I ran the performance tests again with script formatting them
differently – they are aggregated on test function, not on zsh
version. This allows easier interpretation. A new conclusion is that
zhalloc patch doesn't help for string_test. So the first patch (36834)
speeds up string operations and also introduces increased memory
usage.

Best regards,
Sebastian Gniazdowski

On 22 October 2015 at 14:49, Sebastian Gniazdowski
<sgniazdowski@gmail.com> wrote:
> On 19 October 2015 at 19:21, Bart Schaefer <schaefer@brasslantern.com> wrote:
>> We leave that empty arena at the end, but then on the next zhalloc()
>> it's still not big enough, so another one gets allocated.
>>
>> You could try removing that "else" to let the trailing block be freed
>> again, but then we're back to needing to compare speed as well as space
>> across shell versions.
>
> I removed the else, code works longer and memory is still allocated.
> This is seen in mem-test3.txt, where there are many samples and large
> numbers, and in perf-test.txt, where there noted is long running time
> (numbers are milliseconds)
>
> I wrote perf-test.zsh to automatically perform performance tests:
> https://github.com/psprint/zsh-tools/
>
> With it and with mem-test.zsh I tested following zsh variants:
> - 5.0.2 osx
> - newheaps patch (zsh-newheaps-only)
> - newheaps and zhalloc patch 36836 (zsh-newheaps-zhalloc)
> - zhalloc and the previous patch 36834 (zsh-head-both-patches)
> - all three patches (zsh-newheaps-three-patches)
> - a 5.1.1-dev-0 clean zsh
> - the empty arena else removed
>
> My conclusions:
> - 36834 introduces high memory usage; I would suggest to use only two
> optimizations – newheaps and zhalloc; my znavtools are instant fast
> with them, what's slower is searching (wrote search_test function,
> results attached); basically half of pattern search speedup is lost
> with 36834, however what's gained is beautiful memory usage, without
> firefox-like values for RES (~700 MB); newheaps is responsible for
> instant-responsivity of my script, 36836 for much faster searching –
> that's a good compromise
>
> - searching with patterns through very large arrays (700k) allocates
> 264 MB of memory regardless of zsh version; this probably tells
> something
>
> - still not sure why 36834 allocates much memory for the string test
>
> - newheaps makes function calls longer, but it's not a substantial
> difference – 21196 ms vs. 18549 ms
>
> Best regards,
> Sebastian Gniazdowski