Azul's Stephen DeWitt remakes the J2EE server

IDG News Service - Stephen DeWitt is a remarkable salesman. Not only was he able to talk Silicon Valley investors into sinking money into a hardware company during the darkest days of the dot-com implosion, but he's also the guy who convinced Sun Microsystems Inc. CEO Scott McNealy to wear a penguin suit.

DeWitt remembers a few anxious moments before McNealy's February 2002 speech at Sun's financial analyst conference in San Francisco. Sun's CEO addressed the analysts dressed as Tux, the Linux mascot, in order to show Sun's support of the open-source operating system. As DeWitt remembers it, McNealy turned to him just before heading onstage, penguin head in hand, and asked, 'Are you sure I should do this?'"

DeWitt's response? "Put the head on and waddle out there."

It wasn't the first time McNealy had listened to DeWitt. Two years earlier, Sun had purchased DeWitt's company, Cobalt Networks Inc., for $2 billion.

But just months after the penguin suit incident, DeWitt had left the company, and he found himself making the rounds along Sand Hill Road in Menlo Park, Calif., trying to stir up interest in a new venture, Azul Systems Inc. Again, DeWitt proved to be a compelling salesman, and today Azul's first systems are being tested by beta customers. The Mountain View, Calif.-based company is on track to launch its product by the end of March, according to DeWitt.

With a custom-designed low-power processor that has 24 processing units, called cores, Azul has built a system that will change the way data centers run Java 2 Enterprise Edition applications, according to DeWitt. And, if negotiations with Microsoft Corp. pan out, .Net could be next.

Following is an edited transcript of an interview with DeWitt.

What convinced you that Azul needed to design not only its own systems, but its own processors as well? There are a number of reasons why we didn't use off-the-shelf. [Existing server microprocessors] are designed to minimize single-threaded latency. Period. They are designed to get an instruction and execute it as quickly as you can, because all of those microprocessor architectures were born at a time when development environments, binaries, operating systems and all of the programming languages that drove application development exposed instructions in a serialized fashion down to the hardware.

Now the developer world has moved to a multithreaded parallel world, and that whole multithreaded parallel world sits on top of the single-threaded architecture and goes, "I can't scale." So how does the industry deal with that? "Let's add another processor on a chip." So