Walker isn't convinced. "We could run Java code in a z9, but it would make it the world's most expensive Java CPU," he says.

Barnett agrees -- partially. "If you have Java or workloads that need high-speed access to mainframe data, running it on a mainframe partition is a viable choice," he says. "But ... for generic Linux or Java workloads, it still isn't an obvious consolidation platform."

IBM is hoping that others will follow MIB Group's example. "IBM is pushing one box, multiple architectures," says Gartner's Vecchio. Guru Rao, an IBM fellow and chief engineer for eServer, says consolidating a three-tiered architecture on the mainframe when data resides there makes sense because communications between the front and back end don't have to go over a latency-prone TCP/IP network. On the mainframe, he says, "you can communicate with each of these spaces using instructions as opposed to TCP traffic."

DiAngelo acknowledges that rewriting applications isn't always practical. "Doing a rip-and-replace is a big thing," he says of the five-year project. "There are things you can't afford to re-engineer, and they will probably always sit in the place where they were developed."

The transition also requires more horsepower for an application that consumes up to 300 I/Os per transaction and up to 130,000 transactions per day. "Java requires more CPU power than assembler, [and] as you move from proprietary VSAM to a generated database system, you lose efficiencies. With WebSphere, MQSeries and DB2, you have to crank the dial up," DiAngelo says.