Re: Lilypond benchmarking

[David Kastrup]

"Phil Holmes" <address@hidden> writes:

> ----- Original Message ----- 
> From: "David Kastrup" <address@hidden>
> To: <address@hidden>
> Sent: Saturday, September 21, 2013 1:37 PM
> Subject: Re: Lilypond benchmarking
>
>
>> "Phil Holmes" <address@hidden> writes:
>>
>>> David made the comment that we'd no information on the performance of
>>> the latest development release on large project, so I thought I'd do a
>>> little benchmarking. This has been done on windows vista 64 bit.
>>>
>>> I've used 4 benchmarking tests: a) \repeat unfold xx c''4; b) \repeat
>>> unfold 500 { c''4 c' \f c''' g } (this gives the skylining code
>>> something to do, which the simple one in a) doesn't); c) the Finale to
>>> Act I of the Mikado, which I created as code about 3 years ago, and
>>> runs to 496 bars and up to 30 voices and d) The full score for the
>>> Mikado, about 150 pages but set as a number (about 20) of separate
>>> \score blocks.  The main problem I've got is laying the results out in
>>> a text-only email, so I've attached them as a little image.
>>>
>>> Summary: 2.12 was very slow and unreliable on large scores.  2.14,
>>> 2.16 and 2.17.26 are similar: it look like current devel is slower
>>> where there's a lot of interleaving of notes and dynamics to be done,
>>> which is probably to be expected with the more sophisticated skylining
>>> code.  I'd conclude there is no fundamental performance problem with
>>> our current build.
>>
>> The numbers for 2.17.26 are generally about 30% slower than 2.16.
>> That's quite more than the skyline code as such should be accountable
>> for.  Definitely looks like we should bother with some profiling.
>
> It's actually faster with test a), which is why I think it's skylining
> stuff.

That may be as it will, but 30% would be excessive anyway.

-- 
David Kastrup