On Sat, Feb 5, 2011 at 9:37 PM, Johan Corveleyn <jcor...@gmail.com> wrote: > On Sat, Feb 5, 2011 at 7:14 PM, Stefan Sperling <s...@elego.de> wrote: >> On Sat, Feb 05, 2011 at 06:47:35PM +0100, Branko Čibej wrote: >>> I would not worry about existing clients -- simply mark the existing >>> APIs as deprecated, but keep them and do not attempt to improve their >>> performance. >> >> Neglecting performance of backwards compat code is an interesting idea. >> It allows us to focus on the new APIs first and foremost. >> >> The existing APIs will continue to work using the node walker and issue >> queries per node as they do now. We could consider optimising them later, >> before or after 1.7 release. The required changes are mostly mechanical. > > I agree with most of what's been said here. I think it would be a pity > to use WC-NG in a way that provides far from optimal performance. > > FWIW, I just did a quick run of your per-directory proplist query vs. > the per-node version, on my Windows XP platform, to have another data > point. The performance improvement is significant, but not > earth-shattering (around 20%). > > Just tested with a freshly checked out working copy of svn trunk: > > 1) Per-node queries (r1066540). Looking at the third run, to make sure > everything is hot in cache: > > $ time svn proplist -R . >/dev/null > > real 0m1.532s > user 0m0.015s > sys 0m0.015s > > > 2) Per-dir queries (r1066541). Looking at the third run, to make sure > everything is hot in cache: > > $ time svn proplist -R . >/dev/null > > real 0m1.218s > user 0m0.015s > sys 0m0.031s > > > 3) For comparison, I also tested with SlikSVN 1.6.13. This is still > more than twice as fast: > > $ time svn proplist -R . >/dev/null > > real 0m0.578s > user 0m0.015s > sys 0m0.046s
I should've added one more test, to put things in perspective :-), namely the test with r1039808 (the per-wc query version): $ time svn proplist -R . >/dev/null real 0m0.328s user 0m0.015s sys 0m0.031s (but, as you probably already know, this execution is not cancellable, which is pretty annoying :-), especially when I forget to redirect to /dev/null, but let it write on the console). Cheers, -- Johan
