Despite Different Approaches, Mainframes Still All-Powerful : The Cutting Edge: Computing / Technology / Innovation
One of the enduring notions in the world of personal computing is that the desktop (and laptop) computers of today are as powerful as the mainframe computers of just a few years ago.
But is that really true? And how do personal computers stack up against today’s mainframes?
For answers, I called upon Jack Eberth, program manager for IBM’s large-scale computing division in Fishkill, N.Y. His unit has announced a new generation of mainframe hardware and software--additions to its ES/9000 family of mainframe computers and S-390 system architecture, which is a combination of hardware and software products.
IBM is the world’s greatest mainframe proponent, of course, but it’s also betting heavily on microchips for its future, making it the only company that truly covers the computing gamut, from biggest to smallest.
As it turns out, it’s not very easy to compare personal computers and mainframes, because the two take more or less opposite approaches to computing.
For instance, stating the processing speed differences in millions of computer instructions executed per second (MIPS) isn’t a very accurate comparison because the actual work accomplished by an instruction on one kind of system differs from that accomplished on another.
But for what they’re worth, the numbers are dramatic. The fastest Intel Corp. Pentium-based PCs these days can execute about 166 MIPS, and the fastest IBM/Motorola/Apple PowerPC 601 chips can do 300 MIPS.
The highest-performance IBM system can execute 15 billion instructions per second by simultaneously distributing the computing load among 32 of its fastest ES/9000 9X2 mainframes. A single ES/9000 9X2 runs at about 400 MIPS and costs $15 million to $20 million.
One should not believe, however, that a $5,000 Pentium PC is only 60% slower than a $20-million IBM mainframe, nor that a PowerPC-equipped Macintosh can run three-fourths as fast as the mainframe.
One difference lies in keeping the processors actually supplied with the maximum amount of work they can do every second of the day and night.
A single user, even one doing the most intensive kinds of graphics manipulations, doesn’t keep a processor supplied with millions upon millions of instructions to execute every second of the day.
That’s why mainframes were designed from the beginning to simultaneously serve lots of people, each with a relative handful of requests to process.
*
The design philosophy behind a mainframe computer, as Eberth explained it, is to keep its very expensive central processing unit, or engine, full of data so that few of its valuable nanoseconds are wasted with nothing to do.
That’s contrary to the design of a personal computer, where a relatively cheap central processor is placed at the disposal of one person. And if it sits there doing nothing while that person mulls what to say next in a memo to the boss, so what?
Because personal computing is, well, personal, performance begins to bog down as soon as you start linking personal computers together to try to make PCs and Macs do jobs that are simple for mainframes, such as hosting a central database.
*
Mainframes do those things much more robustly than the best PC local area network because they are designed from the beginning to do so.
For instance, the big IBM mainframes have 256 channels for data to flow in and out of, and it races along at 17 megabytes per second. That kind of capacity can be carved up in many ways before it gets slowed down to the 1 to 3 megabits per second (remember, 8 bits equals one byte) typical of a PC local area network. One channel on a big mainframe is about 45 times faster than the PC LAN.
In a PC, you may have 4 or 8 or 16 megabytes of random access memory where programs and data are stored temporarily while they await action by the central processing chip.
In the big IBM mainframe there are two kinds of memory. The most expensive, called central storage, contains up to 2 gigabytes (2,000 megabytes, in PC terms) of ultra-fast microprocessor chips dedicated to memory tasks. There can be up to another 8 gigabytes of expanded storage memory. It runs just 3% slower than the fast stuff but costs only half as much.
The mainframe really is a main frame, a big steel box that houses the components that make up the computer. The largest IBM model requires 251 square feet of floor space and weighs about 16 tons.
The fastest models don’t have microprocessors like the ones we’re used to in the personal computer world. Instead, the actual computing is done by hundreds of bipolar chips, each 63 layers thick, of metal circuits lithographically etched onto thin slabs of glass ceramic material. Because the circuits are much closer together than those of even the most densely packed microprocessor, they run much faster. And they generate a lot more heat.
These top-of-the-line computers are water-cooled, but the water doesn’t get near the chips. Heat is transferred from the chips to helium gas that passes very close to each chip in an intricate plumbing system that won’t hurt the electronics, or people, if it should leak. Water circulates through radiator-like devices a safe distance away to cool the helium.
The high-powered bipolar chip is nearing the end of its economic life after serving the mainframe industry for years. Because of its multilayered construction, adding more performance requires adding more layers, which raises the price exponentially.
The CMOS (complementary metal oxide semiconductor) microprocessor, which can be mass-produced more easily and runs cooler, although more slowly, is poised to take over. IBM offers some models of the S-390 family with CMOS microprocessors, which contain 6 million transistors. It is actually a package of four separate chips grouped together on a circular connector with a heat dissipator on top. Among the new S-390 systems are a couple that can house up to 192 of these microprocessors running in parallel.
In comparison, IBM’s PowerPC 601 chip, which it shares with Motorola and Apple, has 2.8 million transistors. Intel’s Pentium chip has 3.1 million. Both also are CMOS technology and are packaged as single chips. (Strictly speaking, the Pentium chip is BiCMOS because about 10% of its circuits are bipolar, added by Intel to boost performance. Also, the PowerPC is a RISC chip because it uses reduced-instruction set computing, compared to the complex instruction sets built into the mainframe and Pentium, known as CISC chips.)
Another critical requirement for feeding data to hungry mainframe processors is enough disk storage to contain the huge databases required in banking, ticket, credit card and financial trading systems.
*
In the personal computer world, we call it hard disk storage. IBM calls it DASD (direct access storage device) and it’s like PC hard disks, only larger. Currently the disk platters are 10.8 inches in diameter, but it’s expected that 3.5-inch units--the same diameter found in most PCs and Macs--will become standard this summer. However, mainframe DASD has a lot more capacity than we’re used to in the PC world. Typically it comes packaged in 50-gigabyte units consisting of four physical hard drives and a controller.
What does all this power add up to? According to a study last year, Eberth said, the typical large IBM mainframe installation has 300 gigabytes of data storage and handles millions of data access requests per day. But if your PC is beginning to sound puny, remember, it can be connected to a mainframe. The days of dumb terminals are fading. PC LANs now are the preferred way to connect to a mainframe, giving you the best of both kinds of computing.
Computer File welcomes your comments but regrets that the author cannot respond individually. Write to Richard O’Reilly, Computer File, Los Angeles Times, Times Mirror Square, Los Angeles, CA 90053, or message oreilly@latimes.com on the Internet.
