General Electric's road to better data center efficiency is paved with virtualization. Now, as GE looks to build an internal or "private" cloud computing environment, virtualization will once again play a key role.
GE has tens of thousands of servers in 10 primary data centers, and the number of servers in those facilities is growing 20% annually, amid increasing demand for IT resources from its business units and consolidation of servers from secondary data centers. GE has been investing in making its data centers run more efficiently. One example: 60% of new servers put into operation utilize virtualization.
CTO Greg Simpson isn't satisfied. He's now looking to handle GE's ever-expanding number of virtual machines by managing servers as a pool of resources rather than discrete units. "You get efficiencies when you start to manage that as a single entity," he says. "You can flex that capacity across applications, back it up, monitor it, and manage it as one large entity."
GE is in the early stages of a three-year project to implement technologies that give it the flexibility, automation, and manageability Simpson seeks. It's evaluating whether to have one expansive internal cloud or multiple discrete clouds dedicated to, say, Web serving or financial systems. Simpson's leaning toward the latter scenario, but either way, he wants to offer IT resources on demand and to charge business units for what they consume.
It's too early to say which products will serve as the foundation for GE's private clouds, Simpson says, but emerging virtualization and systems management technologies will play a role. He pushes virtualization aggressively--he talks of a "virtualization stimulus package" within GE--for all but the apps requiring extremely high I/O on dedicated servers. GE already uses a lot of VMware technology, so the vendor's vCloud technology will likely play a big role in GE's private cloud strategy.
Like many other companies that are well into virtualization, GE struggles with VM sprawl. Its server admins still manage VMs individually, and Simpson wants to move to managing them as a group. He points to Sun Microsystems' Containers feature in Solaris OS, which lets IT create multiple virtual instances of Solaris that can be managed as one entity, rather than having each VM as its own entity. GE expects to have multiple operating systems, including Solaris, in its cloud environment.
One of the challenges GE faces is finding the tools to build and operate a private cloud environment. While there are plenty of products (see story, "Why 'Private Cloud' Computing Is Real -- And Worth Considering"), Simpson says most are immature, especially when it comes to automated management.
"We're trying to avoid building special GE code, because we don't have the energy to figure out how to do this right all by ourselves," Simpson says. "We think we can help cloud vendors get there faster. We want to be able to take advantage of that and be on the leading edge."
Simpson plans to test Cisco's Unified Computing System, a new blade server system that combines virtualization and server management plus network and storage access, using a single management interface. Cisco's system "allows you to balance CPU, I/O, and memory," Simpson says. I/O and memory requirements are two reasons GE hasn't virtualized certain apps before now, he adds.
Today, GE buys more servers than would be necessary if workloads were better optimized, Simpson admits. A goal is to manage the amount of computing power available to specific apps, increasing and decreasing cycles as needed. "What we can do is manage the capacity of that cloud, take excess capacity, and provide it where it's needed," he says. But Simpson isn't yet testing technology to automatically scale computing power up and down, based on an application's importance and need. The capability just isn't here yet, he says.
GE's IT organization charges business units a flat rate per physical server that covers power, cooling, and hardware. A private cloud, theoretically, would let GE charge for IT resources based on consumption or per user, though Simpson admits he hasn't figured out just how that would work. Some business managers might resist usage-based pricing, preferring a flat rate.