Re: [PATCH 0/3] readahead drop behind and size adjustment

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Sun, 2007-07-22 at 18:33 +1000, Rusty Russell wrote:
> On Sun, 2007-07-22 at 16:10 +0800, Fengguang Wu wrote:
> > - It will avoid large-file-reads-thrashing-my-desktop problem,
> >   so most desktop users should like it. But sure there will be counter
> >   cases when a user want to keep the data cached.
> > - File servers may hurt from it. Imagine a mp3/png file server. The
> >   files are large enough to trigger drop-behind, but small (and hot)
> >   enough to be cached. Also when a new fedora DVD iso is released, it
> >   may be cached for some days. These are only the obvious cases.
> 
> I'm still not convinced of the latter case.  If a second user accesses
> the file before it pages are thrown out, from my understanding there
> won't be readahead but the pages will be put back on the active list
> (ie. the dropbehind is undone).
> 
> AFAICT this patch truly biasses towards releasing pages from
> sequentially accessed files.  So file servers where memory pressure
> comes from lots of such files should be fine (bias will work correctly
> against least-served files).
> 
> The main problem I can see is a low-memory server where we are currently
> swapping out pages from non-sequential files (eg. gigantic running java
> apps), and now performance might decline because we'll discard file
> pages first.
> 
> > So I opt for it being made tunable, safe, and turned off by default.
> 
> I'd like to see it turned on by default in -mm, and try to come up with
> some server-like workload to measure the effect.  Should be easy to
> simulate something (eg. apache server, where clients grab some files in
> preference, and apache server where clients grab different files).
> 
> Peter?

I guess we could do that. Just populate the server with a lot of
randomly sized files, and make a client do a random pull or one with a
poisson distribution or something like that.