D&C Lug - Home Page
Devon & Cornwall Linux Users' Group

[ Date Index ][ Thread Index ]
[ <= Previous by date / thread ] [ Next by date / thread => ]

inodes and filesystem chat was Re: [LUG] User limits



Neil Stone wrote:
> 
> Yes, Linux can still run out of inodes... I fell foul of this a while
> ago.. and how annoyed was i when i found out...?

It's a filesystem thing. ext2 (and ext3 presumably) have static inode
allocations (usually plenty if you don't have lots of small files).

ReiserFS uses a dynamic allocation strategy, which is why I probably
haven't seen it in a LONG while (that and knowing to allocate extra for
certain types of filesystem activity when appropriate).

(I read the documentation a long time ago and decided to user Reiser on
new systems wherever possible for reasons that I've long forgotten and
which probably don't apply anymore anyway)

I guess "out of inodes" is the kind of error you only see once or twice
if you learn from your mistakes. But hey I'm capable of making this
mistake again<sigh>.

For various reasons been pondering the much underrated IDE RAID
solutions recently (some of which come with SCSI interfaces, it is the
disks that are ATA). Seems some good value arrays are around which can
easily keep your 120MB/s SCSI interfaces more than busy - fibre channel
anyone?

Those who bemoan ATA for this kind of thing seem to have forgotten what
the "I" in RAID is about ;-)

Led to an interesting discussion on some Linux kernel list about
write-back cache on disks, and journalled filesystems.

The gist of which is since many modern disks have write-back cache
without battery backup, and it is often enabled by default (hdparm is
your friend), lovingly journalled filesystems can still fail to write
stuff to disk safely during a power outage (or similar failure).

Not a problem for servers like the one I've been installing where the
SCSI disk's individual writeback cache is disabled, and a nonvolatile
128MB write-back cache is used in the RAID controller.

Let me not comment on the advice we received to use software RAID, but
it would be fun to benchmark it, and see how 'interesting' the results
compare, but I'm guessing with no writeback cache at all, dog slow is
all that is possible.

Do RAID in hardware where you can, although the support I'e had so far
from DELL on their PERC RAID stuff doesn't inspire me to recommend that
particular approach.

Makes me wonder whether I should use ext2 for my desktop systems, and
just sit out the "fsck"'s - although perhaps taking inodes into account
we should use ReiserFs and force more thorough consistency checks than
the default ;-)

The kernel people seems to be pondering if you are likely to get enough
warning to flush the caches, in the time it takes for the power to fail,
sounds dangerous to me.

The IDE disk on this box is too old for such refinements, so I'm safe ;)

Attachment: pgp00019.pgp
Description: PGP signature


Lynx friendly