Server Power Consumption Doubled Over Past 5 years

Watt's up writes "A new study shows an alarming increase in server power consumption over the past five years. In the US, servers (including cooling equipment) consumes 1.2% of all the electricity in 2005, up from 0.6% in 2000. The trend is similar worldwide. 'If current trends continue, server electricity usage will jump 40 percent by 2010, driven in part by the rise of cheap blade servers, which increase overall power use faster than larger ones. Virtualization and consolidation of servers will work against this trend, though, and it's difficult to predict what will happen as data centers increasingly standardize on power-efficient chips." We also had a recent discussion of power consumption in consumer PCs that you might find interesting.

Comments

Solution

[Ziest]

48 volt DC. Why the hell are we still putting 110 AC into the power supply and steping it down to 24 volt DC. And what do you get when you do that? HEAT. And to compensate for not having a better power system you then get to spend a fortune on HVAC to cool the room that you heat by stepping down the voltage. 110 power supplies make sense in the home but in a data center it is stupid.

Re:Solution

[Ungrounded Lightning]

So how do you get 12 volt, 5 volt, 3.3 volt, and 1.5 volt DC from that?

High-efficiency switching regulators on the blades. (They're actually getting so good that you have less heat loss by putting a local switcher near a power-hungry chip than by bringing its high current in at its low voltages through the PC-board power planes.)

Getting the raw AC->DC conversion out of the way outside the air-conditioned environment saves you a bunch of heat load, as does distributing at a relatively high voltage (such as "relay-rack" standard 48VDC) to reduce I-squared-R losses. And switchers are more efficient with higher raw DC supplies, so going to 48V (about the highest you can while avoiding touch-it-and-die shock hazard - which is why Bell standardized on it) is much better than 12 or 24.

Re:Solution

[hackstraw]

I didn't read the Google paper (or the FA for that matter), but while we are on the subject, this is something that I don't understand.

Why do servers have AC to DC power supplies at all? I don't know about you, but I have my servers on UPSes. So the whole thing goes AC from wall to DC in the UPS to the batteries then from DC to AC to the computer where it converts it back to DC.

I'm not an EE, but why cant AC come from the wall into the UPS and then the UPS spits out DC to the computer?

Granted the UPS power supply needs to be redundant because they are the 2nd most likely thing to fail in servers after disks, but what am I missing here? I know there are telco grade computers that take DC, but these are not available in many options and are typically lower end boxes. But to me, none of these additional conversions to AC and DC an back again with the added likelihood of a failure anywhere in the chain seems a bit non-optimal.

Did you know disco record sales were up 300%?

[Phanatic1a]

In the US, servers (including cooling equipment) consumes 1.2% of all the electricity in 2005, up from 0.6% in 2000. The trend is similar worldwide. 'If current trends continue ...then by the year 2100, server rooms and cooling equipment will consume over 300,000% of all the electricity!

And how much energy did those computers save?

[WoTG]

It's not like we plug in computers to sit around idling all day. They're doing stuff. I can send an email to anywhere on the planet instead of stuffing and envelope to have it carried by truck, boat, or plane. Cars have better power plants than ever before... they didn't get that way with back of the envelope calculations! A lot of forms that I used to submit by fax or snail mail? All gone electronic.

So, computers are using more power than 5 years ago? Who cares? If it bothers you, then get off the grid and fun in your cave.

Re:"Alarming" increase in "alarming" statistics

[AusIV]

I agree. Server power consumption may have doubled over the past 5 years, but what has the increase in data throughput been? Using a mutilated version of Moore's law, I'll assume that each server is doubling it's throughput every 18 months. 5 years is 60 months, so each server should have doubled 3 and 1/3 times, meaning each server is over 8 times more productive than they were 5 years ago (it's closer to 10, but we'll round down, as I'm trying to make this a conservative estimate).

It's also safe to say that there are "more" servers than there were 5 years ago, but I'm not even going to venture a guess on how many more. Assuming we have the exact same number of servers we did 5 years ago, we'd be processing 8 times more data per kilowatt-hour, meaning the cost of processing data has fallen by 75%. My estimates of data throughput may be high, and my server quantity estimate is definitely low, but I'm guessing 75% is a low end estimate.

Even if we are using more energy, we're getting more bang for our buck. I'd rather have data traveling through servers than on planes and trucks in the form of mail. I'd rather have documents be stored in mass on hard drives than have millions of pages of paper going to waste. Suggesting that this increase of power consumption is alarming is absurd.

Ok, so power use doubled...

[Aphrika]

...but how much did performance increase by?

overcapacity, spam, botnets

[bcrowell]

Personally, I use an insanely wasteful server because I don't have any choice. 99% of the time its cpu is 99% idle. However:

• You can't get webhosting with good support and reliability unless you pay for the level of webhosting that gets you your own box.
• I need my server to be able to stand up to a spike in demand caused by ten thousand spams hitting it in three seconds...
• ... or 1000 ssh login requests in one minute from a bot searching for weak pasword...
• ... or a brain-dead bot requesting the same 5 Mb pdf file 10,000 times in one hour, and sucking down 60 Mb worth of partial-content responses.

Similar deal with multi-core CPUs. People are talking about making desktop machines into the equivalent of 1980 supercomputer, and one of the main justifications seems to be that anti-virus software can run all the time without affecting responsiveness. This is nuts. The internet and its protocols weren't designed for a world infested by Windows machines controlled by malware.