RE: problem with 2.6.13.[0-2]

From: David Ronis
Date: Wed Sep 28 2005 - 20:38:57 EST


On Tue, 2005-09-27 at 22:55 -0400, Parag Warudkar wrote:
>
> > DMA is on in both 2.6.12.6 and in 2.6.13.2. Here's what hdparm
> > /dev/hda gives:
> >
> > in 2.6.12.6:
> >
> > /dev/hda:
> > multcount = 16 (on)
> > IO_support = 0 (default 16-bit)
> > unmaskirq = 0 (off)
> > using_dma = 1 (on)
> > keepsettings = 0 (off)
> > readonly = 0 (off)
> > readahead = 256 (on)
> > geometry = 65535/16/63, sectors = 195371568, start = 0
> >
> > in 2.6.13.2 it's the same, except for
> >
> > multcount = 0 (off)
> >
> > Could this be the problem, and if so, would setting multcount to on
> > fix it? (I take it hdparm -m 16 /dev/hda would do it)
>
> I doubt if multcount will make a difference, but you can try.
> Probably next best thing to do is profile the slow kernel to find where it
> is spending time. (pass profile=2 on kernel command line and use
> readprofile or use oprofile).
>
> Did you try hdparm -tT /dev/hda? It's going to be slow, thats for sure
> from what you described but try that and then post dmesg (for both
> kernels) and profile results - may be someone will get a clue from that.
>
> Parag
>
>

Hi Parag,

Here's what I get:

In 2.6.12.6:

/dev/hda:
Timing cached reads: 1140 MB in 2.00 seconds = 569.80 MB/sec
Timing buffered disk reads: 102 MB in 3.02 seconds = 33.80 MB/sec

In 2.6.13.2:

/dev/hda:
Timing cached reads: 28 MB in 2.15 seconds = 13.03 MB/sec
Timing buffered disk reads: 14 MB in 3.30 seconds = 4.24 MB/sec

and after hdparm -m 16 /dev/hda (recall this is the default in 2.6.12.6)

/dev/hda:
Timing cached reads: 24 MB in 2.05 seconds = 11.73 MB/sec
Timing buffered disk reads: 36 MB in 3.11 seconds = 11.56 MB/sec

I ran thing a few times in each case and the results were close. There
was nothing in dmesg.

David


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/