Perhaps you can process that output so that for each sub called, its different run-times are recorded and see which vary and which don't. For example I can see that YAML::Tiny::* does not vary at all. (btw some subs are not present in all tests).

"btw some subs are not present in all tests"

That's the first thing I noticed, and that definitely doesn't appear to be a fair profiling comparison.

yeah that makes sense, I'll put them in a spreadsheet

As for missing subs...some of the sub durations seem to vary slightly,
so they trade places in the top 15 report,
but after anything taking 100ms or less I don't think I care

if your subs can be broadly categorised as disk-intensive, memory-intensive, cpu-intensive you can see which of these 3 groups varies substantially against the others (standard deviation is an indication of variation). For example that YAML* sub being invariant in all three runs while others with similar run-times were not. Perhaps perlperf can help you.

Also, are you 100% sure that your program is deterministic? Are there no random choices? For example, even iterating over the keys of a hash is non-deterministic. What if the algorithm benefits if the longer keys are processed first? Also, regex are also working internally with non-deterministic algorithms (that's my understanding on when a regex consists of a "choice") and therefore may show variability in termination (see https://stackoverflow.com/questions/36420517/is-it-faster-to-use-alternation-than-subsequent-replacements-in-regular-expressi).

If you want to exclude the possibility of non-deterministic behaviour and other program-specific factors, start by doing some benchmarks yourself. See Benchmark on how to profile subs yourself easily.

People from my social circle constantly harass me with their windows-10 too-slow problems. Sometimes I find that they have background updates sucking up resources (sometimes it's a cryptic generic windows process name).

bw, bliako

I shutdown all other programs on my laptop before running this and I got much more stable results...
The time still varied, from 7 to 10 seconds to run the critical portion of my code, but thats much more reasonable
Best I can guess is the other processes were stealing lots of cycle times from my process

Just having gmail open seems to add 3-5 seconds to my total profile time...

Is your software available as a distribution that can be reasonably easily installed?

I don't write Perl code on Windows, but I do write a single piece of Windows software for Perl, and since I'm a test-first-mentality type, I have several minimalistic Windows VMs for testing on, and I'll gladly give your code a spin against these extremely stable systems if you want to see the variances against machines that are designed and designated for doing these kinds of things.

Let me know.

Thanks for the offer! I could provide you a zip file with all the necessary code to run the test but I don't think its needed.
I distribute this code as an executable (which extracts perl locally and then runs it)

Unfortunately the code will be run on general purpose laptops/desktops,
where I have no control over what is stealing cycles from my process
One of the worst cases I encounter is when I am demoing this code to a prospective customer on a web-Ex
And that really seems to slow the whole process down tremendously

I also distribute this on linux and have been playing with the profiling there to get a more stable result
That VM does give better results varying by 1 or 2 seconds.
Although the very useful nytprofhtml is not working for me on that VM :)


I guess one thing I was hoping someone might tell me that there is a way to
run the perl code as a higher priority windows task

You gotta spin up the laptops, I do  perl -le" $fo++ while 1" before doing long benchmarks, and wait until the fans start hissing, otherwise first runner is always slower ...

Interesting... Then you ctrl-c the perl -le and run the actual code?

I found this interesting nugget to enable win10 ultimate-performance-mode on my laptop...

https://www.windowslatest.com/2018/04/20/enable-ultimate-performance-mode-in-windows-10

it makes my laptop fan stay on...
It did not help the code execution time while I had a browser open
I will try with the browser closed.

Not too much of a difference with browser closed
What really helped to stabilize my results was Turning OFF WIFI

Even then...
There is still too much variance to determine if a new looping structure using a large pre-processed list
is better than 2 loops that created the same list on the fly (using foreach and keys)

I will look more into Benchmark as suggested by bliako

I would be very interested to find out what happens when you run this code on another OS.