Question

I am using XDebug as a profiler for a PHP application. I have run into a situation where XDebug changes severely the results in such a degree that they are useless.

Here a simplified example to demonstrate the problem:

function foo(){ $x = 1; }
function bar(){ foo(); }

Test A:

$t0 = microtime(true);
for ($i = 0; $i < 1000000; $i++) foo();
echo microtime(true) - $t0;

Test B:

$t0 = microtime(true);
for ($i = 0; $i < 1000000; $i++) bar();
echo microtime(true) - $t0;

So, these are the results I am getting (in seconds):

          profiler  | profiler  > profiler
          disabled  | enabled   > results
          --------------------------------------------------------------------
          output    | output    > total time   time in foo()   time in bar()
Test A    0.159     | 12.199    > 12.245       0.110           - (not called)
Test B    0.233     | 25.399    > 25.578       0.104           11.068

The increased execution time is expected because of the extra calls to the profiler. The slight variation between the microtime-based output and the profiler results is also expected. I have repeated the tests several times, and the results are always similar.

From the results taken from Test B with the profiler disabled, we can say that the script spends about 0.159 seconds in foo() and 0.074 seconds in bar(). It is evident that time spent in bar() is less than the time spent in foo().

However, when I analyze the results of the profiler (with qcachegrind), the time shown as spent in bar() (= 11.068 seconds) is ridiculously higher than the time in foo() (=0.104 seconds). There is a possible explanation for this: each time a function call is made, the profiler runs some extra code to keep track of the time spent in the call. I believed that it excluded this extra time from the results, but apparently it does not.

[EDIT]
As a result, the profiler says that bar() takes more time than foo() in this program, which is not the case as we have measured with the profiler disabled. It’s not even close! The relative results (the percentage of time taken by each function) are totally wrong. This should not be expected, because, if this is the case, the profiler can not indicate which function takes most of the time. Although the absolute times are expected to have (big) differences, the relative times should not have.
[/EDIT]

This renders the results unusable. Any code that is more modular (with more function calls, wrappers, objects etc.) is severely penalized, although it is not that slower!

So the question is: Is there any way to tell XDebug to ignore or track separately the extra time spent for the profiler calls?

Answer 1

Unfortunatelly, it is not possible. I am no longer using XDebug as a profiler and I suggest that nobody should use it.