On 6/16/05, Nicholas Clark <[EMAIL PROTECTED]> wrote: > Sorry for the delay in this answer
No problem. > On Wed, Jun 01, 2005 at 06:54:30PM +0200, demerphq wrote: > > On 6/1/05, Nicholas Clark <[EMAIL PROTECTED]> wrote: > > > It would be interesting if someone found some good way of benchmarking > > > perl, > > > and compared current blead with released 5.9.2 > > > > Is perlbench not the right way? What does "good" imply for you? > > perlbench seems to be remarkably good at showing no real difference in > speed whatever changes it's being trialled with. > > I think that Jarkko noted that the size of things it manipulates are tiny > compared with CPU caches, so it won't demonstrate any improved efficiency > in cache usage. > > Also, it concentrates on its tests being as backwards compatible as possible, > with most running on perl4, so we're not convinced that it's really testing > constructions that are used these days > > It tests things in a tight loop, so it's testing execution time only. If we > optimise the startup time of perl, it's not going to reward us with better > numbers. > > IIRC it showed no change with the introduction of shared hash keys, or at > least with the introduction of optimisations to take advantage of them, > yet Nick I-S reported that some TK programs were >2% faster, something > a reasonable benchmark should be capable of spotting. > > I guess it's still measuring package variables and procedural code, when > far more of the world is using lexicals and objects these days. So what you are saying in is that we aught to rewrite perlbench for the 5.[689] world? I think that could be an interesting project. I put together a framework for cross testing different versions of the perl regex stuff so maybe ill give this a go. I guess ideally it would be a plug and play scenario. Also, would there be any disagreement that ditching normalization logic is acceptable? IMO its just a waste of time if you are doing wallclock comparisons of two perl executables on the same machine under similar load (which I reckon most of us would be doing). Yves -- perl -Mre=debug -e "/just|another|perl|hacker/"
