On Sat, Sep 09, 2023 at 10:17:56PM +0000, Andrew Doran wrote: > On Fri, Sep 08, 2023 at 12:27:57PM +1000, Paul Ripke wrote: > > > I need to read more code to see when the pools decide to release idle > > pages - because this is remarkably wasteful considering my machine is > > also paging, and "only" has 16GiB RAM. > > > > Memory resource pool statistics > > Name Size Requests Fail Releases Pgreq Pgrel Npage Hiwat Minpg Maxpg > > Idle > > buf2k 2048 830780 0 617706 207284 85623 121661 172466 0 inf > > 6948 > > vcachepl 640 44075984 0 43266390 5228141 5012754 215387 704585 0 inf > > 3547 > > pcgnormal 320 23415927 32666190 23388721 1050250 1045247 5003 89205 0 > > inf 2425 > > ffsdino2 264 43386098 0 42621716 1690136 1577576 112560 281810 0 inf > > 2185 > > ffsino 280 43312409 0 42548027 1850682 1733882 116800 301939 0 inf > > 2143 > > ractx 40 6888877 0 6846710 41237 37592 3645 4679 0 inf > > 1780 > > bufpl 280 557611 0 239663 31776 559 31217 31296 0 inf > > 1362 > > buf16k 16896 431232 0 350158 52798 21935 30863 32706 0 inf > > 942 > > mclpl 2048 120600 0 119876 32540 31337 1203 5270 0 > > 261333 841 > > pvpage 4096 61134 3 53585 40271 32334 7937 12454 0 inf > > 388 > > mbpl 520 205550 0 204337 13673 12993 680 3049 0 inf > > 362 > > pcglarge 1088 844473 0 825312 88124 81378 6746 9809 0 inf > > 359 > > buf4k 4096 28406 0 25833 5510 2654 2856 4001 0 inf > > 283 > > ... > > Totals 1937415183 32666228 1927493251 11060670 10142826 917844 > > > > In use 3585418K, total allocated 5569236K; utilization 64.4% > > The targets are roughly 10% of RAM for vnodes and 15% for bufcache which > adds up to 25%. So this doesn't seem so totally off the mark. DIAGNOSTIC > does add a bit of overhead since it messes up all the kmem cache sizes with > debugging stuff.
Ah, DIAGNOSTIC! I haven't looked into it's implementation, and hadn't realized it messes with this. I see a couple of my systems with DIAGNOSTIC have the unusual patterns. > I have plans to reduce the overhead from vnodes but not enough to claw back > 100s of MB.. Sometimes I think the method for sizing these caches should be > non-linear or something - don't have a good idea yet. Yeah, caches are hard. Memory management is hard. I see this enough at $DAYJOB. -- Paul Ripke "Great minds discuss ideas, average minds discuss events, small minds discuss people." -- Disputed: Often attributed to Eleanor Roosevelt. 1948.
