Too Hot for Humans, But Google Servers Keep Humming | Hacker News Reader

Too Hot for Humans, But Google Servers Keep Humming

datacenterknowledge.com

118 pointscd3482 comments

This is a very interesting case. As an architect though I can't help putting in my 2 cents that cooling isn't simply a binary choice.

Firstly, the best thing to do environmentally would be to use the heat positively, connecting it into a centralised heat-distribution network for other users (for instance, other industries, public heated pools, or even residences).

Secondly, even without power, there is a lot you can do to maximise the heat-loss of a building: Eg. design to capture prevailing winds and utilise them to help heat escape, or internal arrangements to keep cooler 'microclimates' around areas used by maintenance staff and temperature-critical server elements. I am not sure what the design process of these buildings is, but I would bet that this sort of design doesn't get thought about.

Thirdly, you can also actively cool buildings without 'chillers' (refrigeration units), such as pumping earth-cooled water through the building (as 'chilled beams'). This is actually likely to be quite cost-effective for large projects like this, and much more environmentally friendly than burning coal.

Facebook's Prineville datacenter is aligned to capture prevailing winds:

http://www.facebook.com/prinevilleDataCenter

Somewhere I had whitepapers from the Open Data Center Initiative which did touch on some of the issues you mentioned. http://www.opendatacenteralliance.org/ is the alliance page, but, requires membership, has a few whitepapers and webcasts available.

Why would you bet that professionals with huge budgets candidate cost-conscious customers wouldn't think about how to solve the problem they're paid to solve? If they don't, you could be a billionaire by selling your ideas.

I think the big clients (google, facebook, et. al) are getting smarter very quickly now (see for instance cd34's link above). It also requires not just one person, but a whole team of smart architects, engineers and environmental experts to figure this out for every individual server centre.

Also - some of the big successes (heat sharing, using water from outside sources, etc.) also require government involvement, which can be a minefield in itself (note: google seems to have much better success with these government schemes in Europe than in the USA).

> During these periods, the temperature inside the data center can rise above 95 degrees.

:-/ 95 degrees what? They don't use Fahrenheit in Belgium, you know (nor does most of the rest of the world, for that matter).

For anyone wondering, 95F is exactly 35 degrees Celsius.

Also, it would probably have been more relevant to mention the equivalent Belgian and EU institutes instead of the OSHA ... seriously, I don't expect for a second that the regulations would be the same. Roughly, maybe, but it just doesn't make sense to mention the OSHA at all if the servers aren't in the USA!

As someone who spent his first pair of decades in the USA, it's easy to forget that most Americans view the whole world as if it were the USA. It's a habit that seems normal in that environment.

They also frequently say "the world" when they mean "the whole country".

Is it just me, or did this article strike a dystopian tone? Computer centers of the future, perhaps achieving truly useful AI capabilities, consisting of unimaginable numbers of racks sweltering in a heat so hot that people can't survive in the same room with them? Sounds like something Dante would have written if he were alive in 2012.

I don't think so - to me, the tone seems informative: "hey, did you know that data centers might not need all that cooling? here's an example."

Plus, humans can work for about 80-90 minutes at close to 100 F (albeit not comfortably)[1], so we're not really talking about conditions that would kill people. One can extrapolate, of course, but it's quite possible that the servers would break down before the humans do.

[1] http://www.ncbi.nlm.nih.gov/pubmed/9183079

We also seem to get along fine in space, under water, in low oxygen environments, in irradiated environments, in freezing environments...

The point of this is: who cares if the data center is warm/hot? It's as hot as that inside a mascot suit at a theme park, so they build cooling mechanisms into them. Tools exist for a reason, and it is somewhat ironic that we prematurely assume that our creations require the same environment for "comfortable" survival that we do. The animistic tendencies we have may not just be cultural.

It's not like we get along fine in space et al because we put up with discomfort. Have you actually considered your examples? Space is survivable in part because the suit keeps you warm. Under water is much the same way; stay in the water long enough and you need a drysuit or wetsuit to stay alive. Freezing environments? You wear seal furs.

