Slashdot Mirror
How Internet Data Centers Waste Power
Rick Zeman writes "The New York Times has extensively surveyed and analyzed data center power usage and patterns. At their behest, the consulting firm McKinsey & Company analyzed energy use by data centers and found that, on average they were using only 6 percent to 12 percent of the electricity powering their servers to perform computations. The rest was essentially used to keep servers idling and ready in case of a surge in activity that could slow or crash their operations. 'Worldwide, the digital warehouses use about 30 billion watts of electricity, roughly equivalent to the output of 30 nuclear power plants.' In other words, 'A single data center can take more power than a medium-size town.' This is the price being paid to ensure everyone has instant access to every email they've ever received, or for their instant Facebook status update. Data Center providers are finding that they can't rack servers fast enough to provide for users' needs: A few companies say they are using extensively re-engineered software and cooling systems to decrease wasted power. Among them are Facebook and Google, which also have redesigned their hardware. Still, according to recent disclosures, Google's data centers consume nearly 300 million watts and Facebook's about 60 million watts. Many of these solutions are readily available, but in a risk-averse industry, most companies have been reluctant to make wholesale change, according to industry experts."
I have no idea how the URL got mangled when Timothy moved the anchor text to a different part of the article, but here's the correct link:
http://www.nytimes.com/2012/09/23/technology/data-centers-waste-vast-amounts-of-energy-belying-industry-image.html?hpw&pagewanted=all
Or use a mainframe running lots of Linuxes... Can cut the power to 10% while delivering the same computing power. Mainframes have a very good power management this days.
in this letter and comment.
The most ironic point: "Should we discover (as we undoubtedly would) that tens of thousands of copies of today's NYT were printed, delivered, and sold to subscribers who never read Glanz's report, do we conclude that the NYT needs a new and less-wasteful business model?"
Worst case, if we just include first world people, it's only about a 100W per person
Rough engineering estimate, a watt continuously is a buck per year.
For commercial I'm completely unimpressed. That's like the depreciation on my desk and chair, or the dept "free" coffee budget for a month. A tiny fraction of the overhead lighting power, which is a tiny fraction of the HVAC power, which is a tiny fraction of my salary. In terms of environmental degradation, the gasoline I burn to commute is worse than my share of the corporate data center (based on a KWh being about a pound of coal, so 16 pounds of coal per week, and commute four times per week is about 4 gallons or about 24 pounds of gasoline)
For residential I'm amazed. They need to make $100/yr off my mom who doesn't even have internet access just to pay the electrical bill. I donno if they can make $100 of me per year and I'm always on the net doing "stuff". One interesting comparison WRT advertising is "one million page views per year = one thousand dollars per month or about a penny per pageview". Donno how true that is anymore. But it would imply that just to pay the electric bill the average person would have to visit 27 web pages per day, every day, which seems pretty high across an entire nation.
"Science flies us to the moon. Religion flies us into buildings." - Victor Stenger
That's already how many datacenters do it. Still, that capacity takes about 2-10 minutes to come up to speed so you still need somewhat of a buffer. What they need is instant-on servers which the big guys are experimenting with but the problem is not Google or Netflix or Facebook, they run a pretty efficient operation, it's the rest of the business world who'd rather buy an IBM or HP honking piece of metal that converts 20% of it's power to heat before anything remotely useful has been done than experimenting with what they need and could do to improve on such designs.
Custom electronics and digital signage for your business: www.evcircuits.com
A server is a sort of bulked-up desktop computer, minus a screen and keyboard, that contains chips to process data.
Schroedinger's Brexit: The UK is both in and out of the EU at the same time!
Google's numbers are especially tame. 300 million watts (total) is far below one watt per user (gmail alone has at least 350 million accounts). Certainly if you use Google services on your 30-watt laptop, you use more power than Google uses to serve you. According to Google, "in the time it takes to do a Google search, your own personal computer will use more energy than Google uses to answer your query."
Since Google offers almost all services for free, it has a strong incentive to minimize resources per user. I expect the paid services are the ones that use the bulk of the energy, but all data centers together are still a tiny fraction of total worldwide power usage.
