Slashdot Mirror
Building an Energy Efficient Datacenter?
asc4 asks: "The company I work for is a webhosting and colocation company. As our power utilization grows, we have begun searching for ways to make our datacenter more efficient. The biggest hit from the utility company comes in the peak usage charge, which penalizes (rather severely) for the highest sustained burst of usage during a billing period. Due to the nature of the colocation business, we can't control how much or when client devices use power, so I'm wondering: is there's something we can do at the datacenter level to help smooth out our power consumption, over the course of a given period of time?"
"In these days of hybrid cars, Energy Star devices, and in general more eco-friendly power consumption, it seems like there must be some products out there that can help make datacenters more efficient, as well. Could fuel cell technology be something to look into? Would flywheels or capacitors help? How about using more efficient AC units than what are available from the big names? What are others doing to reduce peak power consumption in high-drain datacenter environments?"
It's not necessarily cutting power consumption, but will reduce monthly bills and is eco-friendly. I'm thinking like solar or wind assist (depending on your geographical location)
During off-peak time, pump water uphill to a holding reservoir - a big swimming pool on the roof might do.
:)
Heat the water with the waste heat from the cooling units.
Sell access for swimming - nice warm water (well, here in Canada we like it warm
During peak hours, drain the pool back via generators to make electricity. (make sure you tell people first)
Use warm water to cool more - generate steam.
Run steam through turbines to generate electricity.
Use electricity to pump more water to pool on roof
continue as needed
Been there, done that, paid for the T-shirt
and didn't get it
If you can, install solar panels on your roof. It will smooth the peak a little, and also reduce your overall expenditure. If you are in a sunny location, the investment can often be recouped after only a couple years. Most utilities will even subsidize such ventures.
If that's not an option, server consolidation and virtualization for the people whom it is appropriate for are the only other options I can come up with...
I am going to guess you have 3Phase power perhaps through more than one primary link. Do they charge
based on the peak of one phase or the average of all. If you aren't balanced on your phase input into your building, you may be able to rebalance and see some benefit there. If you have one or two large UPS systems that are pulling equally across all three phases, make sure that the output of the UPS system is also balanced, that could end up bringing your input usage down.
This of course wouldn't help with your peak usage, but something to consider anyways.
Short of that, you would be looking for something that could store power and charge that at a regular rate. But then you could end up possibly shorting your demand on the output side based on the available power in that 'system' at peak times.
I am going to guess your best bet is to look at phase and load balancing through your power distribution network and make sure you have placed your clients. If I was in a similar situation, I would set up a collection of load coils across each hot lead in your power distribution network and graph the values on a tight schedule (in order to catch peaks) and determine what is responsible for your peaks.
Don't know if any of this would help, but it is discussion, mod accordingly.
In many states, you can save substantial amounts of money by agreeing to scale back energy utilization during critical times. In New York, NYSERDA (www.nyserda.org) is the agency that administers the peak load reduction program.
Conformity is the jailer of freedom and enemy of growth. -JFK
I used to work for a company where I was in charge of building automation and peak demand limiting. We used several strategies for this. 1. Use thermal storage where possible. The only real source you can control is the cooling/heating for the building, and you want to build uip as much of what you need during low periods of usage, like in the middle of the night. If you're in a cold climate, store heat, and if you're in a cold climate store cold. Use water large water tanks to achieve this. It will cost you to install them initially, but they will pay for themselves in a surprisingly short period of time. 2. Monitor the usage and trim where you can when you're hitting peak demand. Turn off lights, coooling units, etc., for the short time that it's required. Pre chill or heat the building ahead of time. 3. Run your backup generator to supplement existing power if you have seasons where usage is much greater than at other times of the year. If you have to run it every day of the year it won't help due to maintenance and fuel costs. But if you need it for short periods to chop the peak then it's well worth it. Again, it will more than pay for itself. The power company may even pay you to supplement them with it. 4. Look for alternative methods to heat or cool, or even generate power. You'd be surprised at what's available now for that.
