JP Morgan Goes Grid

As part of an ambitious plan to slash data center costs and boost performance, J.P. Morgan Chase & Co. has replaced a supercomputer and a mixture of high-end servers with a 4,000-processor grid built from low-cost servers.

Speaking at the High Performance on Wall Street conference in New York today, Adrian Kunzle, J.P Morgans managing director of architecture, explained that the grid, which links 10 data centers across three continents, now forms the bank’s technology backbone. “We have broken the silo mentality,” he said. “We’re starting to get an organization that’s more accustomed to sharing.”

By 2007, predicts Kunzle, J.P Morgan will be able to run a credit risk scenario in 30 minutes that would have previously taken eight hours.

The bank implemented the grid three years ago in an attempt to consolidate a complex mix of technologies. At that time, J.P. Morgan relied on a supercomputer to provide much of its number-crunching, as well as a hodgepodge of high-end servers from four different vendors. To make matters even more complicated, J.P Morgan also employed four separate operating systems.

J.P. Morgan was not getting what it needed from its systems. “We were getting somewhere near 30 percent utilization,” Kunzle said. This led the bank to standardize “on cheap hardware” from just three vendors and a single operating system. As a result, said Kunzle, J.P. Morgan can now tap into 70 percent of its overall data center capacity.Although he would not reveal which vendors are involved in all this, Kunzle did confirm that the bank has replaced expensive SMP (symmetric multiprocessing) enterprise servers with low-cost, 1U servers. This, he said, has brought about a “tenfold decrease in costs,” despite the fact that J.P Morgan has actually increased its overall capacity from 1,000 to 4,000 processors.

J.P. Morgan is not the only firm going down this route. A number of organizations are currently eschewing expensive monolithic systems to build high-powered clusters and grids from standard pieces of hardware. One of them is Intel Corp. (Nasdaq: INTC). Michael Shevlin, the vendor’s director of business development, said today that Intel has already implemented a 70,000 processor grid to support its own business processes. (See Pittsburgh Picks Big Ben, Luebeck Looks to Clusters and Statoil Builds Dell Cluster.)

At least one analyst thinks grid's the way to go. “It’s much cheaper for them to buy the 1U boxes and just link them all up,” Bob Cohen, analyst at the Economic Strategy Institute, tells NDCF.

Kunzle added that the shift to standard low-cost 1U servers has also made his life a lot easier. ”If a 1U server fails and you have a job running across 40 [of them] then it doesn’t really matter."

J.P. Morgan has also turned to InfiniBand as the interconnect between the separate processors, or nodes. This, according to Kunzle, is an ongoing project, and he feels InfiniBand fits in nicely with the bank’s drive to push down costs. The technology, he says, has, “a price point only marginally higher than Gigabit Ethernet, but much less than 10-Gigabit Ethernet.”The director has also been won over by InfiniBand’s performance. “It is operating at over 10 times the bandwidth of Gigabit Ethernet. It is also much lower latency.”

Kunzle also has some advice for other organizations looking to win internal users over to grids. “We went out to look for triggers in the organization, pain points, groups that were having problems, and used them to sell the grid."

But he warns that setting up strong service-level agreements with these users is key to getting them onboard. “If you’re taking someone’s CPU away from them for half the day, you have to be able to guarantee that they can get their work done in the other half of the day.”

— James Rogers, Site Editor, Next-Gen Data Center Forum