Edward Ned Harvey <[email protected]> writes: > > From: Luke S Crawford [mailto:[email protected]] > > > > http://www.supermicro.com/Aplus/system/1U/1042/AS-1042G-TF.cfm > > Incidentally, Luke (or anyone else) ... Do you own any of these systems? I > guess I don't care if it's Supermicro, or any other brand. But the > supermicro link is here.
> As long as it's AMD processors. I haven't got a quad-socket box, but I have some single socket and dual socket boxes. I've got one of these that hasn't hit production yet: http://www.supermicro.com/Aplus/system/1U/1012/AS-1012G-MTF.cfm (and two more that have) also, I've got a dual socket system using the following motherboard: http://www.supermicro.com/Aplus/motherboard/Opteron6100/SR56x0/H8DGU-F.cfm that uses the same power supply (and otherwise is almost exactly like) http://www.supermicro.com/Aplus/system/2U/2022/AS-2022G-URF.cfm which will be going into production within the next few days. All of these are filled with 8 core 115w chips (this is a cost optimization... if you are willing to pay more than 2x as much, you get 12 cores for the same power draw.) I've got a kill-a-watt (and several remote PDUs, which is what I normally use, but as we're doing this on pre-production equipment, I can use the kill-a-watt to make the test the same as yours.) How many disks do you have in the server when you measure? or do you want to do the measurement without disks? > Personally, I think this should be done for every server, ever. I see too > often, that other admins overload UPS's or circuit breakers, or waste money > by overprovision cooling or UPS's. All of these are bad scenarios, and all > of them are easily avoidable for almost no cost, in terms of time or money. I think it's also important to monitor power load in real time. (I mean, testing for max draw before you put things in production as you suggest is /essential/ but I think it's also important to see things in real time. It picks up mistakes like plugging the new server into the wrong pdu, and covers your ass if your 'fully loaded' test wasn't as good as you thought.) generally speaking, when I'm paying for power by the circuit (many data centers charge you per circuit, regardless of utilization) I attempt to keep a circuit a tad under 75% utilization. 75% is where my alarms go off (and it's the max recommended sustained draw on any circuit.) > For the heck of it, here are my "basic instructions," that I think should > always be habitually completed. > > I use a kill-a-watt. When the server is new, I write a 3-line python > script, which is an infinite loop of random number generation. I launch as > many of these simultaneously as required, to fully max out all the CPU's. I > measure both the VA and W consumption of the server, and record it all in a > spreadsheet. What's the three lines? I want to use the same three lines you do. my understanding is that fully maxing out a cpu is somewhat more complex than that... you want to exercise more of it's instruction set, thus burnmmx and related programs, though I don't know how relevant burnmmx is to a modern CPU. Now, a random number generator is a lot better than nothing, and probably 'good enough' for what we're doing... but if we are comparing, I want to use what you use. > I assume it's 120V, so you know the A by knowing the VA. I > keep track of which servers are plugged into which UPS, and how many A are > available in the circuit feeding the UPS. (I also measure the "charging" > current of the UPS.) I always fluff everything by about 20%. And I > estimate approx 3 BTU's cooling are required per W. I'm smaller than you are, apparently. The cost of cooling is rolled into the cost of my power. Also, as far as I can tell, adding a UPS, when you are already at a datacenter with a good UPS and generator doesn't add anything besides something else to fail > Many times before, I have also maxed out the disk, memory, or network > utilization, and consistently find that the idle power consumption equals > the fully active power. It's only the CPU or GPU that seems to vary the > power consumption of the box significantly. I believe that heavy seeks vary the power used by the disk by a good bit. (sequential transfers, not so much. It's waggling the read-write head about, as far as I can tell, that varies the power draw.) -- Luke S. Crawford http://prgmr.com/xen/ - Hosting for the technically adept http://nostarch.com/xen.htm - We don't assume you are stupid. _______________________________________________ Tech mailing list [email protected] http://lopsa.org/cgi-bin/mailman/listinfo/tech This list provided by the League of Professional System Administrators http://lopsa.org/