If you're right on the edge of getting nailed for peak load, you could run the aircon aggressively before the peak load period and try to coast through it with the unit off. Chill the place to 60F, shutdown the aircon a few minutes before peak load and see how long you can go before turning it back on. Economizers can work well at reducing your aircon load. We pull in cold outside air at 5AM and cool the building down to 65f. This saves us about 2 hours of aircon running during summer days.
"Eve of Destruction", it's not just for old hippies anymore...
For smoothing out power usage, there are a number of different options -- aside from alternative energy, you could do rolling brownouts throughout your datacenter and rely on UPSes or generators to keep things going -- but you *will* take a hit in reliability. Every switchover -- one mains circuit to another, mains to battery, etc. -- carries some risk.
I've watched an entire datacenter go out on what was supposed to be a controlled switchover -- power company needed to do some work, pulled the plug (with the datacenter's consent), the backup generators start... and then die. The UPSes kicked in, but could only supply 15-20 minutes of power. Everything failed over to a backup datacenter, whose link then decided to go out to lunch.
Total cost of the outage was measured in tens of millions of dollars.
Just keep this in mind when doing the business justification calculation (cost savings from lower energy bills, minus upfront cost of equipment, minus risk of additional downtime times cost of downtime, minus cost of maintaining the equipment). Unless energy prices go *way* up -- like oil hitting $250/barrel -- I'd be surprised if this would pay for itself.
"Could fuel cell technology be something to look into?"
No. Fuel cells are a way of transporting energy, not creating it. This is such an important concept to grasp that cannot be understated.
We are in deep trouble, energy wise. There is no immediate solution (within the next 30 years) that can help us. We need to get used to that concept, fast. Doing "your bit" for the environment is simply not enough.
Welcome, too, China and Inda. Welcome to the powerdown.
Invoicing, Time Tracking, Reporting
I take it you have quite a server farm.
Intel sells a lot of crap, so take some of it and use a methane generator to produce power.
I have one of these (1.2GHz) and with 1 large HDD, encoder card, network, DVD etc - it idles at less than 20W and maxes at about 60 (encoding, playback, DVD all going, CPU 100%). Burst power when switched on seems to be about 72. This is less than the processor alone on a high spec box.
This will only work with non-CPU intensive operations. However IO seems to be pretty good on these boxes, so an IO bound server would probably not suffer too greatly using a VIA mobo.
I smell a Nobel Prize in Physics here.
Some things are easier to do in the design phase. but something can be done now.
/. earlier this week, keeping the building cooler in the morning and warmer in the afternoon can drop your peek time costs.
First, pre-cool the room. There was a good article on
Second, install a solar power system. Kinda pricy, but if you have a large roof you can generate some solid power. And don't think that being in the north excludes you from solar power. Uni-Solar has a great sun index map showing what level of solar output and electrical output you can expect in any given area.
Third, going with solar, a battery array or some other type of power storage. By using the solar pannels to juice up the batteries, you can pull power from the batteries at peek time, but charge them all day.
Fourth, sub-teranian cooling. Once you get a little ways under the surface of the ground, the temperature becomes a pretty consistant mid/high 50's. Using sunken water tanks you can run 60 degree water through a radiator in your HVAC system. I know there are companies that can install these system but I can't recall any names off the top of my head.
Fith, solid state storage. If you can swing paying $50/gig as opposed to $1/gig for storage space you can dramaticly cut down on your both your cooling bill and your electric bill.But at $50,000 per ter vs $1,000 per ter, it's going to take a while to recoup the costs.
Sixth, custom server cases/cabinets. Traditional closets are great for cramming a lot of servers into a small area, but they about suck for heat management. You could fund a research project at any number of engineering schools to create a better storage solution.
