[LUNI] file system limitations
Steven Lembark
lembark at wrkhors.com
Wed Nov 20 04:32:01 CST 2002
-- Erik Lickerman <elickerman at ameritech.net>
> You make excellent points. Fortunately I have a little time to decide.
> Good backup is a must no matter which filesystem I use. I am using ext3
> in my development system. I'll figure it out.
You would do well to check the performance of ext3 with
legit numbers and distributions of files in development.
If this means borrowing a 200GB driveset for a week then
go for it. Things you need to test in this case would be
how to recover from the journal filling up under heavy
load, how to recover from a full jurnal if the system
crashes during a non-journalling period, and what the
peak disk load will look like with journalling (hopefully
on a separate drive).
Best arrangement for this would probably be LVM on hardware
RAID5 with a stripe == file page == disk page == 4096
(e.g., mkfs.extX -b 4096). This avoids the RAID5 write penalty.
If you can get multiple controllers you can also use mirroring
on the journal filesystem -- this cannot afford the speed penalty
of RAID5 if one of the drives goes down and doesn't need the
same space so mirroring is a better bet. With 10K files and
possibly frequent updates you are going to need the journal-
from-hell for this.
LVM is helpful for managing larger spaces, also helps if you
want to break the filesystem up into separate groups for
simpler backup. With linux there isn't any seriuos hit from
multiply mounting the LV's. This allows you to have a 3-level
filesystem with, say, 8 LV's that can be backed up in parallel
(assuming that you have the drives for it). It also allows
for parallel fscking in cases where the journal fries or if
you decide to stick with ext2.
If this sounds paranoid, I've been bit on data warehouse
projects by full and damaged journals, broken disk drives,
mops, bleeing maintinence workers, and operator failure
when I was sozzled. If whatever you design cannot handle
all of these cases gracefully then you'll eventually get
strung up by disgruntled users; if the thing doesn't
survive an atomic blast then their ghosts will haunt you...
--
Steven Lembark 2930 W. Palmer
Workhorse Computing Chicago, IL 60647
+1 800 762 1582
More information about the luni
mailing list