HAMMER in real life

[Matthew Dillon]

:Matt used to use hardlinks for some sort of historical arrangement; after
:a certain point, the total number of hardlinks was too much to handle.  He
:might have mentioned this somewhere in the archives.  I don't know if this
:would bite you the same way with gmirror.
:

    Here's a quick summary:

    * First, a filesystem like UFS (and I think UFS2 but I'm not sure)
      is limited to 65536 hardlinks per inode.  This limit is quickly
      reached when something like a CVS archive (which itself uses hardlinks
      in the CVS/ subdirectories) is backed up using the hardlink model.
      This results in a lot of data duplication and wasted storage.

    * Since directories cannot be hardlinked, directories are always
      duplicated for each backup.  For UFS this is a disaster because
      fsck's memory use is partially based on the number of directories.

    * UFS's fsck can't handle large numbers of inodes.  Once you get
      past a few tens of millions of inodes fsck explodes, not to mention
      can take 9+ hours to run even if it does not explode.  This happened
      to me several times during the days where I used UFS to hold archival
      data and for backups.  Everything worked dandy until I actually had
      to fsck.

      Even though things like background fsck exist, it's never been stable
      enough to be practical in a production environment, and even if it were
      it eats disk bandwidth potentially for days after a crash.  I don't
      know if that has changed recently or not.

      The only work around is to not store tens of millions of inodes on a
      UFS filesystem.

    * I believe that FreeBSD was talking about adopting some of the LFS work,
      or otherwise implementing log space for UFS.  I don't know what the
      state of this is but I will say that it's tough to get something like
      this to work right without a lot of actual plug-pulling tests.

    Either OpenBSD or NetBSD I believe have a log structured extension to
    UFS which works.  Not sure which, sorry.

    With something like ZFS one would use ZFS's snapshots (though they aren't
    as fine-grained as HAMMER snapshots).  ZFS's snapshots work fairly well
    but have higher maintanance overheads then HAMMER snapshots when one is
    trying to delete a snapshot.  HAMMER can delete several snapshots in a
    single pass so the aggregate maintainance overhead is lower.

    With Linux... well, I don't know which filesystem you'd use.  ext4 maybe,
    if they've fixed the bugs.  I've used reiser in the past (but obviously
    that isn't desireable now).

    --

    For HAMMER, both Justin and I have been able to fill up multi-terrabyte
    filesystems running bulk pkgsrc builds with default setups.  It's fairly
    easy to fix by adjusting up the HAMMER config (aka hammer viconfig
    <filesystem>) run times for pruning and reblocking.

    Bulk builds are a bit of a special case.  Due to the way they work a
    bulk build rm -rf's /usr/pkg for EACH package it builds, then
    reconstructs it by installing the necessary dependencies previously
    created before building the next package.  This eats disk space like
    crazy on a normal HAMMER mount.  It's more managable if one did a
    'nohistory' HAMMER mount but my preference, in general, is to use a
    normal mount.

    HAMMER does not implement redundancy like ZFS, so if redundancy is
    needed you'd need to use a RAID card.  For backup systems I typically
    don't bother with per-filesystem redundancy since I have several copies
    on different machines already.  Not only do the (HAMMER) production
    machines have somewhere around 60 days worth of snapshtos on them,
    but my on-site backup box has 100 days of daily snapshots and my
    off-site backup box has almost 2 years of weekly snapshots.

    So if the backups fit on one or two drives additional redundancy isn't
    really beneficial.  More then that and you'd definitely want RAID.

					-Matt
					Matthew Dillon 
					<dillon at backplane.com>