Slashdot Log In
When Does Powering Down Servers Make Sense?
Posted by
timothy
on Thu Oct 30, 2008 02:35 PM
from the after-coffee-is-served-of-course dept.
from the after-coffee-is-served-of-course dept.
snydeq writes "Powering down servers to conserve energy is a controversial practice that, if undertaken wisely, could greatly benefit IT in its quest to rein in energy costs in the datacenter. Though power cycling's long-term effects on server hardware may be mythical, its effects on IT and business operations are certainly real and often detrimental. Yet, development, staging, batch processing, failover — several server environments seem like prime candidates for routine power cycling to reduce datacenter energy consumption. Under what conditions and in what environments does powering down servers seem to make the most economic and operational sense, and what tips do folks have to offer to those considering making use of the practice?"
Related Stories
[+]
IT: IT Vs. the Permanent Energy Crisis 285 comments
snydeq writes "Organizations looking to remain profitable in the face of escalating energy costs may lean even harder on IT to achieve energy efficiencies in the years to come, InfoWorld reports. But instead of limiting IT's efficiency role to the datacenter, companies will tap IT's vast knowledge of company networks, equipment, and work processes to uncover efficiencies across the organization, in some cases tipping facilities management into IT. 'There is a lot IT can do to fix its own 2 percent [of the company's carbon emissions] and make it more efficient, but the big opportunity for IT is to take a leadership role in tackling that other 98 percent across the business,' says one analyst. And by taking charge of the organization's energy strategy now, IT will be in prime position to alter its relationship with management and reap benefits in the boardroom in the years ahead, analysts contend."
[+]
Hardware: 10 IT Power-Saving Myths Debunked 359 comments
snydeq writes "InfoWorld examines 10 power-saving assumptions IT has been operating under in its quest to rein in energy costs vs. the permanent energy crisis. Under scrutiny, most such assumptions wither. From true CPU efficiency, to the life span effect of power-down frequency on servers, to SSD power consumption, to switching to DC in the datacenter, get the facts before setting your IT energy strategy."
This discussion has been archived.
No new comments can be posted.
The Fine Print: The following comments are owned by whoever posted them. We are not responsible for them in any way.
Full
Abbreviated
Hidden
Loading... please wait.
Only when it makes sense (Score:5, Funny)
Like when someone posts your domain name on slashdot!
You can't take down a server that's already off-line.
Re:Only when it makes sense (Score:5, Funny)
You can't take down a server that's already off-line.
Nuke the entire site from orbit.
It's the only way to be sure.
Parent
The moment when.... (Score:4, Funny)
you see the Windows logo appear? (sorry, couldn't resist)
Re:The moment when.... (Score:5, Funny)
Give it 30 seconds and it'll do it on its own (sorry couldn't resist either)
Parent
Simple (Score:5, Funny)
The best time to shut down the servers is right before you quit your job. Password-protecting the BIOS first adds value too.
Re:Simple (Score:5, Funny)
Password-protecting the BIOS first adds value too.
A real pro puts epoxy on the BIOS battery & any motherboard jumpers.
If I can't have this server, neither can you
Parent
Re:Simple (Score:5, Funny)
A real blackhat flashes a custom BIOS with the password set, so pulling the battery just resets the BIOS back to his default password.
Parent
WOL (Score:5, Insightful)
Re:WOL (Score:5, Funny)
Mod parent redundant.
Parent
Business needs and Risk (Score:5, Insightful)
It's pretty much up to your business....what must run 24/7, what systems are likely to get accessed in off hours, and how likely is that, and how critical are they? With redundant systems, can there be any downtime while they are powered up, or should it be immediate failover? If you use virtualization the redundancy should be easier to manage in many cases...you may be able to immediately offload to running systems and power up backup systems and then bring the VMs up there.
It's hard to get very specific without knowing your business and what you are running and what the needs are.
Re: (Score:3, Interesting)
Your correct.
