On Wed, Feb 11, 2009 at 01:34:23AM +0100, Henrik Nordstrom wrote: > To get an even better trace the best is to build Squid with valgrind > support (--with-valgrind-debug), let it run for a while under valgrind > and then view the memory statistics page. > > squidclient mgr:mem
Thanks, that's a neat trick. It turned out the Valgrind reports were pretty bogus, though; I sent Amos a graph obtained with pprof which was a lot more reasonable. (Turns out Valgrind isn't so good in tracking down blocks that are still reachable, for some reason.) Basically it's something in the delayed read system that's causing problems; you can have a look at http://home.samfundet.no/~sesse/pprof9897.0.ps It's easiest read with gv. Let me know if there's anything in the profile that's confusing :-) /* Steinar */ -- Homepage: http://www.sesse.net/