Date: Mon, 03 Oct 2011 20:33:55 +0200 From: Attila Nagy <bra@fsn.hu> To: Artem Belevich <art@freebsd.org> Cc: freebsd-fs@freebsd.org, Adrian Chadd <adrian@freebsd.org>, delphij@freebsd.org Subject: Re: is TMPFS still highly experimental? Message-ID: <4E8A0013.4070008@fsn.hu> In-Reply-To: <CAFqOu6jEoH8oGpy6i_7emUfrNvSG1jtypYza6jgCWVvXB%2Bh4RQ@mail.gmail.com> References: <CAOfDtXMm9K_fbOmvG2gvXxDfKakkgpPt9MLifqDxa4wCibMExg@mail.gmail.com> <alpine.GSO.1.10.1110011122030.882@multics.mit.edu> <CADLo83-s_3H8PbbxOPPxbe0m10U0U5JW-feB294dFs%2BQ3iTWvg@mail.gmail.com> <CAGMYy3ssi%2BkAuufDTHA1z6u7jRrZwRRkCCkcO94GHNGF9Rku_w@mail.gmail.com> <20111002020231.GA70864@icarus.home.lan> <4E899C8E.7040305@fsn.hu> <CAFqOu6jEoH8oGpy6i_7emUfrNvSG1jtypYza6jgCWVvXB%2Bh4RQ@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On 10/03/2011 04:58 PM, Artem Belevich wrote: >> For me, the bug is still here: >> $ uname -a >> FreeBSD b 8.2-STABLE FreeBSD 8.2-STABLE #5: Wed Sep 14 15:01:25 CEST 2011 >> root@buildervm:/data/usr/obj/data/usr/src/sys/BOOTCLNT amd64 >> $ df -h /tmp >> Filesystem Size Used Avail Capacity Mounted on >> tmpfs 0B 0B 0B 100% /tmp >> >> I have no swap configured. The machine has 64 GB RAM. >> vm.kmem_size=60G; vfs.zfs.arc_max=55G; vfs.zfs.arc_min=20G > I'm curious -- does your ARC size ever reaches configured limit of > 55G? My hunch that it's probably hovering around some noticeably lower > number. Yes, in some minutes. Current counters: kstat.zfs.misc.arcstats.c_min: 21474836480 kstat.zfs.misc.arcstats.c_max: 59055800320 kstat.zfs.misc.arcstats.size: 45691792856 > On my ZFS setups a lot of memory seems to be lost due to > fragmentation. On a system with 24G of RAM and rac_max=16G, I > typically see more than 20G of memory wired. > With kmem_size=60G, ARC is likely to use up most of available kmem > space and that's probably what affects tmpfs. Besides, with kmem_size > that close to arc_max you may be risking meeting "kmem too small" > panic, though, considering that your kmem_size is rather large chances > of that are smaller than on a system with smaller amounts of memory > and kmem_size. Sounds plausible. BTW, it may be possible that the ARC limits are not needed anymore, they are here from the times, where on a 64 GB machine ARC hovered around 2-5 GBs without setting these (arc_min was even higher then). BTW, the user space programs fit into around 1-2 GB RAM on this machine typically. Well, most of the time. :) > I'd start with doubling kmem_size and, possibly, reducing arc_max to > the point where it stops putting pressure on tmpfs. > I know there are several differences, but it would be very good to have similar behaviour with UFS. I guess it's quite evident that tmpfs can eat the file system cache, and I know it may be not so trivial to solve this with ZFS. :) Will try it, thanks.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?4E8A0013.4070008>