2008-11-01 02:34:19

by Wu Fengguang

[permalink] [raw]
Subject: Re: readahead in the 2.6.25+ kernel

Hi John,

On Fri, Oct 31, 2008 at 09:00:30PM -0400, John Galloway wrote:
>
> Hi,
>
> I'm looking into read ahead issues wrt Ingres, particularly wrt mmap.
> I gather (but am not sure) your readahead patches have been incorporated
> into the 2.6 kernel around 2.6.15, with some simplfications it seems??
> Are you still maintaining a distinct patch set?

The ondemand readahead patch series have been merged into linux 2.6.23/24.
So I no longer maintain out-of-tree readahead patches.

> And can you help me out about how linux readahead will effect mmap() as
> opposed to read()? i.e. is it enabled for mmap() as well?

No, mmap readahead is always a dumb and simple 'read-around' one, and
it seems there is no general interest on improving it.

Currently you have two options to affect mmap readahead performance:
- tune readahead size with command `blockdev --setra'
- tell kernel to use the ondemand readahead routines with
madvise(MADV_SEQUENTIAL)

Note that MADV_SEQUENTIAL is safe to use at least for the current
readahead implementation, it won't blindly do aggressive readahead on
seeing MADV_SEQUENTIAL. The flag only switches the readahead routine
from the dumb read-around one to the somehow smarter readahead logic,
which _detects_ sequential patterns for us.

As a long term solution, if you get favorable numbers with MADV_SEQUENTIAL,
that could be a 'first impetus' to improve the default mmap readahead code ;-)

Cheers,
Fengguang