Re: A fix for Re: Severe I/O performance regression 2.6.6 to 2.6.7 or 2.6.8-rc3

[Andrew Morton <akpm@osdl.org>]

Berkley Shands <berkley@cs.wustl.edu> wrote:
>
> Taking the 2.6.6-bk6/mm/readahead.c (a one line change from vanilla 2.6.6/mm/readahead.c)
>  and inserting it into 2.6.7 and 2.6.8-rc3 restores the thread read ahead functionality
>  and througput seen in the 2.6.6 kernel. This does however reduce the single stream read
>  performance from 426MB/Sec (under 2.6.8-rc3) to 386MB/Sec (same as 2.6.6).
>  The major functionality change was to try to detect "random" read performance,
>  if randomness is detected, then NEVER read ahead. In this case, two threads
>  using two different file descriptors to the same file read alternating strides
>  of the file. Since each thread did not do SEQUENTIAL reads, both were treated as random
>  accesses.

Let me attempt to clarify here.  You have two fd's open against the same
file and you are reading in the following manner:

fd1:   0-11k        33-44k        55-66k
fd2:         11-22k        44-55k

yes?

If correct, I think you're expecting a bit too much of the readahead
system.  If you have a real application which is using such access patterns
perhaps it should be taught to use POSIX_FADV_WILLNEED.