-Rick
"Most people in the U.S. wouldn't know they live in a tyrannical state if it walked up and grabbed their junk." - MyFirs
1. Cool down the centre during the night when hydro is at its cheapest.
2. During the day raise the thermostat so the AC does not kick in too soon.
3. If you have windows use the blinds on the sunny side. Thermal load is a royal pain. Where I work it hit 27c inside even though it was -14c outside. The north side was running at about 21c.
4. Put all non-essential equipment on powerbars and turn off the bars. Most monitors and other electronics still draw a bit of current for 'instant on'. That takes hydro and dumps more heat for the AC to handle.
Panic now, beat the rush!
Get a professional electrician in that knows about peak charges.
Older installations used to use giant flywheels, but not to limit peaks. They were used for power conditioning and limited power backup.
I'd do an extensive survey before trying anything else. Buy or rent a power meter that does logging and graphing. Check everything out for a month - each phase and the current draw on each phase, and current draw on each rack (each computer if possible).
Proper sequencing of cooling can drastically affect your power consumption. Never start your cooling motors when you're drawing a lot of power - motor startup is a huge peak. After doing a survey of your power needs you may be able to identify times when you can avoid turning the cooling system on which will lower your peak. For instance, before the daily peak, cool the data center down a few degrees more than usual. Then shut off one or more cooling system until after the daily peak. This can be tricky to correctly manage and implement, especially since it has to be automatic and failsafe.
Alternately, shop around for your power. check with a few competitive companies and see if they offer a better deal.
-Adam
Being a datacenter you would undoubtedly have a generator backup to your UPS solution.
Would it be cheaper/feasable during these peak times to "test" the generator... ie turn the mains power off and run on diesel?
A Tale of 2 idle hands
If you've got a datacenter large enough that energy efficiency is a problem, I recommend you move the whole shebang to a location where energy is more plentiful. Upstate NY, which has plenty of hydroelectric power, would be a good choice. Nowadays, thanks to the internet, you don't have to keep your datacenter next to part of your operation.
Solar panels kick out small voltage through out the day. True they will likely peek at the same time as your peek electricity. But the amount of power they put out at peek is not going to be much compared to your total consumption. So instead of using the power as it comes through out the day, where in the morning you may save 10kWh for say, 8 cents per kWh, you can instead store that juice in a battery for peek time and save 9kWh (due to loss) and 20 cents per kWh. Yes, it would cost extra for a battery array, which is why I listed it seperately, but with it, you could replace more of your most expencive power with the cheapest, instead of replacing a smaller amount of power through out the day.
-Rick
"Most people in the U.S. wouldn't know they live in a tyrannical state if it walked up and grabbed their junk." - MyFirs
I smell something alright ;)
-Rick
"Most people in the U.S. wouldn't know they live in a tyrannical state if it walked up and grabbed their junk." - MyFirs
Seriously. Try killing the flourescents and not allowing "lighted" maintenance during certain peak times.
On the other hand, that might be a dumb idea.
Don't think that a small group of dedicated individuals can't change the world. It's the only thing that ever has.
Switch to natural gas to run the air conditioners. Your peak electricity hit is in the middle of the day when the air conditioners work hardest, but the peak natural gas hit is in the middle of the night when the exterior temperature is coldest. Price wise that works to your advantage.
Moderating "-1, Disagree" is simple censorship. Have the guts to post your opinion.
"is there's something we can do at the datacenter level"
Yeah use the Ultrasparc T1 CPUs, use lower power scsi disks including compactflash disks for boot and OS, keep all lights out when you dont need em, add heavy wall insulation unless youre living far north, add lots of ram in all machines so the disks can be powered down etc.
"Give orange me give eat orange me eat orange give me eat orange give me you." -Nim Chimpsky
I built a data center power monitoring system about a year and a half ago for exaclty this purpose (I installed it in my house and posted a writeup to slashdot... the article is now here). This system monitors every branch circuit in the data center and allows you to assign circuits to customers so you can track usage by customer. The first data center it was installed in was a colocation facility and their intention was to start billing for power like they do bandwidth. That is, you purchase power in 5amp blocks and when you spike, you pay for another 5amp block, etc... the thought being that if customers start having to pay for power, they will optimize it just like any other expense.
