The Open Compute Project is driven by Facebook as a way to innovate in datacenter excellence for all.
Tom Furlong, Director of Site Operations and Jay Park Head of Datacenter Design highlighted the Facebook 2.0 DC which includes
- fully custom servers
- fully custom cooling for the datacenter
- completely new power distribution within the datacenter
Facebook have multiple DC's around the USA and use Akamai to serve content locally and Tom Furlong summed up the idea with the quote "if you own the DC and you own the application, you should own the server".
The Prineville datacenter
- has no chillers
- has no datacenter UPS
- has per rack type UPS
- has a custom evaporative cooling system
Each chassis
- has 270Volt AC power connected (to reduce leaks from multiple conversions)
- has a48Volt DC connected from the rack's UPS providing 90seconds of power backup
- is 1.5u (to find the sweet spot between impedance of cooling and efficiency / number of servers)
- bare bones mother board
There are two server types
- AMD with upto 386G of RAM (for the memcache clusters)
- Intel with upto 288Gb
Whats interesting is that Facebook have benchmarked and found things like
- 40% capex saving (no cooling / no DC UPS)
- 30% opex saving (servers are designed to be worked on more efficiently meaning a greater server to datacenter engineer count)
- (there were many others though couldnt write them down in time, suffice to say Tom Furlong said there were cost savings across the board)
Whilst there I seen the Rackspace EMEA Operations Directly - Jacques must be looking at ways to cut down Rackspace infrastructure costs.
Its such a shame that realestate.com.au isnt big enough to have our own datacenter as it would be great to get involved in the Open Compute Project - at least I got a cool tshirt from the talk! :)
PlanetMySQL Voting: Vote UP / Vote DOWN