Skip Menu |

Preferred bug tracker

Please visit the preferred bug tracker to report your issue.

This queue is for tickets about the Dumbbench CPAN distribution.

Report information
The Basics
Id: 81157
Status: new
Priority: 0/
Queue: Dumbbench

People
Owner: Nobody in particular
Requestors: ribasushi [...] leporine.io
Cc:
AdminCc:

Bug Information
Severity: Unimportant
Broken in: (no value)
Fixed in: (no value)



Subject: Inconsistent results of identical benchmark - what am I missing?
I opted to file an RT instead of contacting you by email, so that the conversation is archived for future generations. I am attaching a standalone script which attempts to test two identical copies of a handmade accessor (defined on line 37) subjecting it to 200 cycles of 5 get/set operations (line 64). I tweaked/tightened the Dumbbench options to the best of my understanding (lines 12-28). The machine is silent (no active processes) and the CPU is in a low-power/no speed variations/no turbo state. Since you use Time::HiRes::time() calls everywhere (awesome!), I was able to replace your wallclock measurements with cputime measurements, by monkeypatching Time::HiRes (lines 210-213). With all this I get rather large differences in my resutls. While one could have attributed this to noise, it does not seem to be the case as the *difference* is reproducible. The benchmark starts randomly placing either of the two on top. But further iterations will *always* have the former "winner" on top, every time for the rest of the run. I've ran hundreds if not thousands of benches, and I can't seem to find the culprit. Note - I am not questioning the 2 percentage-points error in the results - I understand this is somewhat expected. I am questioning why there is no random component to this error - why does it always fall a certain reproducible way. See example below. Cheers rabbit@Ahasver:~/devel$ perl bench_wtf Perl 5.016002, take 1: Rate HANDMADE1 HANDMADE2 HANDMADE1 748.422+-0.074/s -- -1.4% HANDMADE2 759.414+-0.076/s 1.5% -- Perl 5.016002, take 2: Rate HANDMADE1 HANDMADE2 HANDMADE1 742.16+-0.074/s -- -1.4% HANDMADE2 752.902+-0.073/s 1.4% -- Perl 5.016002, take 3: Rate HANDMADE1 HANDMADE2 HANDMADE1 738.118+-0.073/s -- -1.3% HANDMADE2 747.668+-0.074/s 1.3% -- Perl 5.016002, take 4: Rate HANDMADE1 HANDMADE2 HANDMADE1 735.658+-0.073/s -- -1.4% HANDMADE2 745.831+-0.074/s 1.4% -- Perl 5.016002, take 5: Rate HANDMADE1 HANDMADE2 HANDMADE1 733.735+-0.073/s -- -1.7% HANDMADE2 746.405+-0.073/s 1.7% -- Perl 5.016002, take 6: Rate HANDMADE1 HANDMADE2 HANDMADE1 737.488+-0.071/s -- -0.9% HANDMADE2 744.129+-0.073/s 0.9% -- Perl 5.016002, take 7: Rate HANDMADE1 HANDMADE2 HANDMADE1 735.209+-0.071/s -- -1.7% HANDMADE2 748.179+-0.074/s 1.8% -- Perl 5.016002, take 8: Rate HANDMADE1 HANDMADE2 HANDMADE1 735.55+-0.072/s -- -1.4% HANDMADE2 746.062+-0.074/s 1.4% --
Subject: bench_wtf
Download bench_wtf
application/octet-stream 5.3k

Message body not shown because it is not plain text.