:: While I built the system, I don't own it, sell it, or work for the company that does. Anyone looking for more info should visit TrendPoint Systems.
To this end, the system was designed to let customers login to their accounts and see their power usage (with one minute resolution for a year... the gui is a java applet with real-time graphs, etc..) as well as set alarm thresholds, notifications, etc. The customer I built this product for recently completed a new data board that gathers meter grade current, voltage, watts, power factor, and kwh readings so all this history is now available to the user (and colo). There are independent threshold alarms sets for both customer and admin across all data sets, including panel level, per-phase rollups, etc.. (which amounts to almost 1400 alarm points for an 84 circuit panel).
The folks that use the system have told me it's actually almost more useful for capacity planning and load balancing given the increasing power density of customer cabinets these days.
*disclaimer*
- If your server room is not enclosed on the roof of the room (just using plain false-roof tiles) make sure they are atleast insulated very well. The more A/C escapes, the more it has to work.
- Make sure there is enough air-flow through your server racks (best placements and setup ideals very from person to another), best not to have the rear right up to a wall. Middle of the room or offset (5 feet or so from the well) allows for good ventalation.
- Keep server room lights off unless needed with the exception of a low-heat emergency lighting.
- If you have raised flooring and the a/c comes through the bottom, place the racks behind vent openings (so the air is rising to the front of the rack, getting sucked in by the fans in the front) instead of having the rack on the vent itself.
- Upgrade older servers if possiable. Older servers (expecially the old HP NetServer series) are a lot less efficient as newer servers. Not componet (CPU, HD) but also overall engineering.
- Turn off monitors when not in use. LCDs are not as bad but better to be safe then sorry. If you do not need it running, just leave it off.
- Do not allow people to keep the server room door open, may sound simple but you wouldn't beleive how many times I've seen this. If the doors don't close automatically, get automatic closers for them!
- Make sure the doors are weatherstripped.
- Multiple airconditioners! I have a small server room that runs on three airconditioners. Two always run, one does not, this rotates weekly. Also great for redundancy.
I'm sure there are many more things you can do. Hiring outside consulants who have worked with issues such as this are always benifitial. Be sure to get second/third opionions.
Wow, spelling really sucks when you haven't slept for 72hrs. (I really, really hate Exchange. Expecially when custered.)
I've been saying this for years, any outfit that already has a DC infrastructure should be installing photovoltaics on the roof. In a traditional PV installation, inverters and output wiring are a big part of the expense, but if that work is already done, the payoff period is a lot shorter.
Plus, in the event of a grid failure, your generator doesn't have to work quite as hard, which translates to slightly longer runtimes on the same fuel tank.
The available solar resource depends largely on latitude and weather patterns, though. Do some research and talk with the PR and marketing people about advertising your facility as "greener". If I were in the market for colo services, I'd lean slightly towards an enviro-conscious outfit, especially if they had a clue about reliability.
Disks (and other mechanical parts) will consume a lot of energy, but you don't need to replicate every single physical disk - if the data is under two gigabytes, RAM disks should be fine. In the event of a hard drive failure, backing up off RAM disk is no different from backing up from physical disk, so what's the difference? A single SAN-based disk pack, copied into RAM on the servers, would be the least power-consuming design - especially if you powered the hard drive off except when syncing up.
It costs power to task swap, so the more active tasks there are, the more swapping (if the tasks are all being given fair time) and therefore the more CPU time is taken by kernel activity, therefore the more power is being used up on housekeeping. You should be able to reduce the power consumed by heavy kernel activity by load-balancing.
If you're going to load-balance, you don't need high-power server-rated or desktop-rated CPUs. Mobile CPUs will take less power, you'd just need a larger cluster to load-balance over. If using Linux, also look at CPUs other than Intel - many MIPS and MIPS64 implementations are pretty low-power.
