Newsgroups: php.internals
Path: news.php.net
Xref: news.php.net php.internals:64121
Mailing-List: contact internals-help@lists.php.net; run by ezmlm
Received-SPF: error (pb1.pair.com: domain ralphschindler.com from 209.85.214.170 cause and error)
Message-ID: <50B8C8BB.7010409@ralphschindler.com>
Date: Fri, 30 Nov 2012 08:54:51 -0600
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.8; rv:17.0) Gecko/17.0 Thunderbird/17.0
MIME-Version: 1.0
To: internals <internals@lists.php.net>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
Subject: APC + Phar Performance (x-post from pecl.dev)
From: ralph@ralphschindler.com (Ralph Schindler)

Hey all,

After reading this post ( 
http://www.reddit.com/r/PHP/comments/13uwgk/phar_performance/ ), and 
having my own curiosity of the current state of things, I wanted to dig 
in an see what the performance of APC with phar files was.

I've found some oddities in performance, and I am hoping before I do any 
deeper investigation, someone might know something.

The app is here:
https://github.com/ralphschindler/test-phar-performance-apc/

The library I am attempting to introduce to the app is Zend_Db.phar.  I 
chose this b/c I fully know the code base, and the phar clocks in at 
519844 bytes uncompressed.

It breaks down like this:
   /index.php - baseline app without addition of library
   /index-with-filesys.php - Zend_Db library extracted to filesystem
   /index-with-phar.php - Zend_Db as phar included

Without APC, on my machine, I get a slight drop off, but acceptable. 
With APC, the drop off in performance is very noticable.  Even with 
stat=0, only in a very specific usage scenario (described below) does it 
actually nullify phar vs. filesystem.


Without APC (enabled=0)
-----------------------

$ ab -H "Connection: close" -n 500 http://test.dev/index-with-filesys.php

Requests per second:    111.19 [#/sec] (mean)
Time per request:       8.994 [ms] (mean)


$ ab -H "Connection: close" -n 500 http://test.dev/index-with-phar.php

Requests per second:    87.49 [#/sec] (mean)
Time per request:       11.429 [ms] (mean)


Then APC 3.1.13 (Beta) is introduced, the app is clearly faster on the 
filesystem, but not nearly as fast one would expect in the phar version:


With APC (enabled=1, stat=1)
----------------------------

$ ab -H "Connection: close" -n 500 http://test.dev/index-with-filesys.php

Requests per second:    474.47 [#/sec] (mean)
Time per request:       2.108 [ms] (mean)


$ ab -H "Connection: close" -n 500 http://test.dev/index-with-phar.php

Requests per second:    238.84 [#/sec] (mean)
Time per request:       4.187 [ms] (mean)


With APC (enabled=1, stat=0)
----------------------------

$ ab -H "Connection: close" -n 500 http://test.dev/index-with-filesys.php

Requests per second:    529.02 [#/sec] (mean)
Time per request:       1.890 [ms] (mean)


$ ab -H "Connection: close" -n 500 http://test.dev/index-with-phar.php

Requests per second:    246.71 [#/sec] (mean)
Time per request:       4.053 [ms] (mean)




With APC (enabled=1, stat=0) (With include_path hack)
-----------------------------------------------------

If this line:

   require '../vendor/Zend_Db-2.1.0beta1.phar';

is swapped out with this (to ensure the require is relative to include):

   set_include_path('.:' . realpath(__DIR__ . '/../vendor'));
   require 'Zend_Db-2.1.0beta1.phar';

I get the following:

$ ab -H "Connection: close" -n 500 http://test.dev/index-with-phar.php

Requests per second:    522.68 [#/sec] (mean)
Time per request:       1.913 [ms] (mean)



Can someone help me make sense of all this?  Here are my questions:

   * The phar files are listed in the apc_cache_info() with >1 hits, are 
the stat's to those phar files really much slower than to filesystem files?

   * Why does the include_path trick affect performance with regards to 
phar files?




-ralph