This is lovely. Let's worry about problems that don't exist, as if we don't have enough catastrophes to worry about.
Power is money. As long as there is a somewhat unhampered economy in the locus of data centers (and there is), then every entrepreneur will attempt to economize power usage. You don't have to worry about it because the entrepreneurs that use power efficiently will eat the lunch of those that do not, ceteris paribus (all other things equal).
Ipso facto this problem will solve itself. Case closed.
In fact, now that I speculate on the possible reasons for publicity like this to be drummed up, it is to campaign for government regulations that will instruct entrepreneurs how they 'must' handle such a problem. Unfortunately nobody can write such regulations because they cannot foresee every circumstance and possibility, much less predict the future. Nobody on this earth can even tell a single other person what ideal type and amount of preparation is for power efficiency considerations. This is why we have economic calculation.
If such regulations are enacted, ipso facto they will cause the problem itself.
Liberty.
I'm part of the problem. Wish I wasn't, but I don't seem to have any choice.
I run a small web site, and if it goes down, there are various consequences in my personal and professional life that can be extremely annoying and embarrassing. To stay sane, I need the site to have good uptime. Over the years, this has caused me to gradually migrate to more and more expensive webhosting, now ~$100/mo.
The average load on my dedicated server is extremely low, so it's basically like one of the extremely wasteful boxes described in TFA. My site is basically I/O-intensive: I serve big PDF files. In terms of CPU, I'm sure the site would run fine on a low-end ARM, or as one of a dozen sites running off of the same Celeron chip. So by comparison with either of those hypothetical, energy-efficient setups, virtually all of the electrical power is being wasted. I'm a small fry, but there are millions of sites like mine, so I'm sure it adds up. (It would be interesting to know how much of total server-center power consumption comes from the "long tails" of the distribution such as Google and Facebook, and what percentage from cottage industries like me.)
There are basically two problems. (1) Nobody will sell me high-reliability webhosting on low-end hardware. The only way to get energy-efficient hardware is to get cheap webhosting. I've tried cheap webhosting. Cheap webhosts have low reliability and nonexistent customer service. (2) Sometimes you get spikes in demand, and you want some excess capacity to be able to handle it without crashing the server. Maybe you get slashdotted. Actually, in my case one thing that has been a problem is that some people apparently run IE plugins that are supposed to accelerate large downloads, by opening multiple connections with the server. When these people hit my server and download a large PDF, the effect is very much like a DOS attack. My logs show one IP address using 300 Mb of throughput to download a 3 Mb PDF. I've written scripts that lock these bozos out ASAP, but on a low-end machine, these events would bring my server to its knees instantly.
Find free books.
Our energy supply is finite, and so our energy usage should be measured in units of energy, not dollars.
Prices are not based on market forces or total costs, they are based on government policies.
And our money supply itself is schizophrenic, as in disconnected from reality. It's value fluctuates by moods, it's continually debased by printing more, it's backed only fractionally, and then only by the good faith and credit of future taxes on today's kindergartners
Measuring energy with dollars is like scoring sporting events by the applause of drunken fans.
It's value fluctuates by moods,
Sort of how people decide to use apostrophes.
I don't really understand this hostility. I read the New York Times online, everyday. I don't get a paper delivered to my door. Those few, those happy few who actually read this new york times article, read it online.
The circulation is a million pulp, half a million online.
This is one major reason that companies (even very large companies with "money to spare") are moving towards virtualization with incredible speed.
I'm not going to go digging for numbers right now, but the statistics show that something like 100 percent of Fortune 100 companies use virtualization, and perhaps 85-90% of Fortune 500 companies.
The larger virtualization solutions will actually take the servers that are idle, migrate them to another host machine, and power down/suspend the "extra" machine(s) that was/were being used during their core business hours.
Virtualization also allows for spikes in cpu/network, and then can take that power back when everyone goes home (a print server, an intranet web server, a domain controller, etc). So, physical machines actually DO get turned off when they aren't being taxed, and with more and more "software defined networking" the interconnects between systems can be scaled and moved also.
