DC or not DC
Tuesday Nov 21, 2006
Last week I was at the marvelous SC06 conference in Tampa, Florida. I had a great opportunity to talk with many customers, partners and competitors. I learned a great deal about emerging technologies and who is offering them.
I was pleased to see that the issues of power and cooling that we hear about so often with our large enterprise customers were also visible at a show about large super computers (its hardly surprising considering how similar these systems are becoming). Lots of interesting power and cooling products to follow up on.
I had a most interesting discussion with some folks at Rackable about DC power and how their systems are more power efficient. There seemed to be two big advantages to their solution.
Firstly they claimed that the newer very high efficiency (90%+) server power supplies on the market only achieve this efficiency if they are very highly utilised. Their solution to this it to install larger rack level power supplies and share power between server which this enables them to load the power supplies much closer to their maximum ratings.
Secondly they claimed that by moving the AC to DC conversion to outside the server and putting at the top of the rack reduced failures since the power supplies in each server didn't require moving parts. A stated secondary advantage is that less heat is generated in the server.
I was really impressed with their solution and resolved to look further into the benefits of DC power in the data center when I got back to the UK.
Now I'm back I'm no longer impressed, although their technology is still interesting its not as good as I was lead to believe:
The claim that efficient power supplies are only efficient with high loads is simply not true. Take a look at 80plus.org (a site which champions efficient power supplies), look at any of the power supplies they have tested. If you pick any wattage and any form factor they all look pretty much like this:
80plus.org Efficiency of the Enhance ENH-0840A power supply
Notice that this supply is most optimally efficient around 50% load, quite the opposite to what the Rackable folks told me.
There is also an extra loss in efficiency by having an extra power conversion in the top of the rack:
Sun's Bechtolsheim is unconvinced of the merits of DC, though. The crux of his argument is that DC requires two conversions: one from outside AC to 48-volt DC for distribution within the building, and a second, within servers, from 48 volts to 12 volts. Even if each conversion is 90 percent efficient, wasting 10 percent of power as heat, the combination makes the overall process 81 percent efficient.
By contrast, a single AC-DC conversion with the server uses power supplies rated at 90 percent efficiency, Bechtolsheim said.
I've not yet had a chance to look at the heat output from an normal power supply, but I'm willing to bet its not that great when compared to the CPU and memory of a system. Further I expect that the sum total energy loss from power supplies from a rack of normal 1U servers would be about the same or less than an equivalent number of DC supplied systems and their proportion of the loss for a shared power supply.
With respect to moving parts there is a lot to be said for getting rid of them. Perhaps this is the best feature of the Rackable design ...
One thing I have realised is that blade chassis designs which can benefit from a shared power supply but that distribute all the required voltages (12V, 5.5V, 3.3V etc) throughout the chassis directly to the blades actually would be more efficient and more reliable.
I know what server form factor I'd buy if I were a customer 
|
|
|
|
What I like about Rackable's and Verari's approach...