Mailing List archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[vdr] Re: Filesystem pros & cons



> The problem may not be the filesystem but the buffer cache handling. For
> consistent tests you have to write files that are at least twice as
> large as main memory. I got 40+ MB/s with ext3 and xfs. This is
> certainly enough even for my server with 4 DVB cards.

some tests might help to decide this, i just did:

time dd count=1000000 if=/dev/zero of=/data1/test.null
1000000+0 records in
1000000+0 records out
real    1m29.269s
user    0m2.510s
sys     0m30.240s

which equals to 5,7 MB/s (hdparm shows me 13.5 MB/s)
(thats on a Maxtor 94098U8, 40GB, P2-450, 128MB-RAM, reiserfs 0.5.x, 
partition heavily used before and about 80% full, 2.4.20 stock)

It's also not nice, that during that process the sysload varies from 0-100%.

The question is specially how much fragmentation is going on on the partition 
which decreases read and write speed over lifetime until defragmentation is 
done. To my knowledge only ext2 has such a utility of the named filesystems 
(does this work on ext3?) which would run hours on a multigigabyte partion...

it would be interesting to collect _comparable_ numbers for all filesystems 
under different conditions:

1.fresh&empty partion
2.same filled up to 90%
3.after that reempty to 50%

3. is meant to simulate gaps - of course this depends on how large the gaps 
in comparison to typically written files are, worst is for sure randomly 
deleted blocks all over the disk which is fortunately very unlikely to happen.
It's generally a good idea to separate files which large difference in size 
on different partitions - better don't store lot's of small files on your 
video-partition.

What i want to say is, that it's not important on how fast files are written 
and on a fresh partition but on a heavily used one, thats what needs to be 
compared.

To make it comparable one needs to know that on the same hard-disk and other 
hardware + software for all filesystems. The test in ct-magazin did'nt cover 
"used" partitions AFAIK and is quite old right now - has someone have some 
numbers with later versions of the filesystems?


-- 
Info:
To unsubscribe send a mail to ecartis@linuxtv.org with "unsubscribe vdr" as subject.



Home | Main Index | Thread Index