Honey, I Shrunk the Mainframe
by Thom Daniel Kolton
If you've glanced through a computer magazine recently, you may have been astonished by the number of articles being written about on-line transaction processing. OLTP is really hot. What might not astonish you is that the OLTP systems being developed today are not mainframe-based, but run on Windows and Unix.
This isn't to say that much of what is written is still more hype than substance. Availability and reliability remain acknowledged issues for enterprise deployment, and the next generation OLTP systems are still on the drawing boards back in Redmond and Redwood Shores. But benchmarks of existing systems are quite impressive.
To date, Microsoft's SQL Server 6.5 database management system has achieved 8,000 transactions per minute on a four-processor Pentium Pro server and Oracle has achieved 30,000 transactions per minute on a four-node, 32-processor cluster of Digital Alpha servers ("Thinking Big ", Information Week, April 7, 1997). This calculates to 133 and 500 transactions per second respectively. Pretty impressive results overall, although these benchmarks omit response time averages.
How they achieve these rates is, perhaps, less impressive. The use of parallel processing and partitioned databases is the basis for these numbers. IBM uses the same methods to establish their DB/2 database benchmarks. What this means, of course, is that transactions, in whole or part, must communicate with other databases via a network, raising the importance of the network on a par equal to the transaction processors themselves and creating a potential Achilles heel in the process.
It is tempting to dismiss such an architecture as nothing more than many disparate CPUs that communicate with each other. It is equally tempting to boast about the potential processing capabilities of multiple loosely-coupled TPF systems with databases strategically partitioned to minimize the amount of network volume. However, industry chatter is on Windows NT and Unix, not on any mainframe OS.
In New York two months ago, Bill Gates announced Microsoft's plan for a system to process a billion transactions day. Quick!, the calculator: 1 billion divided by 24 hours divided by 60 minutes divided by 60 seconds = 11,575 transactions a second, on average. Considering that this number is more than a 100% increase over SQL Server's current capabilities, this may not be cause for alarm in the immediate future. However, it must be conceded that Microsoft has the money, the cream of the technical gurus, and the marketing skills to make a significant impact on the OLTP market. And Larry Ellison of Oracle is no slouch either. Together, they represent a formidable competition to the staid competitors.
IBM has taken to calling the mainframe a large "server", but seems confused on how to position it for such a role. Now Messrs. Gates and Ellison are demonstrating that a server can be a small mainframe and are actively and successfully positioning their products as such. They are making inroads into what was once IBM's domain while IBM continues to loose customers to competitors.
A real headache for IBM is that companies are realizing their business doesn't require a high volume transaction processor (HVTP). Many TPF customers generate transactions rates which could be easily supported by smaller, cheaper, and more flexible systems. One V.P. from a large international financial concern privately confessed that their volume is only 200 transactions per second.
The major obstacle for these companies in moving away from TPF is the existing investment of millions of dollars in software, all written coincidentally for a proprietary operating system, which means that programs cannot be ported. Even TPF products soon to be released are built upon proprietary APIs, furthering the investment and, forcibly, the commitment to TPF. Perhaps this is IBM's strategy for customer retention. Perhaps the author is just being cynical. In the previous issue of this magazine, Jerrie Stewart, TPF Product Line Manager, provided some insight into TPF's direction for the future. She discussed the need to offer more standard APIs, better tools, and better connectivity to other platforms, the very same things companies have been demanding from IBM for years. Her statements would be laudable were it not for IBM's long track record of technological advances in TPF. Case in point: by the time ISO-C arrived, the rest of the industry was already coding in C++.
Ms. Stewart candidly talked about how companies are reassessing TPF's role within their organizations and that it is her job to see to it that TPF continues to play a role. That must surely be a difficult and demanding position to find oneself in. Whether Ms. Stewart can stymie the continuing loss of customers to the likes of Oracle will be interesting.
Yet, one wonders whether anyone at IBM has the responsibility to market TPF as a viable operating system to non-TPF customers, or whether IBM has perhaps a secret strategy which precludes selling TPF at all. When virtually no one in the Information Technology industry knows anything about TPF, doesn't that just make you want to scratch your head? Contrast this with Microsoft and Oracle who want to be everywhere.
In the latest rash of OLTP articles, IBM had plenty of opportunity to respond, clarify and make corrections to obvious factual errors. They were mute. But I could hear Bill Gates deliver his "One Billion Transactions a Day" speech all the way from New York. Perhaps by not telling us something, IBM is trying to tell us something.
Thom Kolton is the President of Mid-Atlantic Systems Design Corp. and developer of the Greyhound RDBMS.