Hi,
I have some memory reading some similar question somewhere (here?) but I'm
not sure there was an answer.
I do observe strange behaviour if read performance fo my IDE harddisk as
reported by hdparm (or doing linear reads with a self written program):
My FUJITSU MPG3409AT E is supposed to make slightly above 30MB/s. However,
it's connected to a PIIX4, which can only do UDMA33. So I expect something
between 25 and 30 MB/s maximumn speed.
I get it. But not over the whole disk.
Doing a read speed measurement on /dev/hda, I constantly get ~16 MB/s.
Not bad, but less than I'd expect. Measuring single partitions, some show
the same, some show significantly more, 26MB/s--18MB/s, depending on the
position of the partition on disk. Those look good!
There are enough partitions to see a clear pattern: Those with mounted ext2
filesystems perform better. Umounting them does not harm, they just need to
have been mounted once. reiser or (v)fat however don't improve anything.
swap does, as does a ext2 over raid5.
Kernel 2.4.3pre7; Dual iPIII-700 system; i440BX MoBo.
Is this to be expected? Blocksize issues? Readahead behaviour? What's
changed on ext2 mounting ... ?
Regards,
--
Kurt Garloff <[email protected]> [Eindhoven, NL]
Physics: Plasma simulations <[email protected]> [TU Eindhoven, NL]
Linux: SCSI, Security <[email protected]> [SuSE Nuernberg, FRG]
(See mail header or public key servers for PGP2 and GPG public keys.)
On Tue, Apr 24 2001, Kurt Garloff wrote:
> I get it. But not over the whole disk.
> Doing a read speed measurement on /dev/hda, I constantly get ~16 MB/s.
> Not bad, but less than I'd expect. Measuring single partitions, some show
> the same, some show significantly more, 26MB/s--18MB/s, depending on the
> position of the partition on disk. Those look good!
>
> There are enough partitions to see a clear pattern: Those with mounted ext2
> filesystems perform better. Umounting them does not harm, they just need to
> have been mounted once. reiser or (v)fat however don't improve anything.
> swap does, as does a ext2 over raid5.
You wouldn't happen to have 4kB ext2 filesystems on those? When ext2
mounts, it sets the soft blocksize to that then, I would expect this to
give at least some benefit over using 1kB blocks (as your IDE partition
otherwise would have).
--
Jens Axboe
On Tue, Apr 24, 2001 at 10:58:58AM +0200, Jens Axboe wrote:
> On Tue, Apr 24 2001, Kurt Garloff wrote:
> > There are enough partitions to see a clear pattern: Those with mounted ext2
> > filesystems perform better. Umounting them does not harm, they just need to
> > have been mounted once. reiser or (v)fat however don't improve anything.
> > swap does, as does a ext2 over raid5.
>
> You wouldn't happen to have 4kB ext2 filesystems on those?
Sure I do.
> When ext2 mounts, it sets the soft blocksize to that then, I would expect
> this to give at least some benefit over using 1kB blocks (as your IDE
> partition otherwise would have).
Why? Are the request sizes larger this way? This would mean that the
overhead is very significant, turning a max of 26MB/s into 16MB/s!
If so, shouldn't we try to get the same effect also for the whole disk or
other filesystems? Most notably reiser?
Regards,
--
Kurt Garloff <[email protected]> Eindhoven, NL
GPG key: See mail header, key servers Linux kernel development
SuSE GmbH, Nuernberg, FRG SCSI, Security
On Tue, Apr 24 2001, Kurt Garloff wrote:
> > You wouldn't happen to have 4kB ext2 filesystems on those?
>
> Sure I do.
>
> > When ext2 mounts, it sets the soft blocksize to that then, I would expect
> > this to give at least some benefit over using 1kB blocks (as your IDE
> > partition otherwise would have).
>
> Why? Are the request sizes larger this way? This would mean that the
> overhead is very significant, turning a max of 26MB/s into 16MB/s!
Because you'll be doing I/O on 4kB entries at least, and the overhead of
merging a 128kB request into the queue is much smaller this way. But
yes, 16 -> 26 seems quite a large win, I wouldn't expect this much
(unless you get no merging in any of the cases, then I suspect the 4kB
would be an even bigger win)
> If so, shouldn't we try to get the same effect also for the whole disk or
> other filesystems? Most notably reiser?
I would expect reiser to do the same. Setting the soft block size higher
is not required though, users are free to submit 4kB buffer heads even
for devices that have 1kB set. By default, it doesn't happen though.
--
Jens Axboe