Slashdot Mirror
Cooling Challenges an Issue In Rackspace Outage
miller60 writes "If your data center's cooling system fails, how long do you have before your servers overheat? The shrinking window for recovery from a grid power outage appears to have been an issue in Monday night's downtime for some customers of Rackspace, which has historically been among the most reliable hosting providers. The company's Dallas data center lost power when a traffic accident damaged a nearby power transformer. There were difficulties getting the chillers fully back online (it's not clear if this was equipment issues or subsequent power bumps) and temperatures rose in the data center, forcing Rackspace to take customer servers offline to protect the equipment. A recent study found that a data center running at 5 kilowatts per server cabinet may experience a thermal shutdown in as little as three minutes during a power outage. The short recovery window from cooling outages has been a hot topic in discussions of data center energy efficiency. One strategy being actively debated is raising the temperature set point in the data center, which trims power bills but may create a less forgiving environment in a cooling outage."
Other publications have noted it was number 3, too.
DT
Is this thing on? Hello?
Well, back to rejecting software patent applications.
If you want 100% uptime (which is impossible, but you can put enough 9s in your reliability to be close enough), you need to have your data distributed across multiple data centers, geographically separate, and over provisioned enough that the loss of one data center won't cause the others to be overloaded. It's important to keep your geographical separation large because you never know when the entire eastern (or western) seaboard will experience complete power failure or when a major backhaul router will go down/have a line cut. Preferably each data center should get power from multiple sources if they can, and multiple POPs on the internet from each center is almost mandatory.
I read the internet for the articles.
Believe it or not, but in one of those "life coincidences", pi is a safe approximation. Take the number of watts your equipment, lighting, etc., use, multiply by pi, and that's the # of btus of cooling. Don't forget to include 100 watts per person for body heat.
It'll be 90F degrees outside, and you'll be a cool 66F.
Kevin Smith on Prince
I've never understood why data centre designers haven't used a different cooling strategy to re-circulated cooled air. After all, for much of the temperate latitudes for much of the year the external ambient temperature is at or below that needed for the data centre so why not use conditioned external air to cool the equipment and then exhaust it (possibly with a heat exchanger to recover the heat for other uses such as geothermal storage and use in winter)? (Oh, and have the air-flow fans on the UPS.)
The advantage of this is that even in the worst case scenario where the chillers fail totally during mid-summer there is no run-away, closed loop, self re-enforcing heat cycle, the data centre temperature will rise but it would do so more slowly and the maximum equilibrium temperature will be far lower (and dependant upon the external ambient temperature).
In fact, as part of the design for the cluster room in our new building I've specified such a system, though due to the maximum size of the ducting space available we can only use this for half the heat load.
Agrajag: "Oh no, not again!"
A few weeks ago the A/C dropped out in one of our computer rooms. I like the resulting graph: http://leebert.org/tmp/SCADA_S100_10-3-07.JPG
Most large refrigeration compressors have "short-cycling protection". The compressor motor is overloaded during startup, and needs time to cool. So there's a timer that limits the time between two compressor starts. 4 minutes is a typical delay for a large unit. If you don't have this delay, compressor motors burn out.
Some fancy short-cycling protection timers have backup power, so the the "start to start" time is measured even through power failures. But that's rare. Here's a typical short-cycling timer. For the ones that don't, like that one, a power failure restarts the timer, so you have to wait out the timer after a power glitch.
The timers with backup power, or even the old style ones with a motor and cam-operated switch, allow a quick restart after a power failure if the compressor was already running. Once. If there's a second power failure, the compressor has to wait out the time delay.
So it's important to ensure that a data center's chillers have time delay units that measure true start-to-start time, or you take a cooling outage of several minutes on any short power drop. And, after a power failure and transfer to emergency generators, don't go back to commercial power until enough time has elapsed for the short-cycling protection timers to time out. This last appears to be where Rackspace failed.
Dealing with sequential power failures is tough. That's what took down that big data center in SF a few months ago.
(Disregarding your blatant karma whoring by replying to the top post while changing the subject)
There's several good reasons why the servers are located where they are, and not, say, in Alaska.
The main one is light speed through fiber, and a cable from Houston to Fairbanks would induce a best case of around 28 ms latency, each way. Multiply by several billion packets.
This is why hosting near the customer is considered a Good Thing, and why companies like Akamai have made it their business of transparently re-routing clients to the closest server.
Back to cooling. A few years ago, I worked for a telephone company, and the local data centre there had a 15 degree C ambient baseline temperature. We had to wear sweaters if working for any length of time in the server hall, but had a secure normal temperature room outside the server hall, with console switches and a couple of ttys for configuration.
The main reason why the temperature was kept so low was to be on the safe side -- even if a fan should burn out in one of the cabinets, opening the cabinet doors would provide adequate (albeit not good) cooling until it could be repaired, without (and this is the important part) taking anything down.
A secondary reason was that the backup power generators were, for security reasons, inside the server hall themselves, and during a power outage these would add substantial heat to the equation.