On 30/12/13 23:28, James Salsman wrote: > Tim was asking about benchmark fairness, so here, read this: > http://armservers.com/2012/09/11/benchmarks-versus-the-real-world/
Yes, that seems pretty clear. They say that you can replace underutilised Intel CPUs with ARM CPUs, but agree with Intel's conclusion that if the CPU is fully utilised, Xeon is better than ARM in terms of performance per watt. Of course, there are other ways to deal with underutilised CPUs. For example, we have 16 memcached servers with 24 cores each, all with negligible CPU utilisation. They could have, say 4 cores each instead, "right-sizing the compute infrastructure" in Calxeda's lingo, which would greatly reduce the power requirements without the cost of deploying a new system architecture. Maybe if the workload was such that servers with 1 or 2 Xeon cores would still be underutilised, ARM would be worth a look. But we don't appear to have that situation at the moment, at least, not at a sufficient scale to warrant an investment of staff time. There are much larger inefficiencies that we don't have time to deal with. In eqiad, we have about 4700 cores running MediaWiki. Those are fully utilised except for essential headroom, so they wouldn't be appropriate for ARM, according to Calxeda's article. Neither of Calxeda's articles gives a figure for capital cost, so that a performance per dollar figure can be calculated, whereas Intel does provide that information. The obvious conclusion is that the cost is embarrassingly high. Calxeda only tells us that their server is cheaper and slower than the Intel one, they don't claim to have a lower capital cost for a given processing throughput. -- Tim Starling _______________________________________________ Wikimedia-l mailing list [email protected] Unsubscribe: https://lists.wikimedia.org/mailman/listinfo/wikimedia-l, <mailto:[email protected]?subject=unsubscribe>
