Well, I have searched a bit more about the zone_reclaim_mode thing, and obviously I am not the first one having problems with it. It seems, there was some kind of logic introduced to the kernel, which decided that our machine must be a NUMA architecture (which it is not), and as a consequence, decided to switch zone_reclaim_mode = 1. But now, it seems, that some people have discovered that this makes no sense with some of the modern processors, which - although - seem to be NUME from software ... are not NUMA. And thus, they want to go back to always set zone_reclaim_mode = 0 by default. Here is a lengthy discussion about the topic (so no need for me to start it again at LKML - and obviously it is not Ubuntu specific :-) http://lkml.org/lkml/2009/5/12/586 (and yes, our hardware has Core i7 CPUs which seem to make many kernels think it is NUMA - it is often mentioned in the LKML thread). Andras Fabian -----Ursprüngliche Nachricht----- Von: Craig Ringer [mailto:craig@xxxxxxxxxxxxxxxxxxxxx] Gesendet: Mittwoch, 14. Juli 2010 05:11 An: Andras Fabian Cc: pgsql-general@xxxxxxxxxxxxxx Betreff: Re: AW: AW: AW: AW: AW: PG_DUMP very slow because of STDOUT ?? On 13/07/10 22:16, Andras Fabian wrote: > I think I have found the solution. Yes, I now can get constantly high throughput with COPY-to-STDOUT, even if free -m only shows me 82 Mbytes (so no, this solution is not cleaning the cache). Always around 2 3/4 minutes. > > I have compared all the /proc/sys/vm settings on my new machines and the old machines (which never had problems), and of course found some differences, some new settings etc. (of course, lot of changes can happen between 2.6.26 and 2.6.32). And there was one, which stood out from the mass, because its name reminded me of some functions which I have always seen in the kernel stack while having congestion_wait: > > - zone_reclaim_mode > (yes, in the kernel stack there was always also a call to "zone_reclaim"). > > Interestingly, on the old machine this was set to "0" and on the new machine - obviously per Ubuntu default - to "1" ... What these all means is shortly described here: > > http://www.linuxinsight.com/proc_sys_vm_zone_reclaim_mode.html > > Then I though, lets give it a try. Nd I set it to "0" on the new server too ... and voila, it is running at high speed in COPY-to-STDOUT. I can even switch back and forth between 0 and 1 and see, how congestion_wait comes back or disappears. > > Now, someone with big kernel know-how could try to describe me in detail, what exactly could be at odds here. > > But for me it is now obvious, that I will put in my start up settings a change of "zone_reclaim_mode = 0". > > And tomorrow I will see how my nightly backup runs with this setting. It sounds like it's time for a post to the Linux Kernel Mailing List, and/or a Launchpad bug against the Ubuntu kernel. Make sure to have your asbestos undewear on if posting to LKML ;-) -- Craig Ringer Tech-related writing: http://soapyfrogs.blogspot.com/ -- Sent via pgsql-general mailing list (pgsql-general@xxxxxxxxxxxxxx) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-general
