Greg Stein <gstein@lyra.org> writes:
> On Mon, Aug 18, 2003 at 10:09:52PM -0500, cmpilato@collab.net wrote:
> >...
> > I don't know if "suck up all of memory" is a realistic concern -- my
> > estimates have it maybe growing to 30M for a 20,000-path checkout --
> > but still, until the algorithm is smart enough to bound growth and
> > intelligent roll things out of the cache, it's a non-starter (which
> > is, I'm sure, something you agree with).
>
> All right... but what is the cache to do? What was measured as "poor" such
> that a cache might actually benefit? And could the "poor" measurement be
> fixed algorithmically rather than patched cache-ly?
The idea was that the work of walking of the DAG to get to a node-rev-id
was ailing us. So the cache was of a hash of path->dag_node_t. Would
could reintroduce the 'id_root' again, but frankly I think that's a
horrible idea.
Anyway, we need more solid numbers that include times spent in APR and
libc in order to truly get a sense for our painful spots, so I've
squirreled away this patch for a while.
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@subversion.tigris.org
For additional commands, e-mail: dev-help@subversion.tigris.org
Received on Tue Aug 19 14:38:31 2003