The point is, talking about how we can survive in freezing environments does NOT demonstrate that climate control (or rather, local microclimate control around your body) is just creature comfort.

I guess I didn't get to my point clearly. What I'm saying was "Who cares what the temperature in the data center is? We'll just put some guy in a cold suit if it is more efficient." Humans usually adapt when it is in their favor to.

I wonder how much money companies waste by keeping their server rooms as cold as a refrigerator. Some IT guy who thinks servers should be cold probably told them that.

For a long time it was common knowledge that computers last longer if you keep them cold. It's only been in the past few years that evidence has been shown that moderately high temperatures do not appreciably shorten equipment life.

It's also conceivable to me that it used to be more true in the past than it is now. Computer equipment used to be incredibly fragile.

Also, heat generated per server was rising until recently.

Since air conditioning cost is proportional to heat generated, it could have well made financial sense in the 80s and 90s to aggressively cool servers. Not nearly as much heat had to be displaced, and the cost of each server was higher.

From then, institutional inertia and infectious repititis (a brilliant term coined by Amory Lovins) kick in, and it just becomes The Way data centers are built.

Until a bunch of cool kids like Google and Facebook show up on the block, who don't care about the old rules quite so much, and they start to question the conventional wisdom.

Energy costs have risen quite a lot. Cooling is a lot more expensive.

Google (see links elsewhere in this thread) have a lot of equipment, and a lot of data about that equipment, and a lot of smart people to turn that data into information.

Did anyone else think of the scene in Sunshine where Chris Evans gets stuck under the computer that's being lowered into fluid and freezes to death?

The mechanism for making "the hot aisle" temporarily habitable fails, you get stuck on/in something and can't get out in time.... Life imitates art.

I think you're overstating the dangers a bit. A hot aisle might be uncomfortable and not a place you'd want to spend many hours, but it isn't going to kill you quickly. If by some series of amazing coincidences you managed to get stuck in one with no way to call for help, you could just start unplugging servers until someone came to investigate and rescue you.

speaking as someone who has been inside a "hot" datacenter of a few thousand servers (after a failed ac system which didn't trip the alarms for a few hours), uncomfortable does not really begin to cover it. The extremely dry, hot air was quiet difficult to breathe, and being inside for more then a couple of minutes at a time was pretty much impossible. Definitely not the same experience as sitting in a hot aisle. It has been a number of years, but I believe the thermometers read close to 100F at the time.

If I'm not mistaken, humans can live in 100F dry air non-stop (with enough water to drink, of course). 100F is about human body temperature.

Could it be that hot datacenters have some poisonous gases evaporated into the air?

I wouldnt be worried about more than a few minutes. Millions of people live in climates where it gets above 100F for a significant part of the year. I used to go for hour long runs over lunch in Phoenix during the summer. Typically 100-107F. Dry/arid is actually preferable to humid ("yeah, but it's a dry heat").

I'm not saying that 100+F degree air is not livable, just relaying non-medical personal experience. I'm not doubting there were other factors, other then the heat, but it was more then just unpleasant. I've been to arizona in the summer (though I'll admit, I didn't go for long runs), but being inside that datacenter was considerably worse. Could have been the shock/drastic temperature difference though, I suppose.

Of course, in "hot" data centers today the environment is no more dangerous than a nice sauna, but this development makes me imagine what it might progress to in 10 or 20 years. This is the first small step across that threshold between "comfortable for humans" and "comfortable for machines" in data centers, and the first such step is usually the hardest one. If it's more efficient to run the servers in an even hotter environment, or in one that's less-habitable in some other way, you can probably expect it to happen, now that the baseline technology and ideas are here.

Not that I'm trying to seriously forecast anything, or have any special knowledge of data center design; this is just a pleasant exercise in near-future sci-fi.

http://www.youtube.com/watch?v=xeJ0LLsr3sM

Chris Evans in Sunshine dies as he gets stuck under a computer tower which is being lowered back into its freezing cold fluid.

Years ago a data center outside DC in Silicon Alley had an A/C failure. Our equipment was reporting temps of 140F on intake. We were allowed 10 minutes in, one person doing work, another observer, both required to have squirt water bottles and a two-way radio. We were instructed not to carry anything and to use a cart. We were told not to remove equipment that weighed more than a few pounds.

