Tuesday, May 17, 2005

Green Server Farms

Odograph today passes on a combination server and router that uses a scant 3.1 watts. The topic of server energy use is one that I've been interested in for some time, because my employer uses dozens and dozens of servers at a colocation facility (or colo) in downtown LA. Yesterday Slashdot ran a piece about green buildings in the context of colo power use, something APC executive Richard Sawyer talked about as a strong suit of that company. Companies with big server farms have started to notice those electric bills:

Energy efficiency in the data center has come up a lot recently. Has IT evolved to the point where it can consider energy efficiency without sacrificing uptime or performance?

Sawyer: Yes. A lot of the new equipment is getting more and more efficient. The Delta conversion products we use for our large scale [uninterruptible power supply] systems are extremely efficient. They're 95% efficient down to very low load levels. Even at a 25% load level, we're running about 94% efficient, as opposed to the old legacy style systems that were 60%-70% efficient.

With the cost of electricity increasing like it is -- a kilowatt hour going from $0.08 up to $0.13 per kilowatt hour in the Northeast-- it becomes a major factor. For a one megawatt load (1,000 kilowatts), to supply that system with electricity costs close to $1 million per year. That's a million dollars you can't spend on IT equipment. Even an efficiency difference of 10% works out to $100,000 per year. That buys a lot of IT applications or head count in your division. It's becoming a larger factor in the total cost of ownership. Obviously, the more efficiency you can put in, your ongoing expenses become more manageable.

With the price of electricity in California around $0.14/kWh, the dollars saved are even higher. And then there's the matter of the servers themselves. Colo-based server farms tend to have similar issues as power generation: capacity is sized, not by average load, but by peak response load. That means when there isn't a lot of load on the servers, they're still chewing the juice, wasting dollars. And even most servers aren't really that efficient just on a CPU-cycle basis. As one poster on the Slashdot thread noted, most servers spend very little time outside their idle loops (less than 2% on the machine he was using, but based on my own experience, 5% is pretty much the maximum you can expect).

Another factor here is that colos look at electricity as part of the package of providing services; there's no meter on the outside of a customer's cage or rack. I can only imagine this is likely to change given the ever-increasing costs of power. Servers now are optimized for performance; what would be useful is for the colo industry (already beset with thin margins as it is) to force customers to buy their own electricity, or bill them for it discretely, and better power management built into the servers along with traffic routing, such that machines could be kept in a low-power idle state when not needed.