On 10/27/2012 10:49 PM, Віталій Тимчишин wrote:
It can be that some query(s) use a lot of work mem, either because of high work_mem setting or because of planner error. In this case the moment query runs it will need memory that will later be returned and become free. Usually this can be seen as active memory spike with a lot of free memory after.
Yeah, I had briefly considered that. But our work-mem is only 16MB, and even a giant query would have trouble allocating 10+GB with that size of work-mem buckets. That's why I later listed the numa info. In our case, processor 0 is heavily unbalanced with its memory accesses compared to processor 1. I think the theory that we didn't start with interleave put an 8GB (our shared_buffers) segment all on processor 0, which unbalanced a lot of other stuff. Of course, that leaves 4-6GB unaccounted for. And numactl still shows a heavy preference for freeing memory from proc 0. It seems to only do it on this node, so we're going to switch nodes soon and see if the problem reappears. We may have to perform a node hardware audit if this persists. Thanks for your input, though. :) -- Shaun Thomas OptionsHouse | 141 W. Jackson Blvd. | Suite 500 | Chicago IL, 60604 312-444-8534 sthomas@xxxxxxxxxxxxxxxx ______________________________________________ See http://www.peak6.com/email_disclaimer/ for terms and conditions related to this email -- Sent via pgsql-performance mailing list (pgsql-performance@xxxxxxxxxxxxxx) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