I pulled 10 drives so we could run them to another data center as a 'just in case'. The guy I was with passed out about 60' from the door. The air temp outside was ~72F and they had portable hurricane fans blowing air into the data center.

In that ten minutes, I drank 32 ounces of water. I remember that because I went to pick him up, kicked my water bottle by accident and it fell over because it was empty.

He spent 30 minutes in an ambulance, about three hours in a hospital from 10 minutes exposure. It took me about two hours to recover to the point where I felt I could eat.

Don't lose respect for heat. :)

You guys deserve hazard pay. :)

Saunas have 70C - 100C temperature.

http://en.wikipedia.org/wiki/Sauna

You had only 60C (140F).

Fainting after just 10 minutes in 60C is too fast.

BTW, most likely reason your friend spent time in ambulance and hospital is because your employer wanted to be on the safe side, not because your friend really needed it.

I'm reasonably sure the air temp wasn't 140F in the data center, but, he had Heat Stroke or Hyperthermia. He did collapse and I remember calling for help on the radio when I saw him go down. In the space of about 10 seconds when he said he was feeling weak, he went from looking flushed to pale as a ghost and bam, he was on the ground. He was wheeled off in a stretcher and I remember sitting in the back of the ambulance with him. After 30 minutes, he was still not feeling well and reluctantly accepted the hospital visit.

He and I were business partners and this was one of our management contracts. It was ultimately his decision with my strong recommendation to go to the hospital since he was still weak. The paramedics offered ten more minutes on site, but, he opted for the hospital.

We were young and dumb, now I'm not so young. Faced with the same situation today, I doubt I would do it again and I certainly wouldn't force an employee to do it.

I'm glad you replied to the comment. Heat exhaustion (leading to heatstroke) is serious stuff.

It really happens - even to young people at their peak condition. In dry climates, with mildly hot weather around 90-93 F.

Perhaps I'm stating the obvious, but awareness makes a big difference in these situations. It doesn't hurt to have a little EMT / Paramedic training too.

In fairness, that's not the whole story. When I sauna, I'm not doing anything remotely strenuous. I'm sitting around naked and not feeling the slightest stress. These guys were working in an emergency situation - clothed - and doing physical labor.

I agree: stress from handling emergency situation + physical labor definitely contributed.

But just 10 minutes of it still shouldn't have such serious effect.

Yet it did have that effect. Unless you're claiming the parent is lying? Or are you suggesting it's merely an unlikely event? The fact is people have all kinds of physical issues they might not even be aware of and you can't take that chance with people - not for legal reasons, not for ethical reasons. The best you can do is prepare in advance by training people in the situation to detect if they have any issues with the environment.

The air in a sauna is very very humid, right? And the air in a server room will typically be very very dry.

In a hot dry climate, you do lose water much faster.

I don't think that's right.

When overheated, you sweat. Sweat leads to evaporation, which leads to your skin cooling.

In humid environments, our sweat does not evaporate as well. This leads to less cooling, which leads to more sweating and thus more water loss.

Anyone that's been in a sauna can verify this. High humidity humidity feels hotter. You can demonstrate this by throwing water onto the sauna heater - the sauna will instantly feel a lot hotter.

Hmm... but what happens when you get dehydrated, which happens faster in a dry environment?

Yes: in dry climate you lose water much faster, so you cool off. So you should not faint that fast.

But may be it was mostly due to stress...

Are you actually claiming that the GP is lying? He's an eye witness, was instrumental in getting his buddy out. I'm pretty sure that if he felt his friend didn't need to be in the ambulance he would have omitted the bit about him passing out.

Not lying. Most likely some relevant details are unintentionally omitted. It could be that his buddy had poor health or was too overweight. Or may be was quite stressed out even before going into that hot facility. Or stayed there longer than 10 minutes. Or the temperature was higher. Or the air was slightly toxic (that's actually my best guess). Or all of the above.

But 10 minutes exposure to 60C dry air is not nearly enough for fainting.