A lot of places I know shut completely down at night but leave the servers up and running. Often it is so they can run end of night jobs or just so they can get up and running quickly in the morning.
A lot of it is just waste and a lot of it is just habit.
Now for people that run 24/7? That is totaly up to you.
Re:Business needs and Risk (Score:5, Interesting)
Parent
I go by a few simple rules... (Score:5, Funny)
I'm glad this was posted to "Ask Slashdot" where your audience is highly seasoned professionals that can give you wise, insightful answers...
In the data center that I manage, I use a few simple rules to determine when I power them down.
1) If the server is on fire
2) If there are no users using the server
or
3) If the power company is saying that I haven't paid my bill and they are sending "Hank" over to cut me off
4) Civil unrest, tornado, earthquake, zombies, etc.
Re:I go by a few simple rules... (Score:5, Funny)
I'm glad this was posted to "Ask Slashdot" where your audience is highly seasoned professionals that can give you wise, insightful answers...
In the data center that I manage, I use a few simple rules to determine when I power them down.
1) If the server is on fire
2) If there are no users using the server
or
3) If the power company is saying that I haven't paid my bill and they are sending "Hank" over to cut me off
4) Civil unrest, tornado, earthquake, zombies, etc.
Zombies aren't a good reason for shutting down the servers, that's why our IT guy keeps a shotgun leaned up against the server....at least he says it's for zombies.
Parent
Virtualize! Virtualize! Virtualize! (Score:5, Interesting)
If you virtualized your servers, you could create a managed power-down/power-up scenario. In the morning, your servers would turn on, your virtualized instances would move around (so they have more power for the day's activities), and then at night they'd retreat to a smaller group of servers. The unused servers could shut down for the night. You could even rotate which servers stay on overnight keeping the virtual servers running to spread the wear around if there is some.
Re:Virtualize! Virtualize! Virtualize! (Score:5, Informative)
There are a number of tools and products out there to assist this.
Consider a large (65k+ employees) company that has a several hundred server implementation that they use to process payroll every two weeks. They use a management tool to power them up on Friday, process payroll over the weekend, and shut them down on Monday. The power and cooling cost impact of these several hundred servers *not* running most of the month (6 or so days a month instead of 31) is huge.
Another (and also in use by the same company) strategy is to virtualize the OS instances, spin those up and down as necessary, and then use something like VMWare's VMotion to maximize usage of the physical boxes - and again use another tool to power down unneeded compute capacity.
Welcome to the virtual world...
Lots of prerequisites, but when it works, it's pretty freakin' sweet...
Parent
Re:Virtualize! Virtualize! Virtualize! (Score:5, Informative)
...and again use another tool to power down unneeded compute capacity.
And that other tool is ... VMware! DPM (distributed power management) is built right in, and does exactly what you describe.
http://www.vmware.com/products/vi/vc/drs.html [vmware.com] (scroll to the bottom)
Welcome to the virtual world...
Yup, the game is officially changed.
Parent
Re:Virtualize! Virtualize! Virtualize! (Score:5, Informative)
Actually, the other tool in this case is Cisco's VFrame Data Center. The problem with DPM (and other VMware tools) is that they won't let you move a physical box between ESX clusters. If you have multiple ESX clusters, the physical machine stays with it - powered up or not. With VFrame, the system can be powered down, removed from the cluster, and added to another if/when necessary... including any necessary network configuration (VLAN memberships, etc) and SAN configuration (zoning changes, LUN masking).
Not that I'm complaining about VMWare's solution to this problem - they're actually quite complimentary.
Parent
Re:Virtualize! Virtualize! Virtualize! (Score:5, Interesting)
First, let's assume "several hundred" equals 200, and we have exactly 65000 employees. Let's also assume that these extra servers are on for exactly 48 hours. Let's also assume perfect load balancing and distribution of the process over the servers.
That means that each server processes payroll for 325 employees in 48 hours, or about 7 employees per hour. So, each of these servers is basically the equivalent of a Commodore 64 in computing power. I suggest that the best way to save money at this task is to replace the 200 servers with a single Pentium 4 quad core running at 3GHz.
