Re: Patch nag: Bug #363400

[Alexander Larsson <alexl redhat com>]

On Wed, 2007-03-28 at 12:52 -0600, Hans Petter Jansson wrote:
> On Wed, 2007-03-28 at 15:12 +0200, Alexander Larsson wrote:
> > > > On Wed, 2007-03-21 at 14:28 -0600, Hans Petter Jansson wrote:
> 
> > > > > http://bugzilla.gnome.org/show_bug.cgi?id=363400
> > > > > 
> > > > > Has a gnome-vfs patch that I believe to be correct. It works around a
> > > > > serious performance issue on XFS file systems. It has been tested by at
> > > > > least two users.
> 
> > But DROP_CACHE_BATCH_SIZE is far smaller (512k) than
> > DROP_CACHE_SIZE_LIMIT (20 meg). Now, i didn't do much scientific
> > research to get to the DROP_CACHE_SIZE_LIMIT value, but its a large
> > change that might cause regressions (or be better, who really knows...)
> 
> My gut feeling is that it needs to be lower than 20MB, but this would be
> very hard to measure because it's really only relevant when a) processes
> are competing for cache or b) we are copying files that are being used
> by other processes.
> 
> I basically arrived at the value based on what would be acceptable
> fadvise() CPU overhead (<1%) for the XFS runs that were posted.
> 
> Will you accept the patch if I just set DROP_CACHE_BATCH_SIZE to 20MB?
> It would preserve the lower fadvise() limit and cause cache "pollution"
> of up to 20MB during the operation.

I think its better to have two values for this. Even if we later change
the 20 meg value i think its valuable to differentiate these two
settings, as they are really two different things (one is batching to
avoid many fadvice calls, one is some attempt to guess when caches can
be safely dropped).

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
 Alexander Larsson                                            Red Hat, Inc 
                   alexl redhat com    alla lysator liu se 
He's an oversexed small-town farmboy from a doomed world. She's a sarcastic 
paranoid vampire from Mars. They fight crime!