October 09, 2009, 10:25 PM — On the eve of Oracle Corp.'s OpenWorld user conference, IBM today unveiled a new clustering feature that it says contends will help its flagship DB2 database trump Oracle Corp.'s rival product on scalability, speed and price.
IBM says the new pureScale feature will enable companies to "scale out" their DB2 clusters without paying an enormous performance cost. IBM says Oracle's long-running Real Application Clusters (RAC) technology does significantly affect performance.
"This is an Exadata-killer, in the sense that it is much more economical and scalable," said Bernie Spang, director of product strategy for IBM's information management division, in an interview.
Merv Adrian, an analyst with IT Strategy, calls the pureScale announcement a "well-timed shot across the bow."
Oracle hasn't shown off many Exadata customer wins since its launch a year ago, said Adrian, in part because it had to quickly pull Hewlett-Packard Co. servers out of Exadata in favor of hardware from, Sun Microsystems Inc., which it has agreed to acquire.
"The likelihood is that it will be a head-to-head battle" between Oracle and IBM, he said. "We should see some shots fired back and forth as the two firms jockey for position."
IBM did not disclose the price of pureScale, which calls for customers to deploy Power-based servers in a cluster using high-speed Infiniband networking technology. One or two servers in the pureScale cluster serve as "traffic cops" that control the grid and make it appear as one node to all applications connected to it, said Spang.
The traffic cop servers are very efficient, says IBM, because of a technology called Remote Direct Memory Access (RDMA), which lets DB2 database servers directly query the cache memory of the pureScale traffic control servers as it were on their own motherboard, going around the pureScale server's CPU in the process, said Handy.
According to IBM tests, that allows 64-node pureScale-based clusters to waste less than 10% of their processing power on grid-related overhead. That increases to about 20% for 100-node clusters.