When Vern Brownell left his job as chief technology officer at Goldman Sachs & Co. in 2000, he had a fairly modest goal: Start a company that would manufacture servers that reduced what he saw as the “incredible” complexity of most data centers.
“You can look at it by any number of different metrics,” Brownell says. “One can be the number of components you need for an application server. Most data centers have approached the explosion point in terms of the number of servers and the amount of connectivity and components. People are looking for solutions to reduce that.”
The company he founded, Egenera of Marlboro, Mass., unveiled its first Blade-Frame server in 2001. The blade format reduced the cables, ports and host bus adapters a server uses by “an order of magnitude,” Brownell explains. He cites the example of one large commercial bank that used blades to reduce the number of components supporting its Web servers to 400 from 3,000.
In the six years since Egenera rolled out its first server, computer makers ranging from IBM Corp. and Hewlett-Packard Co. to niche start-ups such as Rackable Systems and Appro, both of Milpitas, Calif., have joined Egenera in the blade market. The systems are being deployed more and more on trading
floors and other compute-intensive environments. “Our focus is really that mission-critical space,” Brownell says. “You’ll see bulge-bracket firms using this technology for their core trading, some of their most important transactional business.”
Much more than a wish for simplicity prompted the adoption of blade servers. Wall Street is also seeking to cut its electric bill.
“It sounds very odd that power and cooling would drive the behavior of these large investment banks, but they are out of power, they are out of space, and blades are far more power-efficient than anything else out there,” says
Scott Tease, worldwide marketing manager for the BladeCenter server line at
IBM Corp.
Indeed, blade servers’ popularity has grown as commercial electric rates have climbed. Many companies are not only looking to save on electricity costs, but they are also seeking opportunities to boast of their “green” environmental initiatives. Says Tease, “It won’t be long before we start looking at every light bulb and every monitor and trying to drive out energy consumption wherever possible.”
The adoption of blade hardware and the search for efficiency in the data center go hand in hand with virtualization, in which a single piece of hardware performs as multiple “virtual machines.” This technology dates back to the 1960s and was common on Unix systems for years, but its current popularity stems from Palo Alto, Calif.-based VMware’s introduction of virtualization software for Intel Corp.’s x86 architecture in 2001. The company was acquired in 2004 by Hopkinton, Mass.-based data storage equipment maker EMC Corp. Since the introduction of VMware’s flagship product, Infrastructure, its ac-
ceptance on Wall Street
has been fairly swift, says
Bogomil Balkansky, the
firm’s director of product
marketing. By the end of
2006, most bulge-bracket
firms had installed it.
“Virtualization has two
primary value proposi-
tions,” Balkansky says.
“One is server consolida-
Vern Brownell tion—you can put multi-
ple applications on the
same server.” The higher utilization rates
have permitted some customers to consol-
idate up to 30 applications on one server.
The second benefit stems from the
ability to share large workloads across
multiple servers, he says. Virtualization
creates what Balkansky calls a “layer of abstraction between the hardware and software,” which permits large tasks to hop from server to server. “On Wall Street,” he says, “customers are more attracted by this better resource management.”
In Tease’s view, virtualization and blades go together “like peas and carrots.” The blades provide a suitable hardware platform for sharing multiple workloads using virtualization, and with the reduction in the components used to connect the blades, jobs can be quickly transferred between multiple pieces of hardware.
Virtualization is also helping IT executives address one of the shortcomings of modern computing: that most PCs and servers sit idle or are not being taxed to capacity most of the time. The consolidation of jobs onto servers and the sharing of workloads among multiple servers allow firms to increase the percentage of a system’s processing they actually tap. “It is far more energy-efficient to run one server at peak utilization than ten servers at 6 percent utilization,” says Tease.
According to Brownell, server utilization was one of the most important issues he sought to address while at Goldman during the 1990s. “Despite the amount that I spent on servers, which was
Continued on page 21
on an enterprise grid by the end of March, says spokesperson Selena Morris. The grid will use servers in disaster recovery sites and data centers, which would normally not be available for business use.
Songnian Zhou, CEO of Platform Computing, says that there are four stages of grid computing, and the most advanced financial institutions are in the third. In the first stage, a single application runs on a cluster of servers. This can dramatically lower costs since it saves the expense of buying new servers, but it does increase the complexity of management. Stage two involves line-of-business grid sharing, with multiple applications on the same cluster. Utilization rates jump in this phase, but
complexity again increases.
In the third stage, Zhou says, the grid spans multiple lines of business and is operated by central IT instead of by individual businesses. At this stage, seen at the likes of Citigroup, JP Morgan Chase & Co. and Lehman Brothers, “you bring in the pros” who can test the grid, validate policies and troubleshoot at a level that is far too complicated to be handled at the business-line level, Zhou says. Ultimately, these firms want to extend grids and their benefits beyond the “firewall” to third-party business partners. He sees this process beginning in about a year.
Thanos Mitsolides, SVP for fixed income and analytics at Lehman, says that in the near term he expects more groups at his firm to move systems onto Platform’s
flagship Symphony system, and in the long term he hopes to see everything on a single grid. “In parallel with this,” he says, “we want users to stop thinking that they own the hardware, but instead that they’re buying CPU time”—the essence of utility computing.
According to Mitsolides, grids require cultural change, because business units must adjust to sharing resources that they traditionally owned. The units are “ comfortable with what they have now ... and being in full control of the machines,” says Mitsolides. “Eventually they should realize they don’t need to be the owner of it all, but that will take some time. Sharing will provide businesses with the flexibility to easily expand their compute power in emergencies and will reduce the need to keep around expensive spare capacity.”
In response to concerns around own-
ership, Platform rolled out a new version of Symphony in March, saying it will address fears that claims on IT resources will be usurped by the grid. It overcomes managers’ concerns that capacity won’t be available when they need it by allowing business groups to borrow server cycles while respecting an owner’s right of first refusal. Symphony also assigns tasks to dedicated resources instead of borrowed hardware, and individual departments can determine how they use shared resources.
Eventually, Mitsolides sees an opportunity for grids to be used for more than computing services: for generic service requirements across the enterprise, such as deal retrieval and market requests. “All of those services have nothing to do with heavy computing, but they are all services that could be moved onto the grid,” he says. ■
References:
Archives