IBM supers shun nukes for biz analytics

BAO. It's the next ERP

BAO times two

The two systems that IBM trotted out this week to make its point about BAO were not data warehouses or search engines in the traditional sense of those products, but they are indicative of the cross-group development effort that IBM is talking about when it discusses BAO.

The first is the Systems S streaming server, which we first told you about here when it went into prototype at Toronto Dominion Bank in early April as the back-end for an options trading system. System S takes a BlueGene/P massively parallel PowerPC-based supercomputer and slaps on some data streaming and processing middleware called InfoSphere Streams continuously searches through data feeds of all kinds (news feeds, stock tickers, video off TV and the Internet, it doesn't matter what) and continuously updates a predefined set of queries setup in the InfoSphere database to give you up-to-the-second information to make decisions.

TD Securities, the options trading arm of Toronto Dominion Bank, took the first prototype of the System S machine and used it as a front-end to its options trading system and showed that it could process 21 times more information than a prior system used by the bank and that the features in the System S software allowed the options trading system to crunch 5 million options valuations per second, which IBM said was 20 times faster than the world record up to that point.

At the investor day briefing, IBM said that it would be opening a European stream computing center in Dublin to help customers figure out how to use the InfoSphere Streams software running on BlueGene/P iron and said further that it was making the InfoSphere Streams code - which manages data feeds and chews on them - available for free to customers on a trial basis (and for an unspecified length of time) so they can play with it.

Presumably, you have to buy a BlueGene/P super to play. The BlueGene/P crams 1,024 of IBM's 850 MHz single-core PowerPC 450 chips into a rack, with a board comprised of four chips and 2 GB of DDR2 main memory shared by an SMP bus. A BlueGene/P rack has about 13.9 teraflops of number-crunching performance and costs about $1.3m. But IBM says that InfoSphere Streams can run on other boxes. (The prototype at TD Securities runs on a BlueGene/P with Fedora 8 Linux from Red Hat).

Governments - and particularly their law enforcement and public safety groups - are very keen on the capabilities that IBM says the System S setup has, as well as retailers, healthcare providers, financial institutions, and transportation companies.

The other machine that is going to weigh in heavily on this BAO expansion is the "Watson" Question and Answer engine, another offshoot of the BlueGene/P that was announced several weeks ago and that IBM is programming to take on the best human players of the game show Jeopardy next year.

John Kelly, who heads up IBM Research, described the Watson machine to the Wall Streeters and investors. He knew that there has been plenty of skepticism about the project, despite the fact that IBM did create Deep Blue and it did beat champion Gary Kasparov in chess matchups when IBM first tried to demonstrate its prowess with massively parallel computing back in the late 1990s. Making a machine called Watson to play Jeopardy against real people is a PR stunt, but IBM wouldn't do it unless it thought it could win and - more importantly - unless it thought the PR stunt was going to drive business in some way in the future.

The Watson has to hear a Jeopardy statement, sift through databases for what it might refer to, and rank possible questions that match that statement, pick an question or decide not to try to say a question, and do it all in under three seconds.

Kelly blew off the skeptics about Watson. "It is not a pipe dream, it is up and running," he said. "It beats me, and it will probably beat most of you in this room."

He went on to say that IBM had every intention of turning around the QA system at the heart of Watson and making it into a business decision engine, helping call centers cope with customers, doctors diagnose diseases, and businesses and governments of all kinds and sizes make decisions. And that is why IBM Research has over 150 mathematicians hammering away on BAO algorithms that can be spliced onto machines like System S and Watson for real-world customers.

It all sounds pretty far fetched, right? But whole layers of middle management are going to be made redundant if what IBM says is possible turns out to be doable. Perhaps IBM should have called it business optimization and analytics to get the correct boa constrictor image for the people currently in charge of making calls and sifting through data in the middles of organizations the world over. ®