Morgan Stanley has come a long way with its use of blade servers since Jeffrey M. Birnbaum (now Merrill Lynch’s chief technology architect) pioneered their use at the firm. [For non-blade-users, here’s a nice definition from whatis.com: A blade server is a server chassis housing multiple thin, modular electronic circuit boards, known as server blades. Each blade is a server in its own right, often dedicated to a single application. The blades are literally servers on a card, containing processors, memory, integrated network controllers, an optional fiber channel host bus adaptor and other input/output ports.] Morgan Stanley is running 30,000 blade servers and it buys 10,000 to 20,000 more each year. Over the past two years, 95% of the firm’s installed server base has been blades.

“For us, blades have always been a net efficiency play,” Russell says. “People talk about being out of space in the data centers. Rarely are they out of space, they’re usually out of power and cooling.”

The firm uses “pygmy racks” that hold up to three groups of blades per rack. The blades draw 150 watts each, whereas a standard rack-mounted server draws 350-400 watts, providing an 85% improvement in efficiency. “Nothing saves cost in data centers like not using much power and cooling in the first place,” Russell says.

A side benefit to using blades, he adds, is that the failure rates of blades are half and in some cases a third that of standard servers.

Like many Wall Street firms, Morgan Stanley runs virtual machines on its blades. But where conventional wisdom indicates that virtualization drives server utilization up and server count down, Russell says this has not been his experience. “One thing we find is, if you make something appear free, you have infinite demand,” he says. “We’re giving virtual machines to some people that we would never authorize to get servers, but those people are doing useful things that we never anticipated, so it’s an enabler of business and operational change.”

The biggest IT challenge to using blades, Russell says, is the rapid proliferation of network switches. “In two years, we’ll have 10,000 network switches” unless something changes. The network switches are also the biggest cost to blade computing, he says. The firm is looking at a hub-and-spoke architecture that might help reduce network cost. [We recently spoke with a company called Teak Technologies that offers rack- and row-level switches, for example.]

In the future, Russell would like to see vendors come out with larger blade racks and rack-level management software. He would also like to see blades that are less reliant on specific processors. “There’s a battle in the processor world,” he notes. “One of the downsides of blades is that you have to place bets. If you bet that AMD’s 4-way processor will win in 2008, then you’re susceptible if AMD slips behind a year.”