Allied Irish Bank (AIB) has found that vector processing, more commonly associated with highly computational scientific tasks such as weather forecasting or designing drug molecules, can blow a hole through its management information systems workload.
The AIB treasury processing system, Opics, processes 120,000 transactions every night, based on more than 3,500 trades a day, which generates about 360,000 financial transactions and updates up to 33 million trading balances in the bank's general ledger. This data is then used as the source to create the myriad management information system reports required by users.
AIB's problem was the volume of transactions passing through overnight that created acute pressure on the amount of time available to analyse the transaction data and generate the multiplicity of reports required for the morning.
"There was tremendous pressure on us," says Padraic Hamrock, head of IT services at AIB Capital Markets. "We were in the middle of the project and we would run out of our batch window." And the problem was only going to intensify as more data feeds were being tapped from other systems to generate an increasingly diverse range of reports, in less and less time.
"The integration of our money markets' business transactions increased our source for financial and management information data by 300%," says Hamrock. "We had to scale back on the reports because we were adding more and more volumes [of data] at the front end."
He faced tough choices - accepting that only a limited number of reports could be produced in time, which meant that service to users would deteriorate, or he had to fix the problem. "Our immediate thought was how to rearchitect the system. That would have been the typical approach."
It was at that point - facing a possible radical redesign of databases and process flows that an alternative was suggested. The bank could simply press a "go-faster" button - courtesy of vector processing. The proposal was to use a vector database, Metasoft, from Aleri, a company founded by former Cray supercomputer staff, to crunch through the vast quantities of data required to produce the multidimensional reports, at extremely high speed.
Hamrock's reaction was initially sceptical.
"When they told us they could do four hours' processing in four minutes we laughed and said, 'Sure'," he recalls.
Nevertheless he gave Aleri a chance to make good their claim. "They had a good understanding of our problem and what we had to achieve, so we just stood back and allowed them to demonstrate their software. We were convinced - it worked immediately," he says.
"It amazed me - and it still does - to slash the processing time so dramatically. You read about it but you don't believe it."
Adopting technology more commonly associated with scientific supercomputing has worked well for AIB for a number of reasons, believes Hamrock.
A key issue has been that despite the massive speed-up in management information systems data analysis, the vector software takes its data straight out of the Sybase database that holds the general ledger, and runs on a 2x64 Alpha computer under Unix. "We were running it on NT until we got the Alpha version," says Hamrock.
Nor did vector processing require Hamrock's staff learn extra skills. "We don't need any specialist skills and it's had no impact on our database administration or maintenance," he says. "It is a 'black box' approach."
That there is vector processing going on inside the box does not concern him. "We don't know how it works and we don't need to know," he says.
To do so, he believes, would be a waste of our time. "We just need to know what the business benefits are and that we can proceed without becoming experts in the underlying technology," he says.
Even his technical staff are "very comfortable" not knowing about vector processing, providing it does the job. AIB already has a buy-not-build IT culture, and highly specialised software applications for complex trading instruments are the norm in capital markets IT.
AIB was willing to consider something as unusual as vector processing for data analysis because of the urgency of the problem, says Hamrock.
Redesigning the system would, he says "have taken another 30 to 40 people and added another nine months to the project, and user reaction to that would have been very poor."
The implementation cycle for the vector option took about six months. "We did no requests for information or supplier reviews, but we did a lot of talking with Aleri initially in the autumn, engaged properly in February, and deployed the software in April," says Hamrock.
Maintaining the existing report generation system in parallel "made us comfortable about deploying [the vector software] in full," he says.
Although Hamrock acknowledges that, as an early adopter of the Aleri technology "it didn't cost us a huge amount because we took a risk on it," opting for vector processing is by far the most cost-effective method of solving the problem, he says. "The cost of rearchitecting and development work would probably have been 10 to 20 times as much."
But however urgent the problem to be fixed, Hamrock made sure he did not risk becoming over-dependent on novel technology adopted into corporate IT from the world of scientific computing.
"Although the product is very stable - we installed it over two years ago and it's still there, we are not critically dependent on it," he says.
"You have to be in a situation where you could switch it off at any time. We could still deliver the essential reports if it was not available - we always have that fallback."
Bringing science into computing
In general, scientific and business computing run on different tracks. "There is not much cross-fertilisation," says industry analyst Phil Payne of Isham Research. "Very often scientific machines are architechted for particular scientific problems."
Xephon analyst Trevor Eddolles agrees. "There hasn't been much of a Nasa-Teflon effect," he points out.
As with the Allied Irish Bank instance of applying vector processing to multi-dimensional data analysis for data warehousing purposes, Mike Thompson of the Butler Group says the trade-off of using special-purpose technology to gain speed is at the expense of flexibility in the system.
He also says that apart from a few niche areas, such as using fuzzy logic for fraud detection, there are few scientific computing technologies that have a business use.
However, technology such as parallel processing and reduced instruction set computing, originally introduced to speed up solving scientific problems, has now become commonplace in business computing as, perhaps, will grid-computing (sharing out the unused cycles on individual computers). The most compelling recent example of business take-up from the academic world has to be the Internet, which changed the rules of the global economy.