Making Your Datacenter Into Less of a Rabid Zombie Power Hog

Nerval's Lobster writes "Despite the growing list of innovative (and sometimes expensive) adaptations designed to transform datacenters into slightly-less-active power gluttons, the most effective way to make datacenters more efficient is also the most obvious, according to researchers from Stanford, Berkeley and Northwestern. Using power-efficient hardware, turning power down (or off) when the systems aren't running at high loads, and making sure air-cooling systems are pointed at hot IT equipment—rather than in a random direction—can all do far more than fancier methods for cutting datacenter power, according to Jonathan Koomey, a Stanford researcher who has been instrumental in making power use a hot topic in IT. Many of the most-publicized advances in building "green" datacenters during the past five years have focused on efforts to buy datacenter power from sources that also have very low carbon footprints. But "green" energy buying didn't match the impact of two very basic, obvious things: the overall energy efficiency of the individual pieces of hardware installed in a datacenter, and the level of efficiency with which those systems were configured and managed, Koomey explained in a blog published in conjunction with his and his co-authors' paper on the subject in Nature Climate Change . (The full paper is behind a paywall but Koomey offered to distribute copies free to those contacting him via his personal blog.)"

Comments

Rabid zombies

[Kohath]

are a renewable resource.

Less powerconsumption = less cooling

[plopez]

I've pointed this out a number of times. But people do not seem to "get it". If you can reduce your power consumption then there is less waste heat and then less cooling cost. Note too that if your applications use lots of disk reads/writes and network IO with the cpu in a waiting state then you can save power by using a lower end gear. E.g. laptop chips and slower memory vs full blown "Enterprise" hardware.

Re:Less powerconsumption = less cooling

[TeknoHog]

My thoughts exactly. My first web server in 1998 was a laptop, and ever since I have wondered why 'desktop' components waste so much power compared to 'mobile' counterparts. Since 2003 my 'desktop' machines have been built with 'mobile' CPUs (Mini-ITX et al) and I keep asking this: why should a machine waste power willy-nilly just because it is plugged in? I also like the quiet of passively cooled CPUs (of course, other components like PSUs can be passively cooled).

Re:Less powerconsumption = less cooling

[sjames]

It works the other way too. If you don't cool the servers at all, eventually they stop consuming power ;-)

Re:Less powerconsumption = less cooling

[ShanghaiBill]

It works the other way too. If you don't cool the servers at all, eventually they stop consuming power ;-)

Eventually. But not as soon as you might think. Modern servers can tolerate heat fairly well, and many data centers waste money on excessive cooling. As long as you are within the temp spec, there is little evidence that you gain reliability by additional cooling. Google has published data on the reliability of hundreds of thousands of disk drives. They found that the reliability was actually better at the high end of the temperature range. This is one reason that Google runs "hot" datacenters today.

Re: Less powerconsumption = less cooling

[jd2112]

if you can run on a low power system as you describe it is probably a good candidate for virtualization.

Re:Less powerconsumption = less cooling

[evilviper]

you can save power by using a lower end gear. E.g. laptop chips and slower memory vs full blown "Enterprise" hardware.

"Enterprise" hardware doesn't mean the fastest... Infact it's the opposite, as enterprise hardware has longer development cycles.

Enterprise gear means things like ECC memory, BMCs monitoring server health, HDDs that won't freeze up for several minutes retrying a single unreadable block error, etc. And if you feel like skimping on it, you'll end up paying much more in the long run, as a single flipped bit in your database can cost a company obscene amounts of money, RAID arrays will report disk failures far more often, you'll be paying for remote hands, and waiting for on-site access far more often, and you'll have no notification nor insight into your servers, as hardware keeps silently failing, rather than alerting and halting to protect your data.

what about cuting down all the ac to dc to ac

[Joe_Dragon]

Why can't there BE UPS with ATX DC out?

Re:what about cuting down all the ac to dc to ac

[petermgreen]

48 volts is high enough that is doesn't need the big fat wires that 12VDC high-amperage connections do

48V while not as bad as 12V still means much thicker cables and/or higher cable losses (most likely some combination of both) than normal mains voltages.

Servers at full load can draw a heck of a lot of power. 500W is not unreasonable for a beefy 1U server, put 42 of those in a rack and you are looking at 21KW.

Feed those servers with a 240V single phase supply and you are looking at about 88A. That is high but managable with the sort of cable sizes you can find at most electrical wholesalers.
Feed those servers with a 240V/415V three phase plus neutral system about 29A in each phase conductor (and ideally nothing in the neutral). Easy to deal with
Feed them of a telco style -48V DC system and you are looking at about 438A. That is getting into the territory of busbars.

and computers would just need a DC-DC converter to convert the incoming voltage to the 12 and 5 volt rail voltages.

To prevent large currents flowing where they are not wanted you'd almost certainly want isolating DC to DC converters in the computers. An isolating DC-DC converter isn't much different from an AC-DC power supply (the main difference is that the rectifier and power factor correction stuff can be eliminated).

Re:And scaling up

[mlts]

This. With the availability and reliability of SANs, virtual machine software, hypervisors, rack/blades, and such, there are a lot of tasks which are best moved to a rack/blades/SAN/VM architecture. Even high/extreme I/O can be handled by virtualization on POWER and SPARC platforms.

These days, for most tasks [1], the question is why not a rack/blade solution. A half-rack with a blade enclosure and a drive array oftentimes can do more than 2-3 racks of 1U machines.

Security separation is getting better and better. Even Microsoft is getting a solution out there that is good enough for prime time with Hyper-V in Windows Server 2012. IBM has had top notch separation (well, since the days of the 1970s and VMs on mainframes), Oracle as well.

To go "green", if a data center hasn't already gone with P2V, they should. There are always exceptions, but this is something to be considered.

This also helps with the next buzzword I'm hearing bandied about from the PHB types -- the SDDC, or software designed data center.

[1]: Ones that do not require specialized high-speed hardware like professional video capture. Of course, there are other tasks that require separation due to heavy I/O such as Netbackup servers. Then, there are servers that have to be separated for security or management reasons. For example, a SDMC for POWER boxes should be on discreet hardware for security reasons. Similar with the VM for vCenter management so it can be powered on and used even if the main cluster is inop.

Of course, having cheap, discrete hardware for large scale operations like Facebook makes rack/blades not as useful, but most data centers will benefit from the P2V move.

Re:And scaling up

[evilviper]

These days, for most tasks [1], the question is why not a rack/blade solution. A half-rack with a blade enclosure and a drive array oftentimes can do more than 2-3 racks of 1U machines.

This is complete nonsense. Blade servers are more expensive, and CAN'T outperform simple 1U servers. 1U servers are packed to the gills with the hottest components that can be kept cool given the amount of space they have to work with. Blade servers, or any other design, can't possibly pack things more densely than 1U servers have been.

And Blades can't compete with virtualization either. It's just not remotely as flexible. You can't oversubscribe the memory of a Blade server, since that memory is physically dedicated to the single OS running on it. You can't oversubscribe CPU, nor boost the CPU when you need more performance.

You get more expense, with less performance, and less flexibility. Blades need to die off already...