Zend\Loader\ClassMapAutoloader - Performance improvement

[fedys]

array_merge() in Zend\Loader\ClassMapAutoloader::registerAutoloadMap() method causes significant performace lost. Replace $this->map = array_merge($this->map, $map); with foreach ($map as $key => $value) { $this->map[$key] = $value; }.

[Ocramius]

@fedys is that true also for larger maps?

[fedys]

Absolutely, in my production environment the RESULTING map has over 2000 records with aprox. 40 custom modules(each provides its map). Performance gaing is 10-15ms after use of foreach(before: 190ms, after: 175ms).

[manuakasam]

http://3v4l.org/gAt9X 10 entries
http://3v4l.org/7Zo0k 100 entries
http://3v4l.org/NSKdT 1000 entries
http://3v4l.org/B7ajo 10000 entries

A little depending on the PHP Version, but most of the time foreach is faster.

[Ocramius]

Just verified it as well here

ArrayTest\ArrayMergeEvent
classmaps-performance
Method Name Iterations Average Time Ops/s Relative
------------------------------ ---------- ------------ -------------- --------- ---------
arrayMergePerformance : [Baseline] [10,000 ] [0.0005519990921] [1,811.59718]
forEachPerformance : [10,000 ] [0.0003462410450] [2,888.16134] [62.72%]

That is... curious...

[Ocramius]

According to @DaveRandom, the slowness is caused by copy-on-write going on at each call of array_merge. That's a lot of stuff being copied.

I've added a test that tries to do a single array_merge with multiple params, and the results are much much better:

ArrayTest\ArrayMergeEvent
  classmaps-performance
    Method Name                                 Iterations    Average Time      Ops/s    Relative
    ------------------------------  ----------  ------------ --------------   ---------  ---------
    singleArrayMerge              : [Baseline] [10,000    ] [0.0003203309298] [3,121.77160]
    singleArrayMergeCallUserFuncArray:            [10,000    ] [0.0003251281500] [3,075.71030] [101.50%]
    arrayMergePerformance         :            [10,000    ] [0.0005345391035] [1,870.77053] [166.87%]
    forEachPerformance            :            [10,000    ] [0.0003967431068] [2,520.52268] [123.85%]

[Ocramius]

Composer doesn't seem to be affected since it has a single classmap when optimizing autoloading.

[ThaDafinser]

👍 great work!

[franz-deleon]

@Ocramius i am trying to figure out the code differences from your testds. mind sharing the actual code from the test above? thanks a bunch!!

[ThaDafinser]

@franz-deleon something like this i think

//singleArrayMerge
array_merge($config1, $config2, $config3, ....$config22);

//singleArrayMergeCallUserFuncArray
call_user_func_array('array_merge', array($config1, $config2, ....$config22));

//arrayMergePerformance
$config = array();
$config = array_merge($config, $config1);
$config = array_merge($config, $config2);
//...
$config = array_merge($config, $config22);

[DASPRiD]

What about the performance of array_replace(), or even $newArray + $oldArray?

[Thinkscape]

What was puzzling to me, is that 5.5.8 is 36% faster with array_merge() than foreach.

Apparently 5.5.6 introduced: "Improved performance of array_merge() and func_get_args() by eliminating useless copying.", so the mystery is solved.

The performance gap might be huge enough to sanction a version_compare(). With each new PHP version it will be even more important.

[Thinkscape]

@DASPRiD hahaha... it seems that array_replace() suffers from the same problem but I expected + operator to be way faster... guess what happened ? :-)

http://3v4l.org/BANT2#v5417

Anyways, + is now also an alternative, as classmaps are always assoc arrays.

[Ocramius]

@Thinkscape the 5.5.8 improvement is magic by @dstogov

[Ocramius]

@franz-deleon sorry, thought I had pasted a link. use https://gist.github.com/Ocramius/8399625 to write additional tests

[staabm]

Also interessting that foreach performance dropped significantly since 5.5.6