Several hundred thousand people from the US have recent experience working in 120-140F weather wearing 50-100 pounds of equipment and heavy clothing, often doing labor a bit more strenuous than pulling drives or racking servers.

Disneyworld in Orlando Florida with an average summer temperature of 91F responds to dozens of heat stroke cases with people wearing shorts and shirts, merely walking around and waiting in lines to enjoy rides.

It is extremely humid in Orlando. Evaporative cooling through sweating doesn't work very well in those conditions.

Yeah, there's a wide variation in training, condition, and the level of care/equipment/hydration, and the results (and there are lots of heat injuries in the military, too). The OSHA limits are clearly not the limits of human performance (either low or high). One of the things is you acclimate to it over a period of a week or two fairly successfully, so if there were a datacenter which constantly operated in that range, you should be able to get staff who can handle it. (I'm sure steel mills and other industrial facilities have high temperatures, too).

If I read your comment correctly, it sounds like you're comparing healthy young people who have undergone extensive physical training and testing to technicians of a highly variable age who have gone through no such selection and training? Or are you really suggesting that any old technician can run around in a desert carrying military equipment with zero training or physical aptitude, or vice versa?

I thought he was suggesting hiring Iraq veterans to manage data centers.

The heat in Iraq goes up to 120F, and some soldiers have been hospitalized, or even killed by the heat. In the extreme case of 140F heat, I'd imagine soldiers would only be allowed out in if there's a very good reason. >120F heat should be treated like a sandstorm - an environmental hazard which you try to avoid.

Fairly certain he is talking about fire fighters folks.

>>> weather

I'm not.

I... don't see you participating in this conversation in the first place? So ok...

Well regardless, firefighters have plenty of experience with running into hot as balls buildings and doing physically exerting things. If he wasn't referring to firefighters, then he should have been.

I was definitely thinking of Iraq because I'd say national guard are a pretty broad cross section of regular Americans, but firefighters, steel mill or industrial workers, etc are all highly relevant too.

Relatively extreme temps are an issue, but can be overcome. If it requires a couple weeks of practice for your staff to get used to working in 100-120f low rh environments but saves huge amounts of energy, it is worth considering.

In a cloud hosting environment, you could eventually have non safe temperatures in active operation, lots of redundancy, and just do all your human maintenance during reduced power/temp windows all at once. Sort of like treating the data center as the inside of a computer. Either run multiple facilities or partition by thermal suite or hae the equipment power management be able to throttle cores for reducing heat.

When Ramadan is during the summer, observant Muslims in hot Middle Eastern countries still have to work, oftentimes in the heat, and fasting includes no water either. Maybe it's because they're used to it, but the point is, 140 degrees is certainly doable with water, even if it's not a lot of fun.

The hottest recorded temperature on earth was 136. Average highs in the Middle East aren't anywhere near 140.

He may not have been very eloquent, but he wasn't saying that Middle East temps reach 140.

(He was saying ME Muslims do x degrees without water, so 140 with water should be tolerable. X is anywhere from 90 to 120+ in certain ME countries).

fasting includes no water either

One of the strongest principles of Islamic law is that necessity overrides prohibition. If you're working outside in severe heat and becoming dehydrated, you're allowed to -- and arguably required to -- drink water, Ramadan or no Ramadan.

I used to work at the CSTJF in Pau, France which is one of Total's biggest cluster (if not the biggest). Their data-center was ranked 91 in the top-100 at the time.

In the event of a fire, they have (had?) a system that would release a gaz in the data-center that would "eat" oxygen so that the said fire would die.

Obviously any living creatures trapped in the data-center would die as well. The evacuation time before release was perhaps something like 30 to 60 seconds (I can't remember that). Just enough to drop anything and reach one of the safe corridor where no gaz is released.

Halon 1301 is the gas that used to be used which was banned in the 90s. Most inert gas fire suppression systems use HFCs or Ketones now and have a ten minute hold time. They work by displacing the oxygen and reducing it to <10% which is not enough to support flame - or life.

In '86, I was in a data center that had expanded to take up 70% more space. During the expansion, emergency lighting hadn't been put up. During expansion, a 470 volt three-phase line had been put in, the insulation had been stripped off while they were pulling it through the conduit. As it heated up, gas or vapors from the melted insulation built up and the 1/4" thick conduit was ripped open with roughly an 8" hole and an explosion people on the first floor heard (computer room was in the basement).

