Re: [PATCH 5/9] readahead: on-demand readahead logic

[Fengguang Wu <wfg@mail.ustc.edu.cn>]

On Fri, May 18, 2007 at 11:23:35PM -0700, Andrew Morton wrote:
> 
> That would have to be the best changelog I've ever seen ;) Thanks for
> persisting with this.

Thank you :)

> > sysbench oltp (trans/sec): up to 8% gain
> 
> Have you given any thought to identifying workloads which may be worsened
> by your changes?  Attempt to deliberately expose any weak spots?

Yeah. All possible downsides I can imagine are:

- CPU overheads
  Only random reads will be hurt.
  That's 1% slow down for _sparse files_, and should be much smaller
  when real I/O is involved.

- Behavior changes
  It do not enforce strict check sequentialness.
  - it is in general a good behavior for interleaved reads and
    clustered-and-intermixed-random/sequential workloads.
  - it might lead to more readahead misses
    E.g. a random read sequence of 0,1,4,12,28,60,92,124,156,188,220
    that is weird enough to start the readahead and hit all the
    lookahead pages.
    I highly doubt the possibility of such patterns happen in real
    world. But if ever it happens repeatedly for some user, he can
    work it around easily by tuning readahead_kb to some other value.

So, it is only a possibility that some random workload may be
worsened. But it's really hard to find one real world example.