Le mercredi 06 octobre 2010 à 14:37 +0200, Florian Forster a écrit : > It'd be awesome if you could test my changes and tell me whether the > plugin now works as expected or if further changes are required. I'll > merge the branch as soon as you give me the thumbs-up ;) >
Hi Florian, I'd be grateful if you could postpone the merge of this plugin for a few days. I recently spotted something which might be worth investigating. On IBM pSeries machines there is a "feature" called Capacity on Demand (CoD), which in short consists in IBM logically disabling some of the installed physical processors, and then charging you for unlocking the extra processing capacity if you happen to need it. Using the phys_cpus_pool number could be misleading in this case, as you might think by looking at the graphs that you still have plenty of power left, when in reality some of the processors you're considering are disabled and unusable. I am looking for a way to take this into account. I am wondering whether pool_max_time doesn't do just this, ie indicate the activated/online/available/paid processing power instead of the physically present cpus... By the way this could also be the cause of the difference between the results when calculating from pool_busy_time vs from pool_idle_time! I will come back to you shortly, probably next week, as soon as I complete a few tests... Regards, Aurélien Reynaud _______________________________________________ collectd mailing list collectd@verplant.org http://mailman.verplant.org/listinfo/collectd