On Wed, Dec 08, 2004 at 10:39:29PM +0100, Michael M?ller wrote:
> On Mon, Dec 06, 2004 at 05:53:33PM -0500, Sonny Rao <[EMAIL PROTECTED]> wrote:
> > On Sun, Dec 05, 2004 at 11:40:21AM +0100, Michael M?ller wrote:
> > > Hi all,
> > > 
> > > I read an article in the German 'Linux Magazin' 11/04 about a
> > > comparision of the different FS. They tested Ext2, Ext3, JFS, XFS,
> > > ReiserFS, Reiser4 and Veritas. Detailed results can be found on
> > > http://www.linux-magazin.de/Service/Listings/2004/11/fs_bench.
> 
> The link only contains test results; no German texts.

True, there were a few articles for November here:
http://www.linux-magazin.de/Artikel/ausgabe/2004/11
 
> > My guess is that they didn't set the readahead high enough for
> > whatever type of device they were testing on 2.6 (It looks like a Raid
> > array, since on 2.4 it gets about 100MB/sec, which I don't think very
> > many single disks can do).  The readahead implementation on 2.6 is
> > certainly different from the one on 2.4.  IO performance on 2.6 is
> > much, much better across the board.
> > 
> > My German isn't great, so I'm not going to try and read the article,
> > but I'd also like to know what kind of array they are using for this
> > test.  Before we can make any conclusions, we should know what the
> > hardware is capable of doing.
> 
> The hardware:
> 
> Pentium 4, 2.8GHz, 512MB, 12 SATA-HDs in a RAID, overall capacity 2TB,
> test partition 200GB
> 
> For the 2.4 tests they used SuSE Linux Enterprise Server 8, kernel
> 2.4.21-138-smp, for 2.6 SuSE Linux 9.1, 2.6.7-mm4 with patches for
> Reiser 4.

Ok, how did they set the readahead size in the tests, just the defaults ?
For a 12 disk array, the default of 128k readahead on 2.6 isn't going
to cut it.

Was it a hardware or software RAID?  RAID-0, RAID-5, RAID-10?
If it was hardware, what type of adapter, was it a PCI-X adapter or
just a regular PCI? 

Given all of that, what is the expected/advertised hardware throughput?
Something like aio-stress would be a good test since a filesystem
isn't required, and we can isolate problems in the block layer/drivers. 

One needs most of these details to make any kind of reasonable
conclusion from the results given.

Sonny

_______________________________________________
Jfs-discussion mailing list
[EMAIL PROTECTED]
http://www-124.ibm.com/developerworks/oss/mailman/listinfo/jfs-discussion

Reply via email to