Power Consumption and the Future of Computing

mrdirkdiggler writes "ArsTechnica's Hannibal takes a look at how the power concerns that currently plague datacenters are shaping next-generation computing technologies at the levels of the microchip, the board-level interconnect, and the datacenter. In a nutshell, engineers are now willing to take on a lot more hardware overhead in their designs (thermal sensors, transistors that put components into sleep states, buffers and filters at the ends of links, etc.) in order to get maximum power efficiency. The article, which has lots of nice graphics to illustrate the main points, mostly focuses on the specific technologies that Intel has in the pipeline to address these issues."

Comments

Big cuts

[Forge]

The thing with power usage is that nobody seems interested in attacking the 2 largest areas of power wastage. (except maybe google)

#1. DCAC conversion.
Your typical Datacenter has a UPS or batteries and inverters (Enterprise scale UPS). What this amounts it is AC power from your utility company converted to DC for storage in a battery and then converted back to AC to supply the Server's power supply, then converted back to DC to actually run the components of the computer.

Ever notice how hot a UPS gets during normal operation? That's power going to waste. The solution is to run our servers at a standardised DC voltage. 48 Volts sounds good since that is already defined for Telecom equipment (correct me if I'm wrong. I am not sure of the figure)

#2. Raised flour and underground AC. A good chunk of datacenter power is used to run the air conditioning. If we abandoned the notion of raised flours and replaced them with say insulated celling mounted ducts with vents faceing each rack.

While we are at it here is another simple power tip. Turn your rows of racks back to back. When they all face the same direction, hot air blows from the back of one machine to the frunt of another, forcing the AC to work overtime. In my design, I would have extraction fans betwean my back to back racks, pumping the hot air outside (or into the office during winter. For those of you who have winter.

Re:Big cuts

[NeverVotedBush]

Lots of errors in your suppositions.

DC/AC conversion? The bigger data centers can't use batteries - too many, too big of a hazard, etc. They use rotational UPS's. These stay AC all the way.

Additionally - power distribution is better at higher voltages. It's that current squared thing. More and more equipment is also going to higher voltage distribution on the boards with local DC/DC conversion at the load. For the exact same reason. Our center distributes at 208 volts.

The argument against a raised floor is bogus. That acts (and is necessary) not only for cabling, but also for air distribution. Heated air rises. Feeding cold air up from the floor to where it flows into the racks to be heated and then recovered at the ceiling is the most efficient way for air. The fact that the floor is not insulated is a non-issue. The whole room is being cooled. The temperature is the same on either side of the floor tiles.

And about the face to face and back to back layout of racks - every single one of our racks is already in that orientation for exactly that reason. We have hot aisles and cold aisles and the temperature difference between them is pretty marked.

The next wave is a move back to "water" cooling. Either plumbing liquid to each rack where in the rack it locally grabs heat from circulated air within the rack, or plumbing into the boxes themselves. This is simply because heat loads are going up and it gets harder (and louder) to pump enough air through a building to cool the more dense newer equipment. Plus people don't have to put on jackets to go out on the floor or yell to be heard in a big data center.

Re:Big cuts

[Firethorn]

I've seen raised floor AC done right. Each rack was sealed, had a vent in the bottom and a vent into the ceiling. The AC pushed cold air into the subfloor, which was then sucked into the racks, with hot air rising into the ceiling. Where the AC pulled the hot air to be cooled again.

Also, 99% of UPS units don't convert AC to DC unless it's charging the batteries. Normally this would only be a trickle charge. If the UPS is providing power, you're in a critical situation anyways, I wouldn't worry about the fact that a UPS isn't particularly efficient, as you're probably spending 99% of your time not on UPS.

As for switching to telephone industry standard 48V power, you'd be converting it again to whatever the equipment wants, much of it 12V or less. 120VAC->12VDC is more efficient than 120VAC->48VDC->12VDC. In addition you run into the problem that 120VAC over 12gauge cable wastes less than half of the power that the same wattage of 48VDC would waste over the same diameter cable. So you'd have to use heavier gauge cable - payback isn't quick for that by any means.

You might be able to get away with it on a rack level, powering all the blades on 48V via rails to a couple of redundant power supplies somewhere in the rack. Either top or bottom, depending upon cooling and other requirements, though the middle might be an interesting choice, as it'd allow you to have half the wattage running over the rails on average(you'd have two runs instead of one).

You want to save power? I'd switch to feeding the racks/power supplies with 240V lines. Half the line resistance for the wattage.

