IBM S/390 General Manager

Letter from my desk...

May, 1997

Mythic Times

A myth, says the dictionary, is "a traditional, typically ancient story; a fiction, or a half-truth." I don't know how myths get started, but my guess is they've been around as long as people have gathered around campfires. Or water coolers. And some perfectly reasonable people believe them.

This month, I'd like to explode a few myths about S/390 CMOS servers and Parallel Sysplex technology. One myth has it that adopting Parallel Sysplex is risky. Another says it's too complex. And another says it's too expensive when compared to a more powerful single system.

First, let's talk about risk: Fact is, S/390 Parallel Sysplex architecture reduces risk. That's because workloads can be automatically re-routed to other available systems. Staying with a single system -- no matter how powerful -- is the riskier path because work flow simply stops during any kind of outage. And in an age of 24x7 network computing, even planned outages (much less unplanned) are unacceptable. So the risk lies not in going to Parallel Sysplex, but in not going there.

What's more, a single-system image can get only so big before it develops its own set of problems. It is, after all, IBM operating-system software running on all those plug-compatible systems, and we designed the architecture. So we know that at present there are technical limitations to our operating-system software running on large single images of over, say, 500 to 600 MIPs. This is why, back in 1994, we decided to support horizontal parallel growth.

Second, the myth about complexity: Parallel Sysplex requires more time and effort than simply installing another single system. That much is true. But with careful planning, it's a straightforward operation, and, once installed, the result is less complexity, not more.

Several S/390 customers have made the migration to Parallel Sysplex in less than four months, including Zeiss, Rite-Aid, and Burlington Northern Santa Fe Railroad. And the migrations continue apace: As of the end of 1Q 1997, there were 1,461 systems enabled for Parallel Sysplex at 857 customer sites, 457 of which were in production mode. They include one of the country's largest banks, a major insurance company, and several large retail chains. And very few have had to alter their applications in going to clustering.

Once customers have clustered CMOS systems, it's much easier to add more servers. As an executive at Rite-Aid put it: "One of our Parallel Sysplex expansions -- to enable an acquisition of 1,000 stores -- took only about as long as it took IBM to get the hardware through the door, and it added 50% in workload capacity."

Third, the myth about expense: It's because of the expense of bipolar, actually, that in the long run everyone at the high end will eventually move to CMOS and Parallel Sysplex. A CMOS Parallel Sysplex configuration needs only about 20% of the physical space, runs approximately 75% cooler, and is 80% more energy efficient than a bipolar uniprocessor of comparable performance. Besides that, the labor savings, in a continuous computing environment, can be enormous. The Yankee Group estimates that, for backup and recovery alone, the savings can be several hundred thousand dollars annually.

With bipolar technology, it was simply costing too much and taking too long to deliver what were fairly small increases in performance. This meant it could no longer offer S/390 customers the long-term performance improvements they needed at a competitive price. That's why we made the switch to CMOS. With CMOS, we've been shipping three times the performance in only two years' time as opposed to the five years that it typically took with bipolar.

What's also interesting is that very few customers, it turns out, need more processing power than CMOS delivers. The Meta Consulting Group recently reported that G3 servers have the capacity to replace more than 90 percent of existing S/390 workloads. This is consistent with the hundreds of customer workloads we have analyzed: Very few need the uniprocessor size of the old bipolar technology.

At the end of the day, customers are going to Parallel Sysplex for some very non-mythic reasons: continuous computing, performance, and capacity. Continuous computing because a single server is inherently risky, no matter how many MIPs it is. What happens when it goes down? How do you handle planned outages?

In a year's time, PC LANs are down an average of 600 hours. UNIX systems are down an average of 80 hours. Parallel Sysplex systems are down an average of five minutes.

As for performance, last October, using DB2's Sysplex query parallelism function on a system composed of two 10-way G3s, our Teraplex Center took just 50 minutes to scan the world's largest single DB2 table -- 6 billion rows of data, weighing in at 750 gigabtes. Without Parallel Sysplex, that would have taken eight hours.

In terms of capacity, no single-system image will be able to handle the tidal wave of transactions coming down the pike in the form of heavy-duty electronic commerce. Parallel Sysplex, on the other hand, provides just-in-time capacity. You add more when you need more, in incremental steps.

Finally, there are also the applications reasons. By the end of this year, for example, IBM will have all its database and on-line transaction processing environments enabled for Parallel Sysplex, including CICS/V SAM, DB2, and IMS.

Maybe ancient peoples developed myths because nothing else could explain the mysteries they saw around them. Or maybe they just liked telling stories!

But I want people to know the whole story about IBM CMOS Servers and Parallel Sysplex. Let us know whether you think we've succeeded.

Linda Sanford