ZFS and Wired memory, again

[Matthias Gamsjager]

I remember reading something similar couple of days ago but can't find the
thread.


On Tue, Jan 28, 2014 at 7:50 PM, Anton Sayetsky <vsjcfm at gmail.com> wrote:

> 2013-11-22 Anton Sayetsky <vsjcfm at gmail.com>:
> > Hello,
> >
> > I'm planning to deploy a ~150 TiB ZFS pool and when playing with ZFS
> > noticed that amount of wired memory is MUCH bigger than ARC size (in
> > absence of other hungry memory consumers, of course). I'm afraid that
> > this strange behavior may become even worse on a machine with big pool
> > and some hundreds gibibytes of RAM.
> >
> > So let me explain what happened.
> >
> > Immediately after booting system top says the following:
> > =====
> > Mem: 14M Active, 13M Inact, 117M Wired, 2947M Free
> > ARC: 24M Total, 5360K MFU, 18M MRU, 16K Anon, 328K Header, 1096K Other
> > =====
> > Ok, wired mem - arc = 92 MiB
> >
> > Then I started to read pool (tar cpf /dev/null /).
> > Memory usage when ARC size is ~1GiB
> > =====
> > Mem: 16M Active, 15M Inact, 1410M Wired, 1649M Free
> > ARC: 1114M Total, 29M MFU, 972M MRU, 21K Anon, 18M Header, 95M Other
> > =====
> > 1410-1114=296 MiB
> >
> > Memory usage when ARC size reaches it's maximum of 2 GiB
> > =====
> > Mem: 16M Active, 16M Inact, 2523M Wired, 536M Free
> > ARC: 2067M Total, 3255K MFU, 1821M MRU, 35K Anon, 38M Header, 204M Other
> > =====
> > 2523-2067=456 MiB
> >
> > Memory usage a few minutes later
> > =====
> > Mem: 10M Active, 27M Inact, 2721M Wired, 333M Free
> > ARC: 2002M Total, 22M MFU, 1655M MRU, 21K Anon, 36M Header, 289M Other
> > =====
> > 2721-2002=719 MiB
> >
> > So why the wired ram on a machine with only minimal amount of services
> > has grown from 92 to 719 MiB? Sometimes I can even see about a gig!
> > I'm using 9.2-RELEASE-p1 amd64. Test machine has a T5450 C2D CPU and 4
> > G RAM (actual available amount is 3 G). ZFS pool is configured on a
> > GPT partition of a single 1 TB HDD.
> > Disabling/enabling prefetch does't helps. Limiting ARC to 1 gig doesn't
> helps.
> > When reading a pool, evict skips can increment very fast and sometimes
> > arc metadata exceeds limit (2x-5x).
> >
> > I've attached logs with system configuration, outputs from top, ps,
> > zfs-stats and vmstat.
> > conf.log = system configuration, also uploaded to
> http://pastebin.com/NYBcJPeT
> > top_ps_zfs-stats_vmstat_afterboot = memory stats immediately after
> > booting system, http://pastebin.com/mudmEyG5
> > top_ps_zfs-stats_vmstat_1g-arc = after ARC grown to 1 gig,
> > http://pastebin.com/4AC8dn5C
> > top_ps_zfs-stats_vmstat_fullmem = when ARC reached limit of 2 gigs,
> > http://pastebin.com/bx7svEP0
> > top_ps_zfs-stats_vmstat_fullmem_2 = few minutes later,
> > http://pastebin.com/qYWFaNeA
> >
> > What should I do next?
> BUMP
> _______________________________________________
> freebsd-fs at freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-fs
> To unsubscribe, send any mail to "freebsd-fs-unsubscribe at freebsd.org"
>