John, > I may be wrong, but it is not that the cached pages are needed by this job or > the last job. > It is that the last job has caused buffer cache to be used, which fills up > memory on one or more NUMA nodes. > When the new job starts it allocates memory - if it finds memory is full on > one NUMA node you will get NUMA misses, > which lead to slower memory access.
That rings a bell. Like many things to which I've become habituated, I've lost their origins in the mists of Time. :) Being a consulting benchmarker, I would be happy to help anyone evaluate and increase the efficiency of their application load. > Me, I come place myself in the camp of one job / one node (or one job / lots > of nodes) A fan of predictable performance, eh? :) Regards, Max _______________________________________________ Beowulf mailing list, [email protected] sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