Networks take power to run. There's no escaping that. Don't run more wire/fibre than you have to (that also includes not running longer cables than you need), and don't use more intermediate network devices than will get the job done properly. Oh, and don't overspec the network for a given technology. CAT6 is good stuff, but if your machines never exceed 10 mb/s on the network, you're going to lose efficiency. The "for a given technology" matters, as different technologies will consume different amounts of power for a given spec. Shop around.
Cooling systems are another mechanical system and so are necessarily power-hungry. You can't put those in RAM, however. Again, shop around. You want the best cooling power per unit of energy. This may turn out, for your system to involve having several fans on a single component. It might equally well work out that you can link ducting together such that a single fan can directly cool many components. Since the energy efficiency is what is important, go for the most energy efficient solution for your system.
Depenmding on the system, it MAY (this is not guaranteed) improve the efficiency to have a variable-speed fan, with the speed controllable by the CPU, and where all components cooled by this system have thermal sensors readable by the CPU. You can then vary the cooling as a function of both temperature and predicted load levels. (Varying according to temperature alone is useless, as the loads on the components will change faster than the sensor readings - but could change in either direction. Since the OS knows what tasks it is currently doing, it should be capable of predicting the likely loads for a much more reasonable timebase.)
Connectors are notorious for high resistance and therefore power loss. If there is something that you're unlikely to change for the productive lifetime of the computer, all power loss through all unnecessary connectors (whih are generally made from poor conductors anyway, just adding to the problem) is power you can conserve simply by improving the connection. If you insist on using connectors, make sure the wires that go to the connectors are soldered and not just held in place by pressure. Also, clean the connectors thoroughly, as buildups of oxide and dirt will increase the resistance. You WILL be better off by removing the connectors entirely and soldering anything that's not going to change in place.
Finally, the data center's power grid. You want very high voltage, very low current. (Power dissipation is proportional to voltage, but proportional to the square of the current.) The industrial powe
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
As a data center manager myself, I can understand your pain. Unfortunately, I'm in charge of a corporate data center rather than a pure hosting arrangement; many of the tricks I've used to manage power consumption wouldn't work for you, but...
I'm able to play building load for the laptops/desktops off against data center consumption, and also able to relocate equipment to other sites to juggle the load. I have the option of passing the cost on to the customers because most of what I do is cost-plus contracts. I know this might cost you business, but it's something to consider. Other things that may help:
This may sound silly, but don't leave systems running with a failed component. A lot of servers run the cooling fans at higher RPMs if there's a power supply or fan failure.
If you're not already using SAN storage, consider switching to it. If you are, make full use of it by having your servers treat the SAN as the boot drive and removing all local drives. Better to have 20 servers accessing a 20-disk RAID on the SAN than those same 20 keeping 2-8 hard drives each spinning.
I'm going to assume that you're a high-availability setup, with UPS and generator coverage for a multi-day power outage. One of the simplest things you could do is set up a system where, if your power draw approaches your previous peak (or acceptable peak if there is a limiting factor), you switch to generator power. Whether this really helps you would depend on how far you need to limit your use, and how much refueling the generator costs you.
The other option may not be workable for you, but... if there's other office space in whatever building you're in, I'd look into renting it as a separate office of your company, and set up a second data center (shared staff, but separate electric service.) Dividing your current electrical load between two bills even 75%/25% would be a great way to limit peak load. This may even work with your current space if you occupy multiple suites in the building.
Pavlov's Dog ate the bell, and now he's barking at Schroedinger's cat all the time... -Me
Southwest Michigan may have more per square mile, but my 10th floor balcony on Lake Shore Drive has no shortage of wind. I see 30-50 mph gusts almost every day of the year due to the layout of the other high-rises around me.
They locate wind farms in mountain passes or other natural high-wind locations; I wonder if turbines located in certain spots of major metropolitan areas would be super-efficient. The plaza south of the IBM building on the river in downtown Chicago has to be one of the windiest places on the face of the Earth... I've literally been blown over there on several occasions.