One of the operators panicked, tripped over the anti static carpet, ripped the Fire Suppression box off the wall and pulled the Halon. He stumbled out the automatic sliding glass doors and hit the EPO leaving the data center completely dark with me about 120' from the door in pitch black.

I had walked those corridors between the drives hundreds of times, but, had difficulty finding the path. Time was running out, so, I stood on one of the minicomputers, moved one of the drop ceiling tiles and stuck my head into the return air plenum and the Halon went off. It was noisy, sort of a giant whoosh with a low rumble and a very odd smell.

I don't know how long I stood there, but, it wasn't more than a few minutes before a flashlight was shining on me with a Fireman in full gear.

After that, we had emergency lighting and oxygen tanks throughout and the Halon system was upgraded. When the tanks were refilled, they did the math and noticed that the amount of Halon we had wasn't enough for the expanded data center and probably wouldn't have extinguished a fire.

Most humorous moment - seeing firemen walk through the data center with axes in hand looking for fire.

I hope the data center owner had to pay a hefty fine for almost killing you.

maybe it is worth keeping an emergency flashlight on you always. One of those small led flashlight that give like 100 lumens or so.

Thoes come free with an andoid phone with physical flash.

Slower to turn on than a real one.

Not by much, and I always know where my phone is. Not necessarily true of a torch that I'm not used to carrying.

Related: I had to get out of a pitch-black room on Thursday, when the IR-triggered lights flicked off because I'd been in the room too long.

So I downloaded a torch.

I love the 21st century.

15 minutes prior, I was working on the mux cabinet that would have put my back 18 inches from that conduit.

Good 'ole Racal Milgo equipment - where we had ~24* terminals on a 9600 baud line and we were happy to have that much bandwidth.

After that incident, there were a lot of changes to the data center, all safety related. It was a combination of an office building that was repurposed, a county government that was expanding very quickly, computer equipment that was taking more room and a government contractor under pressure to keep things running while the expansion was going on.

I don't believe the county government fined itself, but, things did change. It was the only building other than the courthouse in another city where we didn't run plenum rated cable. After the incident, all cable we ran was plenum rated.

It was the wild west back then. 143mb hard drives were the size of file cabinets, 67mb removable disks had 5 platters and weighed about 20 pounds. PCs were just starting to come into play, but, the mainframes and minicomputers took a lot of space. The tape library was easily 300 sq feet. When we needed more space, we asked the county, they would move whatever office group was adjacent to the computer room and the space would be expanded. There wasn't a lot of documentation at the time regarding best practices.

I never really felt that I was in danger, though, at 18, you sort of feel invincible anyhow. Had the computer room been built properly and not been converted office space, it wouldn't have had a drop ceiling and the outcome may have been different. Had the system been sized properly for the expanded room, perhaps things would have been different.

Edited: even though the mux we were using was a 32 channel mux, probably eight or more of the ports would have been 1200 baud for printers.

In the 1980s we had the fire suppressant system in our server room upgraded and the installers test fired a cylinder of Co2. Unfortunately, someone hadn't fitted the discharge 'shower head' properly on the end of the outlet pipe - which for some reason had a 'j' bend in the end and so the head shot up into the false ceiling, followed by all the gas, which pressurised the void and caused the ceiling to explode into the data centre.

Imagine the scene - shredded ceiling tiles and years of dust and grime everywhere! It really looked like a bomb had gone off. Fortunately there was no-one in the room during the test and everything kept running, but the clean-up took quite some time.

Ships engine rooms are required to have to separate CO2 release controls for exactly that reason. One releases the CO2 into a protected space and the other releases it into the protected space.

Did anyone read the last two paragraphs? Seems the article wasn't quite done being edited:

> Before entering the hot aisle, a technician uses a supply trigger, typically a switch located outside the hot aisle, to activate the SmartAire T units. Cool air then enters the hot aisle until a comfortable temperature is established. SmartAire T units maintain this temperature until the technician completes the assigned work and deactivates the units, eliminating any need for rest periods and increasing productivity.

