The supply of electrical power is something that has enormous impact on the design of a data centre. Today’s 1U-high servers seem to suck up electricity as if it were going out of fashion, and their heat output seems to get bigger and bigger. I remember well my visit to the Microsoft SQL Server test lab at Redmond, where they maintain a large data centre that runs almost completely hands-off – all the test loads, OS images and so forth are remotely controlled by the management computers. This technique allows them to soak-test a huge range of product combinations, such as the Serbo-Croat version of SQL Server running on French Windows XP Service Pack 1, if they feel the need. But when I was there, I noticed that a number of the racks were only partially populated. When I asked why, the explanation came back that they’d run out of electrical power into the building, because the 1U servers were taking as much power as the old 2U servers. No more power meant adding no more servers, and also no more air-conditioning either.
Of course, if you’re a large corporation, there are serious electrical engineering people who can and should be helping you to design your data centre, and they’ll be taking responsibility for specifying the power and air-con requirements – which is, after all, why you go to experts such as Arup for their expertise. Things are a little different in the SME marketplace, though. Power consumption measurements are rarely even considered – servers are simply plugged into a suitably sized UPS and left to soak-test (and I can confidently say a “suitably sized UPS” merely because one that’s too small will just trip out and fail to work).
Rarely is the issue of heat production given proper consideration either, especially when it comes to the matter of redundancy. A rather too obvious solution to this problem at a client I visited recently was to plug the air-con unit into the same UPS as the server, in an attempt to ensure that both can keep running in the event of a power failure. Such charming naivety only came home to roost when the power surge caused by the air-con compressor starting up caused the server to reboot, which in a really helpful fashion happened sporadically about once a week or so. There was nothing in the server logs that might indicate the source of the problem, and it was only the loud “whump” made by the air-con unit as it engaged its main pump, plus playing a quick game of “follow the mains lead”, that enabled me to present them with the required explanation.
It can be very useful to actually measure the power consumption of your server racks and cooling solutions but, unfortunately, measuring power by using a voltmeter and ampmeter will result in completely bogus numbers because of the matter of phase shift. Power only equals volts times amps if those volts and amps are in phase. If they’re seriously out of phase, your calculation may well look correct on paper, but it won’t represent the real power usage at all. This is called power factor.
To measure power consumption correctly, you need a proper power meter. In my lab, we use the excellent Voltech PM100, which is, as you might have guessed, reassuringly expensive. However, there’s no quibbling with the results it produces. Integrating the power consumption over a half-hour period will enable a true picture to be gained, because it catches events such as extra fans switching on and off as the hardware cycles up and down in temperature. I also like to make use of the internal temperature-reporting tools within the server management software to keep an eye on what’s going on.