Now, I don't know how the big players are using this (e.g. Amazon, VMware, Rackspace, Google). I can't see inside their datacenters, but one would think that something like AWS would have a huge stake in saving power by turning off idle instances and moving VMs. Not only for the power savings from the server directly, but for the (approx) 30-40 percent more energy that it takes to cool the physical machines.
It's also worth noting that larger companies are putting their datacenters in areas with plentiful (cheap) power. Places like Washington state, with hydroelectric power and a cooler average ambient temperature, allow for a huge savings on power right off the bat. Add things like dynamic scaling of server and network hardware, lights-out datacenters, and better designed cooling systems (look at Microsoft's ideas), and there is a huge power savings across the board.
How much energy does the NYT use to print paper copies of the newspaper, distribute and deliver them, harvest the trees and process the paper? Now compare that with the energy that the online NYT uses. Which allows for more people to view the publication for less energy? I'm positive that it is the electronic version.
Grandpa: My Homer is not a communist. He may be a liar, a pig, an idiot, a communist, but he is not a porn star.
This is the same "problem" that faces airline companies, taxi drivers, power companies, cell network operators. Consumers pay for these services by usage and so total revenue is proportional to average use but the costs are heavily skewed towards capital costs and so are proportional to the peak load that you can service. In that case, there's a fundamental tradeoff -- either we have to degrade service when demand hits the 95th percentile (just as an example) or we have to figure out a way to pay for the extra capital investment that's not needed 95% of the time.
There's a few alternatives you can do:
(1) Overprovision and soak it up into the price structure for all consumers. This is what most power companies do -- they build enough power generating capacity for peak load and then charge a bit more per KWH to make up for the increased outlay.
(2) Overprovision and charge extra at peak. This is the airline solution -- they always have service available but under contention the last few seats are exorbitantly expensive. Essentially those that need peak service are paying to leave a few seats open all the time in case they need them.
(3) Don't overprovision: this is the taxi solution. This means that service degrades significantly under peak demand -- anyone trying to get a cab home on a Saturday night in a major city has experienced this. Those that do get a cab pay the usual fare, everyone else waits around a while. This is also the solution that California has routinely deployed for their inability to provide peak power during heat spells -- same price for everyone but rolling blackouts for the unlucky few.
That's it -- there aren't any clean answers when you are making compromises between peak availability and average efficiency. You've either got to pay for the extra capacity when you don't need it or else you have to suffer when you don't have the capacity when you do need it.
that was the whole point of this article, you stupid twat.
Why are you worried about benchmark scores on servers that typically only run computations 12 percent of the time?
You people eat up artificial gimmicky numbers like nothing. It's amazing.
I think the problem is that while you can run 10,000 linux instances on a single mainframe and maybe it can keep them all chugging along at 12% load (though it seems like it would take a rather sizable mainframe to be equivalent to 12% of 10,000 or 1200 standalone servers), but when your peak load comes and those linux servers that are nearly idle all night long are suddenly 80% utilized, can the mainframe keep all 10,000 instances running along at 80% utilization?
And can it do it more cheaply than on VMWare and Intel? You'd need around 300 4 socket 8 core CPU Intel servers to handle 10,000 instances using up one core each of CPU power, figure around $10M for the cluster and 10 - 15 racks -- can you build the same mainframe for $10M in less space?
I really don't know the answer.
I had a visceral reaction to the article. This is because they are pointing out the obvious and then pretending they are performing some kind of public service and pat themselves on the back.
Do they really think the data centers don't know these things? Do they really think they are not trying to address them? Power costs are pretty high up on the balance sheet and anyone who's been paying attention knows there are millions of dollars spent on researching ways to bring those costs down.
So it's kind of like a guy standing at a car wreck watching the rescuers trying to pull someone from a car and saying, "if you don't get that guy out of there, he's gonna die". No Shit Sherlock.
Just shows that reporters are idiots. Always have been and always will be.
When Fascism comes to America, it will call itself Anti-Fascism, and tell you to give up your guns.