The other explanation—that the software is so unbelievably bad that it really does take 8½ minutes for it to run a single employee—is possible, but would going out and buying "QuickBooks" really cost more than the 200 servers to run this awful beast of a payroll program?
Parent
Re:Virtualize! Virtualize! Virtualize! (Score:5, Interesting)
Parent
Re:Virtualize! Virtualize! Virtualize! (Score:4, Informative)
And given that there's more to processing payroll than printing out a check for each employee (calculation of taxes in every state in the US, every country in the world, etc), updating various accounting systems, etc... there's a lot of work to be done besides "take 8 1/2 minutes for a single employee".
Parent
Re:Virtualize! Virtualize! Virtualize! (Score:5, Funny)
First of all, I'm with you, I also don't understand what it is about these mythical accounting processes that takes so damn long to process.
I guess it's like everything else in the software industry:
- software built by programmers for programmers runs quickly
- software built by programmers for non-programmers is incoherent
- software built by non-programmers for non-programmers is slow as molasses
- software built by non-programmers for programmers is never executed!
Parent
Re:Virtualize! Virtualize! Virtualize! (Score:4, Insightful)
But what happens to all of the servers that fail to start up in time to process payroll? It's late? You pay overtime through the nose for th SysAdmins that have to come in and work 24 hour days to bring the machines online? Seriously, I'm not saying it's a bad idea but I would say that this scenario is probably more like 15 days on 15 days off. You have to build in time on the front end to make sure the machines are up and running in a stable configuration, and probably time on the back end to apply patches and perform metrics on the machines to make sure they are running properly for next month. I'm not sure that this would save anyone any money in the long run because of the load on their staff during spin-up.
Parent
Re:Virtualize! Virtualize! Virtualize! (Score:5, Insightful)
Wait, you know of a place where sysadmins are paid overtime? Or even by the hour?
Parent
Like a car... (Score:4, Interesting)
Re: (Score:3, Insightful)
It's not 03:00 everywhere on the planet nor is it sunday either.
Re: (Score:3, Insightful)
You're assuming the web servers are for an international service.
Re:Like a car... (Score:5, Insightful)
If it's a website, it's international, wether or not the actual products and services are available internationally.
Parent
Re: (Score:3, Insightful)
He didn't say website, he said "web servers" which, despite the name "web", could serve an internal web-based service to a large company that only exists in one region of the world.
Re:Like a car... (Score:5, Funny)
Hola Señor.
Me gusta lo que dices, y estoy interesado en suscribirse a su boletín informativo.
Parent
Re:Like a car... (Score:4, Insightful)
Excepting google and such, I doubt that the vast majority of servers would have such a geographically balanced workload.
Parent
Re: (Score:3, Insightful)
Google likely shunts load to different datacenters based on location.
Simple Answer (Score:5, Insightful)
When you're sure you don't need it to come back up.
Not often (Score:5, Insightful)
How many of us have servers that don't need to be live? Yeah, I guess there might be a development server, but that assumes that you're not developing. There could be a failover server that does nothing when the primary hasn't failed, but in that case you'd want to be damn sure that the failover will come online without difficulty when it needs to.
It seems to me like it would be a pretty rare case when this is applicable. I'd sooner be interested in asking, can they build servers that can selectively power down subsystems that aren't currently in use, sufficiently enough that there's no serious harm. For example, I'd consider putting some of my fileservers' hard drives to sleep over night, but I'd still want the server to be available and the drives to spin back up if I log in from home and need access.
Mostly, I'd say that if you have servers that you don't need to be live, you might not be using your servers efficiently. It may be worth looking into setting up some kind of VM server with various images that can be brought up on command. But hey, if you do have a server that you can turn off without causing problems, go for it.
