> From: [email protected] (Michael van Elst) > Date: Wed, 24 Jun 2026 16:28:25 -0000 (UTC) > > Yes. I'm using a similar patch but haven't found what knobs were > actually useful. But I have some other adjustments to help > memory usage.
That's great, I'm going to try your adjustments and see how they help. One comment, arc_average_blocksize and comressed_arc_enabled are read-only on FreeBSD (RDTUN). I'm also wondering, in sysctl_createv do we actually have to pass sizeof(var) in newlen for primary types? I am running some benchmarks on three systems, with 12, 8, and 2 gigs of ram, running MKUPDATE=no build while periodically running find -type f on all zfs filesystems and Greg's stress-zfs. I tried halving arc_max and _min for testing. In general, when running find, the metadata usage explodes irregardless of arc_max or meta_limit [2;3], but [1] setting arc_max works. And In all cases c is respecting the c_max limit. > From: Greg Troxel <[email protected]> > Date: Wed, 24 Jun 2026 13:44:51 -0400 > > It would be great to at least get sysctls to control arc size into the > tree, vs a number of copies uncommitted. I think arc_max, arc_min, and arc_free_target could be useful. There is a difference between Michael's style, rewriting entire functions for NetBSD, or my approach with many #ifdefs, I guess mine will be shorter but not sure what is better. S. [1]: https://lamprinidis.com/pub/zfs/2g_sys.png [2]: https://lamprinidis.com/pub/zfs/8g_sys.png [3]: https://lamprinidis.com/pub/zfs/12g_sys.png
