IT Pain Points - And How Blade Servers Can Fix Them

Vern Brownell has seen both sides of the computing world. As chief technology officer of Wall Street powerhouse Goldman Sachs, Brownell ran one of the world's largest data centers. Frustrated with the inability of computer vendors to provide systems that solved the problems he faced there, Brownell left Goldman Sachs, and in March 2000 founded Egenera, a Marlboro, Mass.-based company dedicated to making high-end blade servers which address the pain points large data centers face on a daily basis.

What were the problems you faced at Goldman Sachs that led you to start Egenera?
One big issue was underutilized resources. I've seen analyst estimates that the average customer may utilize only 10 to 20 percent of their processing capacity in a typical data center. It's very difficult if not impossible to tap into that unused capacity with a traditional server architecture. In the storage market you have storage area networks (SANs), which give you flexibility, but that transition hasn't happened in the server space. So people still have to worry about what size server they're putting in, and because it's very difficult to resize it to the application, they wind up overbuilding it.

If your application grows, can't you just add servers?
It seems surprising, but at Goldman I found that it would take me a week to a month on average to get a server deployed. That wasn't because the vendor couldn't ship the server. It was because once you got the server in, you had to hook it up to your storage network, to your IP network, you had to get space, you had to get power, you had to connect it to BMC Patrol or whatever operations methodology you were using to back it up, and you had to connect it to your console network. All those things had different kind of SLAs and provisioning schedules, and coordinating them all added up to a project.

That was very frustrating, if you were trying to respond quickly to a business requirement. If you hit a period of peak demand or a particular part of your systems are under high-load, it's very difficult to respond to those things, unless you completely overbuild your infrastructure, which makes the utilization problem even worse.

What other challenges did you face as CTO?
High availability was very difficult to configure in the Unix world. We wanted to build mission-critical applications, so you had to have automated failover, and the state of the art in automated failover today is you buy two servers, one acts as a primary, one acts as a backup, you install clustering software from some other manufacturer like Veritas, and you add dual-ported storage, redundant networking connections, and so on. And you have to hand configure all this stuff, and that is your high availability cluster. But there is something inherently wrong with that picture because you're adding complexity to achieve reliability. And I'm a big believer that reliability is directly related to simplicity.

So what is Egenera's approach?
We call our product BladeFrame -- think of it as a mainframe approach to blade computing -- which we're aiming at the high-end, enterprise customers. The BladeFrame has 24 two-way or four way SMP processing blades, along with control blades and switch blades for networking, all of them fully redundant. The blades are connected internally, through a high speed interconnect, so you don't need separate cables to hook each blade up to your network and SAN. The whole system can be networked with as few as four cables.

Although it sounds like a non-glamorous thing, people are really crying out for ways to get a handle on the cabling issue. It's amazing to me how often a server goes down because a SCSI cable gets pulled out. As a customer, one of my pet peeves was that server manufacturers would claim five nines of reliability, but that's just for the server component, it has nothing to do with how it's cabled up, or how the server architecture forces you to cable up.

So what we tried to do is centralize a lot of the connectivity that you would need to build a large scale distributed Unix environment into our BladeFrame, and do it in a way that was completely redundant, so there's no single point of failure. We also aimed to make it really simple to manage. For example, all of the blades are hot-pluggable, they're UL certified for customer insertion. So we can FedEx a blade to a customer, and they just slide it in, it's a very simple process.

Egenera talks about something you call a processing area network. What is that?
Our processing blades have no disk associated with them, so there's no persistent state. That provides what we call a processing area network, which means you can dynamically allocate processing resources to applications. It's all done through software, if you suddenly hit peak demand for a particular application, you don't have to send someone into the computer room to recable servers. And you don't have to worry about buying the right size server for every application.

What operating system does the BladeFrame run on?
It runs Red Hat Linux, but we're adding .Net Server, so you can mix and match both .Net Server or Linux on the same machine. The BladeFrame already can support different generations of our blades, and different generations of operating systems, but we'll continue that even across generations of technology, so you'll be able to add Itanium into the frame. And because the blades are hot-pluggable, you can upgrade and never have to reboot.