Hey: On Wed, Oct 28, 2015 at 2:11 PM, Bob Weinand <bobw...@hotmail.com> wrote:
> > Am 28.10.2015 um 05:20 schrieb Dmitry Stogov <dmi...@zend.com>: > > What is the general impression from the patch? > Should we commit it (with minor fixes and/or additional optimisations)? > Bob, do you see performance improvement on your app? > > Thanks. Dmitry. > On Oct 27, 2015 4:14 PM, "Dmitry Stogov" <dmi...@zend.com> wrote: > >> Hi, >> >> Please take a look: https://github.com/php/php-src/pull/1604 >> The speed improvement on real-life is insignificant. >> VM size is increased by 10KB. >> >> I don't have a strong opinion, if this should go into master or not. >> >> Thanks. Dmitry. >> > > In general I like the patch, if only for the reason that it splits up the > monster zend_fetch_var_address_helper actually was into two helpers… > > I benchmarked https://github.com/amphp/aerys/blob/master/lib/HPack.php with > the following code: > > // Two example request headers from RFC 7541 > $data1 = hex2bin(str_replace(" ", "", "8286 8441 8cf1 e3c2 e5f2 3a6b a0ab > 90f4 ff")); > $data2 = hex2bin(str_replace(" ", "", "8286 84be 5886 a8eb 1064 9cbf")); > > for ($i = 0; $i < 1e5; $i++) { > $obj = new HPack; > $obj->decode($data1); > $obj->decode($data2); > } > > and got 6.27 vs. 6.01 seconds (4%). With the one or other optimization > (like mentioned in PR: not doing a zend_fetch_class() on each call, maybe > more), it'll be probably even more. > 4% is also a significant improvement. I 'd like to see it to be committed to master branch, we can keep do something opts base on this after it merged. thanks > > (I realize there's a lot overhead, but that's because I try to have a not > too synthetic benchmark; but at least these methods are the bottleneck of > most HTTP/2 traffic in this server app, so, if I can speed these up, it > will be significant.) > > Bob > -- Xinchen Hui @Laruence http://www.laruence.com/
