Newsgroups: php.internals
Path: news.php.net
Xref: news.php.net php.internals:36055
Mailing-List: contact internals-help@lists.php.net; run by ezmlm
Received-SPF: error (pb1.pair.com: domain chiaraquartet.net from 38.99.98.18 cause and error)
Message-ID: <47D21F95.7090704@chiaraquartet.net>
Date: Fri, 07 Mar 2008 23:09:41 -0600
User-Agent: Thunderbird 2.0.0.12 (X11/20080227)
MIME-Version: 1.0
To: internals Mailing List <internals@lists.php.net>, 
 Dmitry Stogov <dmitry@zend.com>
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: 7bit
Subject: profiling followup on addition of streams to include_path - it's
 *faster*
From: greg@chiaraquartet.net (Gregory Beaver)

Hi,

When I posted yesterday's patch to add stream support to include_path
(http://news.php.net/php.internals/36031) I mentioned that I suspected
benchmarking would reveal it to be slow.  My primary goal is to provide
no impact on current users who are using a traditional include_path,
with a secondary goal of improving performance of those who use the new
syntax.  Today I ran callgrind on the thing, with some surprising results.

With the patch, include is *faster* for our traditional users than it is
now.
With the patch, include_once with >1000 unique files is about 3% slower
- not the whole execution, just include_once
With the patch, include_once with 1 unique file included 10000 times is
insignificantly slower (about 0.4%)

For these reasons, I'm really encouraged :).  The next step is to
absolutely ensure correctness and then see if the streams part of
include_path  can be optimized at all (or if it needs it).

Details
======

I just ran callgrind on this script:

<?php
set_include_path('.:/usr/local/lib/php:/home/cellog/workspace/php5/ext/phar');
for ($i = 0; $i < 100000; $i++) {
include 'extra.php';
}

The empty file "extra.php" (zero byte) is in
"/home/cellog/workspace/php5/ext/phar/extra.php" ensuring that we
traverse include_path to find it.

To my great shock, the script runs *faster* with my patch, because it
executes significantly more instruction cycles in
php_stream_open_for_zend_ex without the patch.

Note that this does not measure the cost of *_once.  *_once is a lot
harder to measure, so I created 10,000 files (yikes) via this script:

<?php
for ($i = 1; $i <= 10000; $i++) file_put_contents('test' . $i, '');

and then ran this test script:

<?php
set_include_path('.:/usr/local/lib/php:/home/cellog/workspace/php5/poop');
for ($i = 1; $i <= 10000; $i++) {
include_once 'test' . $i;
}

callgrind reported that php_resolve_path was about twice as slow as the
other version, resulting in a 3% degradation of include_once performance
over the current version (which is much faster than 5.2.x, incidentally).

Finally, to test the _once aspect of include_once, I ran this script:

<?php
set_include_path('.:/usr/local/lib/php:/home/cellog/workspace/php5/ext/phar');
for ($i = 0; $i < 100000; $i++) {
include_once 'extra.php';
}

With this script, it really highlights the most common use case of
include/require_once: attempting to include the same file multiple
times.  The difference in performance was insignificant, with callgrind
reporting a total execution portion of 75.12% for CVS, and 75.57% with
my patch.

So, it looks like the biggest performance hit would be for users
including more than 1000 different files, and would result in
approximately 3% slower performance *of include_once*.  I'm curious how
many of our readers have a PHP setup that includes close to this many
files, because it seems rather unlikely to me that anyone would include
more than a few hundred in a single process.

The surprising news is that users who are using "include" would see a
performance improvement from my patch, so I recommend that portion be
committed regardless of other actions.  This improvement proabbly
results from removing an include_path search in plain_wrapper.

Thanks,
Greg