Re: Strange IDE performance change in 2.6.1-rc1 (again)

["Ram Pai" <linuxram@us.ibm.com>]

Sorry I was on vacation and could not get back earlier.

I do not exactly know the reason why sequential reads on blockdevices
has regressed. One probable reason is that the same lazy-read
optimization which helps large random reads is regressing the sequential
read performance.

Note: the patch, waits till the last page in the current window is being
read, before triggering a new readahead. By the time the readahead
request is satisfied, the next sequential read may already have been
requested. Hence there is some loss of parallelism here. However given
that largesize random reads is the most common case; this patch attacks
that case.

If you revert back just the lazy-read optimization, you might see no
regression for sequential reads,

Let me see if I can verify this,
Ram Pai


On Sun, 2004-01-04 at 06:30, Paolo Ornati wrote:
> On Saturday 03 January 2004 23:40, Andrew Morton wrote:
> > Paolo Ornati <ornati@lycos.it> wrote:
> > > I know these are only performance in sequential data reads... and real
> > > life is another thing... but I think the author of the patch should be
> > > informed (Ram Pai).
> >
> > There does seem to be something whacky going on with readahead against
> > blockdevices.  Perhaps it is related to the soft blocksize.  I've never
> > been able to reproduce any of this.
> >
> > Be aware that buffered reads for blockdevs are treated fairly differently
> > from buffered reads for regular files: they only use lowmem and we always
> > attach buffer_heads and perform I/O against them.
> >
> > No effort was made to optimise buffered blockdev reads because it is not
> > very important and my main interest was in data coherency and filesystem
> > metadata consistency.
> >
> > If you observe the same things reading from regular files then that is
> > more important.
> 
> I have done some tests with this stupid script and it seems that you are 
> right:
> _____________________________________________________________________
> #!/bin/sh
> 
> DEV=/dev/hda7
> MOUNT_DIR=mnt
> BIG_FILE=$MOUNT_DIR/big_file
> 
> mount $DEV $MOUNT_DIR
> if [ ! -f $BIG_FILE ]; then
>     echo "[DD] $BIG_FILE"
>     dd if=/dev/zero of=$BIG_FILE bs=1M count=1024
>     umount $MOUNT_DIR
>     mount $DEV $MOUNT_DIR
> fi
> 
> killall5
> sleep 2
> sync
> sleep 2
> 
> time cat $BIG_FILE > /dev/null
> umount $MOUNT_DIR
> _____________________________________________________________________
> 
> 
> Results for plain 2.6.1-rc1 (A) and 2.6.1-rc1 without Ram Pai's patch (B):
> 
> o readahead = 256 (default setting)
> 
> (A)
> real	0m43.596s
> user	0m0.153s
> sys	0m5.602s
> 
> real	0m42.971s
> user	0m0.136s
> sys	0m5.571s
> 
> real	0m42.888s
> user	0m0.137s
> sys	0m5.648s
> 
> (B)
> real    0m43.520s
> user    0m0.130s
> sys     0m5.615s
> 
> real	0m42.930s
> user	0m0.154s
> sys	0m5.745s
> 
> real	0m42.937s
> user	0m0.120s
> sys	0m5.751s
> 
> 
> o readahead = 128
> 
> (A)
> real	0m35.932s
> user	0m0.133s
> sys	0m5.926s
> 
> real	0m35.925s
> user	0m0.146s
> sys	0m5.930s
> 
> real	0m35.892s
> user	0m0.145s
> sys	0m5.946s
> 
> (B)
> real	0m35.957s
> user	0m0.136s
> sys	0m6.041s
> 
> real	0m35.958s
> user	0m0.136s
> sys	0m5.957s
> 
> real	0m35.924s
> user	0m0.146s
> sys	0m6.069s
> 
> 
> o readahead = 64
> (A)
> real	0m35.284s
> user	0m0.137s
> sys	0m6.182s
> 
> real	0m35.267s
> user	0m0.134s
> sys	0m6.110s
> 
> real	0m35.260s
> user	0m0.149s
> sys	0m6.003s
> 
> 
> (B)
> real	0m35.210s
> user	0m0.149s
> sys	0m6.009s
> 
> real	0m35.341s
> user	0m0.151s
> sys	0m6.119s
> 
> real	0m35.151s
> user	0m0.144s
> sys	0m6.195s
> 
> 
> I don't notice any big difference between kernel A and kernel B....
> 
> From these tests the best readahead value for my HD seems to be 64... and 
> the default setting (256) just wrong.
> 
> With 2.4.23 kernel and readahead = 8 I get results like these:
> 
> real	0m40.085s
> user	0m0.130s
> sys	0m4.560s
> 
> real	0m40.058s
> user	0m0.090s
> sys	0m4.630s
> 
> Bye.