Yesterday I got all 4 nodes of the DELL PowerEdge C6100 up and running. This is a beautiful machine. I am running the BOINC PrimeGrid project on all 4 nodes and have now got enough data to work out the next step in the upgrade.
I have been running PrimeGrid on a DELL PowerEdge T610 server. This amazing machine has a pair of 3.06GHz hex-core processors giving me a total of 24 logical processors (2 processors per core) and it has 48G of RAM (24G per processor). PrimeGrid doesn’t use any RAM worth talking about so it is down to the grunt in the processors. In addition you don’t want hyperthreading enabled running PrimeGrid so the processors are run at 50% giving me an effective 12 cores running the program. In addition – the T610 processors are run in turbo mode at 3.2GHz when running PrimeGrid – so this system is blistering along! Power consumption is 327W and the processors are running well within their temperature limits (as we are talking a server here used to doing heavy lifting). Each core runs through one of the Prime Number tests (a work unit) in around 35 minutes – so every 35 minutes the T610 chomps through 12 Prime Number candidates. These are the statistics for the DELL T610 that the DELL C6100 has to measure up to.
What do the numbers look like for the DELL C6100 as bought? Well the processors are a bit on the slow side as they are only 2.1GHz quad core Xeons, and there is only 1 processor per core, and 2 processors per node. So the computing grunt I have for PrimeGrid with this system is 8 cores per node running at 2.1GHz – I also have 6G RAM per processor, but as I said above the RAM requirements are minimal for running PrimeGrid. What is VERY nice for running PrimeGrid is that with 4 running nodes I have a total of 32 cores to throw at the problem, and these are running at 100% as there is no hyperthreading to worry about. As with the T610, since we are talking servers there is no temperature problem running these processors flat out and they run at 100% well within their temperature limits. The big problem is that it now takes 1 hour and 40-minutes for a core to chomp through a Prime candidate, compare this to the 35-minutes per core on the T610 – and you can see this is not good. And of course, if it takes so long to get through a work unit then the power consumption per work unit isn’t going to look too good either. I have no accurate figure for the overall power consumption of the C6100 with all 4 nodes running flat out, but I estimate we are talking around 800W in total, as the unit runs comfortably from an 1100W power supply. You can see the power consumption figures per work unit are not going to compare favourably against the DELL T610 even with 32 versus 12 effective cores. Plugging all the numbers in, the DELL T610 is costing me approximately 0.0136 kWhr per work unit, and the DELL C6100 is costing 0.111 kWhr per work unit – and that’s not very good. In fact it is very expensive computing.
It is obvious what I need to do to bring the DELL C6100 up to spec, I have to swap out the slow quad core processors for faster hex core processors. The added power consumption for carrying out this upgrade is only 15W per processor, or 120W overall. However, as I only have an 1100W psu fitted, I have purchased a beefier 1400W psu just to be on the safe side. I will swap out 2 sets of processors this month and then check how the system runs. If it all runs as expected I will swap out the other 2 pairs of processors next month.
I love this stuff – it’s now getting really interesting to see what the real cost of computing power turns out to be.