poor performance on big machine

Steve O'Hara-Smith steve at sohara.org
Mon Jul 30 10:30:52 PDT 2007


On Mon, 30 Jul 2007 09:57:11 -0700 (PDT)
Matthew Dillon <dillon at apollo.backplane.com> wrote:

>     I don't think an I/O scheduler will help.  You are maxing out the hard
>     drive with your workload.  MB/sec is irrelevant... the workload is 
>     causing the disk to seek a lot.

	If it is thrashing between a small number of areas then it may be
possible to improve things with a scheduler that buffers and attempts to
cluster the requests by track (OK address proximity). In extreme cases
going from one operation per seek to two operations per seek would be a
big gain.

	Of course this is likely to interfere with any "has been committed
to disc" semantics needed.

-- 
C:>WIN                                      |   Directable Mirror Arrays
The computer obeys and wins.                | A better way to focus the sun
You lose and Bill collects.                 |    licences available see
                                            |    http://www.sohara.org/





More information about the Kernel mailing list