By some estimates, the total value of the applications residing on mainframes today exceeds US$1 trillion. Most of that code was written over the past 40 years in Cobol, with some assembler, PL/1 and 4GL thrown into the mix. Unfortunately, those programs don't play well with today's distributed systems, and the amount of legacy code at companies such as Sabre Holdings in Texas, makes a rewrite a huge undertaking. "We're bound by our software and its lack of portability," Sabre Vice President Alan Walker says of the 40,000 programs still running on IBM Transaction Processing Facility (TPF), Agilent Modular Power System and other mainframe systems.

With a shortage of Cobol programming talent looming in the next decade and a clear need for greater software agility and lower operating costs, IT organizations have begun to make transition plans for mainframe applications. The trick lies in figuring out which applications to modernize, how to do it and where they should reside.

Applications fall into one of three groups based on scale, says Dale Vecchio, an analyst at Gartner. Applications under 500 MIPS are migrating to distributed systems. "These guys, they want off," Vecchio says. As organizations begin peeling away smaller applications, they may move to a packaged application; port the application to Unix, Linux or Windows; or, in some cases, rewrite the applications to run in a .Net or Java environment, he says.

In the 1,000-MIPS-and-up arena, the mainframe is still the preferred platform. Applications between 500 and 1,000 MIPS fall into a gray area where the best alternative is less clear. An increasingly common strategy for these applications is to leave the Cobol in place while using a service-oriented architecture (SOA) to expose key interfaces that insulate developers from the code.

"If you expose those applications as a Web service, it's irrelevant what that application was written in," says Ian Archbell, vice president of product management at tool vendor Micro Focus International. "SOA is just a set of interfaces, an abstraction."

"SOA at least allows you to break the dependency bonds," says Ron Schmelzer, an analyst at ZapThink.

Cobol isn't going away, but it's also not moving forward. While the Cobol code base on mainframes is projected to increase by 3 percent to 5 percent a year, that's mostly a byproduct of maintenance, says Gary Barnett, an analyst at Ovum.

"No one is learning [Cobol] in school anymore, and new applications aren't being built in Cobol anymore," says Schmelzer. "Cobol is like Latin."

Vendors such as Micro Focus have abandoned the idea of evolving the Cobol language for distributed application development. "Micro Focus is not about a better Cobol compiler," says Archbell. Instead, its approach is to "embrace and extend," he says. "We expose things like aggregated CICS transactions as JavaBeans, Web services, or .Net or C# code. It's wrappering." But with so much legacy code, that process won't take place overnight. "It could take 20 years," Archbell says.

Sabre still has more than 10,000 MIPS of applications on mainframes, and Walker plans to migrate everything off over the next few years. The company's TPF-based fare- searching application, used by Travel-ocity.com LP and travel agents, has been rewritten to run as a 64-bit Linux program on four-way Opteron servers.

Sabre migrated the back-end data to 45 servers running M ySQL that each contain fully replicated data. The new system is more flexible and "pretty cheap" compared with the mainframe, Walker says. He questions the conventional wisdom that all high-end applications need to stay on mainframes, noting that the search application was in the thousands of MIPS. "It's pretty obvious that you don't need mainframes to do large-scale transactions," he says, pointing to the successes of eBay and Amazon.com.