On Sun, 25 Jun 2000, Jens Axboe wrote:
<snip>
>
> I also did a dbench run.
>
> 2.2.17-pre6
>
> burns:/opt/software/testing # ./dbench 48
> 48 clients started
> Throughput 12.84 MB/sec (NB=16.0499 MB/sec 128.4 MBit/sec)
>
> 2.4.0-test2-ac1
>
> burns:/opt/software/testing # ./dbench 48
> 48 clients started
> Throughput 30.6317 MB/sec (NB=38.2896 MB/sec 306.317 MBit/sec)
>
> The 2.2 results are very disappointing here, the last time I checked
> 2.2 dbench results were in the ~24MB/sec range.
In 2.2.17pre7, try_to_free_buffers() was only waking up kflushd in case
sync_page_buffers() failed to sync the dirty buffers of a page.
With the attached patch, try_to_free_buffers() wakes up kflushd if the
percentage of dirty memory is high, independant of sync_page_buffers
failing or not. The patch also modifies do_try_to_free_pages to print a
"VM: do_try_to_free_pages failed..." only when _no_ progress is made.
Testing dbench with 48 processes on one of the test machines around here
(IDE disk), I've got 13MB/sec with 2.2.17pre7, 17MB/sec with 2.2.17pre7 +
the attached patch, and 23MB/s with 2.4.0-t1-ac22.
Jens, could you please try to run dbench with this patch on your machine
and report us results? Thanks a lot.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Mon Jun 26 2000 - 21:00:07 EST