Re:Big cuts

[DaleGlass]

Also, 99% of UPS units don't convert AC to DC unless it's charging the batteries. Normally this would only be a trickle charge. If the UPS is providing power, you're in a critical situation anyways, I wouldn't worry about the fact that a UPS isn't particularly efficient, as you're probably spending 99% of your time not on UPS.

That's a cheap, consumer oriented UPS. Datacenters use the kind described, ones that are always doing the AC -> DC -> AC conversion. What this achieves is that instead of the UPS taking over when the line voltage isn't good, the UPS is always providing clean power because everything goes through it. One of the advantages is that this kind has no delay for switching between AC and battery, as there's no switching involved.

Google Distinguished Engineer's point of view

[naeim]

A couple of months ago, Luiz André Barroso of Google gave a talk at Stanford about this very topic. Unfortunately the talk wasn't recorded, but here's a summary: http://cs343-spr0607.stanford.edu/index.php/Writeu ps:Luiz_Andr%C3%A9_Barroso

The future computing device uses less than 10W

[Anonymous Coward]

The future of desktop computing is 24/7 thin clients/home servers using less than 10W and passive cooling without fans, because for a typical 300W desktop 24/7 system you probably would be paying $100/month, more than a thousand a year. This is enough for 90% of users, those who are not after the latest/greatest 3D horsing power, those whose necessities are supplied with an onboard graphics chip such as Intel X3100 or even less. You would be surprised with the amount of computing power such devices have nowadays.

They do not use hard disks, but flash memory/pendrives as storage for the operating system and homedir, and are passively cooled, so they do not use fans, which are noisy and spend more energy. Massive storage (TBs) can be added if necessary, each one using an extra ~15W. A small list with some of them:

1) Linutop: http://www.linutop.com/
It comes with xubuntu, 280euros.
~6W, AMD Geode LX700 433MHz, 256MB RAM, Audio, 100baseT, 4xUSB2.0

2) Zonbu Zonbox: http://linuxdevices.com/news/NS9073106297.html
It comes with Gentoo Linux, $250.
~15W, VIA C7 1.2GHz, 512MB RAM, Audio, 100baseT, 6xUSB2.0

3) Mini Linux PC: http://linuxdevices.com/news/NS6372429785.html
Not sure about which Linux flavor it comes with, but if it runs Linux, it runs Ubuntu, $99.
~5W, 200MHz x86-compatible, 128MB RAM, Audio, 100baseT, 3xUSB2.0

4) OLPC: http://www.laptop.org/laptop/hardware/specs.shtml
Not yet available, but specs are fine for a home server + external storage, ~$100.
~2W (!), AMD Geode LX-700@0.8W 433 Mhz, 256MB RAM, Audio, LinuxBIOS (!), wireless connection, 3xUSB2.0.

Many others: http://linuxdevices.com/articles/AT4923746399.html

Re:1000 watt power supplies

[MP3Chuck]

If you look at the number of reviews, though, compared to something like a more modest 500W PS it would seem that not too many people really use/need a 1KW PS.

In fact, those high-end 1KW supplies might even be better for power consumption since they tend to have higher efficiencies than the cheapo options.

Re:Multipurpose.

[grimdawg]

I should point out that there are words that sound the same, but dynamically reconfigure their spellings to do different jobs as required.

Does slowing down idle CPUs help?

[rduke15]

I'm not sure ho much power this saves, but on all servers which I install, I use this Debian HOW-TO : CPU power management page. Basically, I do:

aptitude install cpufrequtils sysfsutils
cat /proc/cpuinfo | grep "model name"
modprobe p4_clockmod ## depends on your CPU!
modprobe cpufreq_ondemand
echo ondemand > /sys/devices/system/cpu/cpu0/cpufreq/scaling_gover nor
echo p4_clockmod >>/etc/modules ## depends on your CPU!
echo cpufreq_ondemand >>/etc/modules
echo devices/system/cpu/cpu0/cpufreq/scaling_governor = ondemand >>/etc/sysctl.conf

And I see my servers run at 350MHz instead of 2.8Ghz. or more.

Of course, these are all small workgroup or very small Internet servers. It would be of no use for a server which would be at the max speed most of the time.

Anyway, I haven't had an opportunity to meter the difference yet to see how much power that really saves. Does someone know?

Re:Does slowing down idle CPUs help?

[Wesley Felter]

p4-clockmod doesn't help at all; try acpi_cpufreq. The newer Intel processors have C1E, so they automatically drop to the lowest frequency when idle, so there's not a lot for CPUfreq to do.

Video

[Wesley Felter]

Here's a video.