Re:Not often (Score:4, Interesting)
Uh I am the administrator of a server that archives all email for our company. We no longer use this solution for our email archiving, but according to federal regulations this email needs to be accessible for at least another 26 months. The only people who use the server anymore are the various alphabet soups of regulators who came in twice a year, maybe I'm the exception but not the rule but I can't see a reason to keep the server on...
Parent
When they're not being used (Score:3, Interesting)
Why have 16 terminal servers (sorry, couldn't think of anything else) running when no more than 10-20 users are on it after working hours? Then in the morning, power them back on again using WakeOnLan.
And that backup server with a whole lot of disks? Why not only have it running during the night when stuff is being backed up?
Power Management (Score:3, Interesting)
Some servers, like the HP ProLiant line, has power management features [hp.com]. Try experimenting with features like these first.
Virtualization solves that (Score:3, Insightful)
"Servers that sit in idle state for long periods of time are the top candidates for powering down between uses."
Then virtualize it or combine its function with another server. I see this part of the article as a bad example. It starts by saying that virtualization has helped, and then uses an example that virtualization would solve, NOT power-cycling.
Maybe its just me, but when I think of a server, I think of something important that is running, that needs to be accessible on something other than a glorified desktop. If it is important, then it cannot be turned off.
XenServer (Score:4, Interesting)
Wrong way round (Score:5, Insightful)
Some criteria (Score:4, Informative)
1: Can your service be load balanced across several identical servers?
2: Does your services experience predictable but varying load?
3: Can the state used by your service be rapidly replicated (10 minutes) across newly booted systems?
Not all server systems make good candidates for shutdown. Web farms do tend to because they fit the criteria above.
PSU failures (Score:5, Informative)
The problem is the PSU, which fails most often during power-up. Leaving the servers always on has the advantage of avoiding that particular failure mode. Also, other components in the server are prone to failure during power-up, way more often than at steady state. So, powering up your computers is overall a risky moment.
VMWare can be used for this (Score:3, Informative)
VMWare has some cool functionality such that if you virtualize all your machines, at night time when the loads are lower, you can consolidate all your VMs onto a smaller number of physical machines, and automatically turn off the physical machines. Then, in the morning, as the loads increase, you can automatically power on the physical machines and move the VMs back onto these physical servers to handle the load. Not sure what it's called but when I heard about it, I thought it was really cool.
colo (Score:4, Insightful)
Well the most obvious use for power down.... (Score:4, Funny)
... is with electronic voting servers, to force paper ballots and more accurate counting.
Re:Well... (Score:5, Insightful)
What kind of UPS does that? If the batteries are already charged, what would it be doing with the power that's not consumed by the devices - does it also act as a space-heater?
The ratings for UPS's - and any other power supply - are peak loads, if the UPS is being used at 100% capacity.
Parent
Re:Well... (Score:5, Funny)
(My datacenter charges about a buck a VA)
A watt?
Parent
Re:Well... (Score:4, Funny)
I personally think a buck for a Virginia would be a great deal.
On a side note, that seems pretty pricey for a watt - turning on my blow dryer in the morning would cost me $1500 where you live. I'd have to go to work with wet hair everyday.
Ooooh - I could instead use solar energy to dry my hair (stand outside in the sun for a while). Green is good!
Parent
Re:Old gear (Score:5, Interesting)
Analogy (this time, light bulbs).
- Light bulbs fail just as you turn them on. Or off.
- They hardly ever fail whilst switched on.
I think servers are the same. You're in trouble if a server you've had switched on for two years and forgotten about loses power and doesn't come back up. If it'd been switched off every weekend it would have failed earlier -- but probably at a more convenient time.
Parent
Re:Old gear (Score:4, Interesting)
Light bulbs fail when they're turned on because there's a warmup spike as the electricity flows through the wire. This causes the tungsten to heat up more than it will when the circuit is complete, and break the circuit.
If your lightbulb fails it's because the filament has worn down - it's just that you usually find out about it when you turn the light bulb on or off.
Also, your analogy means nothing. Servers are nothing like light bulbs.
Parent