Facebook refuses to employ virtualisation across its infrastructure, because the technology does not scale efficiently, the director of the social networking site's labs division has said.
The social network shies away from abstracting its technology into virtual machines because at large-scale, the economics don't work out and the cost of dealing with hardware failures is too high, Gio Coglitore of Facebook Labs, said at a press event on Tuesday.
"We find within our testing that a realised [non-virtualised] environment brings efficiencies and the ability to scale much more effectively," Coglitore said. "When we look at the operational realities, we very much would like to lose a server and not have it impact our experience, but as your start to virtualise the importance of that server becomes enhanced and at scale becomes difficult."
By way of example, he said that because hundreds of virtual machines, each running different applications, can be run on top of a single server, if that specific hardware fails it is complicated and costly to migrate VMs, compared to starting again with new hardware.
"I prefer to have little segments lost and have a load balancer redistribute that to a wider pool of individual smaller computers," he said, and explained that was more difficult to do if the applications were virtualised. "At scale, fewer and fewer folks find that [virtualisation] is an effective way scale out".
Gio Coglitore was speaking at an Intel event in which the hardware giant detailed its plans for microservers — relatively low-powered, dense servers, taking up one or two rack units.
At the event, Coglitore said Facebook had tested microservers within its infrastructure in a production role but that "the possibilities of deploying [microservers] en mass is a late 2011, early 2012 phenomenon."