...and don't forget Sun, if you feel like paying $20,000 for an 8-core Ultrasparc T1 chip that uses less power than an Opteron. Windows not allowed though. But since the original question is probably talking about managing a datacenter full of customer machines (which you can't control, unless you lease them), I would imagine that just changing the way you do cooling would make a rather large difference. Have you considered piping heat from the racks to the outside through the roof, and using a high specific-heat fluid (say water/ethylene glycol mix) that could be prechilled to an extremely low temperature during off-peak times, then used in a radiator-type arrangement for cooling. I would say that sealing the sides of your rack and replacing the front and rear doors with liquid-cooled radiators/heat collectors should work quite well. If you can pre-chill the air entering the front of the machines, it will be cooler coming out, and if you absorb the excess heat coming out the back, less heat will enter the room, meaning your air conditioners won't have to work as hard. Other good option would be to duct the exhaust heat (if it is significantly above room temp) up and out of the building with a few smaller fans.
--That's the point of being root, you can do anything you want, even if it's stupid.
Have all IT people work from home. :-)
No office space to cool or heat. No coffee machine or water cooler. No overhead. Just house the machines and an small maintance staff.
This has been another valuable and informative opinion from:
Catahoula!
Can you honestly suggest that a shorter Ethernet link will consume less power than a longer one? Sure, there's a tiny difference in cable resistance. But the transmitting end is putting the same amount of energy into it either way, since it doesn't know the difference. Any that doesn't turn into heat in the cable will turn into heat in the receiving chipset. Hence, the same power draw.
You *may* have an argument on very long fiber links. If you can get away with a short-reach transceivers instead of long-haul, you might save a watt or two. But again, if it's the same equipment at either end, the cable length doesn't matter, because any energy that doesn't succumb to cable attenuation just gets dissipated in the receiver. Or in some cases, in an optical attenuator pad just before the receiver, used to weaken a very strong signal so it doesn't destroy a sensitive receiver.
In both cases, the energy actually transmitted through the cable is so infitesimal, resistive losses are negligible and they don't matter anyway, because the energy that makes it through isn't used by the receiver, it's interpreted, then dissipated.
Tell me again how using CAT6 cable on a 10Mb/s link is inefficient? Financially yes, the cable's expensive, but electrically, the signal doesn't care. Neither cable is going to heat up more than the other one.
That being said,
Power cables are another matter entirely! Since they carry lots of current, resistive losses, even in adequate power cables, are measurable and significant. Particularly in 48v DC environments, as compared to 120v or 240v AC systems. To offset this, the power conductors in DC distribution plants are usually appallingly thick. Still, with a thermal infrared viewer, you can find "warm spots" in your power system. Fuses and breakers will always be a little warmer than the cables that feed them, out of necessity. But your power cables should be as short and thick as possible, within reason.
I say within reason because you're still only looking at a few dozen watts throughout the entire datacenter being lost to resistance in power cables. Rewiring the mess to use shorter cables will cost you more in labor and downtime than it's worth, but designing it right from the ground up might be worthwhile. In the meantime, just turn off the lights when you're not using them.
The rule to follow would be "if it produces measurable heat, it's something to look at". Network cables don't. Power cables, power supplies, processors, chipsets, drives, and memory do.
No, I'm not the experts but I refer you to the: Rocky Mountains Insititute . They are a not for profit environmental think tank who work with corporations and governments (Ford, the US Military for example) to increase profits or reduce costs through more efficient environmental practices. They ran a Design Charrette around this specific question. This is where they take their staff members with general energy efficiency expertise and a whole bunch of industry types (data centre types, power company types etc. and worked at redesigning the entire data centre idea from scratch with energy efficiency in mind. There is a detailed report including return on investment figures and detailed financial breakdowns. The information available is extremely comprehensive and free (as in beer). These guys are excellent and slashdotters might also like to look at similar exercises they have done with cars and energy security.
- Just trying to survive until the nanobots make me immortal -