs take advantage of the SMP and multithreading features built into Mac OS X — which can, for instance, let the system decode two video clips simultaneously by running one on each processor. Apple has built SMP support into its CPU-intensive video-editing applications, including Final Cut Pro DVD Studio Pro, iMovie 2, and QuickTime.

PC makers usually restrict their dual-processor offerings to CAD, 3D, or financial-modeling workstations like Dell’s Precision 530 ($4,566 with dual 2.4GHz Xeons, dual 80GB hard drives, and 512MB of RDRAM). But enthusiast and gaming suppliers are dabbling with duallies, too; Alienware’s MJ-12 DDR is $3,162 with paired Athlon MP 2100+ processors, 1GB of DDR, two 80GB hard disks, and Nvidia’s GeForce4 Ti 4600 graphics.

Of course, multiprocessing has historically played its biggest role in the server market, where large enterprises routinely rely on two-way, four-way, and larger systems. Today, multiprocessing is even infiltrating the growing market for “blade” servers — systems integrated onto tightly packed boards for more flexible configuration than traditional rack-mounted servers (insert “twin blade” shaving joke here).

Hewlett-Packard has just released a ProLiant BL p-Class two-way blade server aimed at the dynamic Web hosting, media streaming, and terminal-server-farm markets, and reportedly plans a four-way blade server for the first quarter of 2003. Although HP’s choice of Pentium III processors for the ProLiant BL may not seem state of the art, these compact, low-power systems can deliver much more CPU power per square foot than bigger, hotter hardware. And as with all multiprocessing systems, the gigaflops add up quickly.

Categories: Technology