Expected resaul is that performance should be at least comparable and not dramaticaly worse

Can confirm this is the case. Some further observations:

* In PHP 7 the time increases quadratically with the number of rows. In PHP 5 it increases linearly.
* PHP 7 goes back to reasonable performance with USE_ZEND_ALLOC=0
* callgrind with --cache-sim suggests that we're seeing a crazy number of LL data read misses in zend_mm_alloc_pages (this line: https://github.com/php/php-src/blob/master/Zend/zend_alloc.c#L856) This being callgrind, I don't know whether this is true

My wild guess: We're going through the list of chunks where the first N are all full until we hit one which isn't. At each chunk we experience an LL miss due to aligned address conflicts.

It's not really necessary to bring mysql into this. Just allocating a bunch of strings without deallocating them in between is enough:

<?php

$n = 50000000;
$s = "aaaaaaaaaaaaaaaaaaaaa";
$a = [];

$t = microtime(true);
for ($i = 0; $i < $n; $i++) {
    $s++;
    $a[] = $s;
}
var_dump(microtime(true) - $t);

This gives me the following results with and without ZMM:

nikic@saturn:~/php-src-fast$ sapi/cli/php -d memory_limit=-1 t043.php 
float(30.064018011093)
nikic@saturn:~/php-src-fast$ USE_ZEND_ALLOC=0 sapi/cli/php -d memory_limit=-1 t043.php 
float(3.7853488922119)

Generally the ZMM is somewhat faster than the system allocator, not 20x slower.

Assigning to dmitry...

Just wanted to add this bug also affects Centos 6 & 7 and still present in PHP 7.0.12.

USE_ZEND_ALLOC=0 appears to be a good temporary alternative, but also leads to other weird behavior. Like this bug https://bugs.php.net/bug.php?id=73370

Could have sworn I posted a more detailed analysis here... I no longer remember the specifics. The issue is that we're doing a naive linear search on the chunk list until we find a chunk with enough free pages. This ends up causing quadratic complexity. I checked how jemalloc handles this, they use an RB-tree to find an appropriate chunk. We should probably do something similar as well -- though it might already help if we just move completely full chunks into a separate list, there's really no point scanning them.

The attached patch should fix the problem in PHP-7.0 (without BC breaks).

The fix for this bug has been committed.

Snapshots of the sources are packaged every three hours; this change
will be in the next snapshot. You can grab the snapshot at
http://snaps.php.net/.

 For Windows:

http://windows.php.net/snapshots/
 
Thank you for the report, and for helping us make PHP better.