> Before entering the hot aisle, a technician can use a supply trigger – typically a switch located outside the hot aisle – to activate the SmartAire T units. Cool air then enters the hot aisle until the temperature reaches a comfortable level.

I honestly have to wonder why we don't have robots doing the majority of these tasks. Obviously, a robot couldn't troubleshoot detailed or obscure problems, but for things like provisioning a server, rerouting around bad network hardware, replacing drives, NICs, etc. they should be good enough. CERN already uses robots for the LHC tape backup system. I can't see that this would be that much more complicated.

Humans are very easy to train compared to robots, and aren't that expensive. I'd imagine each facility is laid out slightly differently, with different generations of hardware. To make the robot work well you'd either have to program in each hardware variation or make the software sufficiently advanced to understand the differences itself and adapt. I imagine the cost/benefit wouldn't pan out when you look at how many datacenter techs you replace with how many software and hardware engineers.

Humans are great at adapting, so if you throw a new enclosure at them or a new motherboard design which has the CPU sockets in a different location, they'll be able to learn the new system in 30 minutes.

A former coworker of mine trained to work with factory robots, but he realized that there wasn't much of a current job market for it because humans are still way cheaper (and will be for some time). Instead he went into software.

I can't help but think about the logical extreme of this type of trend, which is to put the server rack itself outside- what components would you need to weather seal or protect against the elements? There are fans and disks that have moving parts, and cable connections you would have to seal, but other than those changes what would you have to do? Maybe it would be even more efficient....

One aspect of this approach they don't mention is how it affects the MTBF of the equipment. Many components suffer increased failure rates when run at a consistently higher temperature. While it may be relatively safe to run machines in a range higher than is comfortable for humans, I bet there's an upper limit past which components really start to fail in numbers.

The article mentions that "Studies by Intel and Microsoft showed that most servers do fine with higher temperatures and outside air, easing fears about higher hardware failure rates." So I think it isn't a big issue. Now, obviously there is going to be some upper limit, but it just hasn't been hit in this case.

We've published a few papers about this that studied (among other things) the effect of temperature on DRAM and hard-drive failure rates. We watch our machines pretty closely. If you're a data geek, Google's a fun place to be.

Disk Study: http://research.google.com/archive/disk_failures.pdf

DRAM Study: http://www.cs.toronto.edu/~bianca/papers/sigmetrics09.pdf

Very interesting! Thanks for the links, I'll check out the papers.

In the picture it sure does look like a row of evaporative coolers - there was no mention of them in the article - so maybe I'm wrong.

Free air cooling uses evaporative coolers. They are not running chillers. Belgium also has a water treatment center bringing in water from an industrial canal for their coolers and thus runs on 100% recycled water.

In Finland, they run heat exchangers with water from the Baltic Sea, again, a chiller-less data center.

Document from Google that talks about their cooling: http://www.google.com/about/datacenters/inside/efficiency/co...

I'm interested in how much "efficiency" can be built in to new data centres.

Is it possible to use the waste heat for other purposes? Would painting the roof white make any difference?

at one of BT's Datacetres one hot summer we played water hoses on the roof to increase cooling via evaporation

If it gets too hot for humans can't we use robots instead?

The other option is not to fix stuff, just leave failed stuff in place. Various people have done this, all depends on costs.

Hope that mindset does not come to pervade the nuclear power industry. :-)

These guys are whimps.

I used to work in an academic building with an ancient HVAC system in which temperatures in my office would soar to 105 degrees in April, between having no AC, sun shining in the big windows, plus the heat dissipation from 5 humans and 20 computers. My deck included a Dell Windows machine and two old IBM machines running Linux and a Linux laptop and a Sun Ultra 10 and a "Pizza Box" 32-bit SPARC machine and two Sun Rays.

(Even if they couldn't charge enough tuition to get a decent A/C system, at least I had access to a pool of last year's hardware.)

We never evacuated... You see, that's why the U.S. is #1 -- people in any other country would bail out at 95, but we stay the course. ;-)

Reply on news.ycombinator.com