Performance Regression in 21 - CSS

[Florian Kirmaier]

Hi John,

That sounds very promising!
Let me know when you have a PR available so I can test it out.

Quite sure I measured 10x, but maybe we measured something slightly
different - or it's hardware-dependent.
(The issue with scrolling happened only in my closed-source project, which
has a behavior similar to ListVIew.)

Florian Kirmaier

On Mon, 1 Jan 2024 at 16:37, John Hendrikx <john.hendrikx at gmail.com> wrote:

> Hi,
>
> I think I've found the issue. Note that on my machine I can't really
> reproduce the problem as clearly (Windows 10) as the performance
> degradation is not quite as bad as mentioned in the issue (maybe a factor
> of 2, and scrolling was unaffected), but it's probably what you are seeing
> as little else changed in the PR that seems to be the root cause.  Still,
> VisualVM did hint at the probable culprit (SimpleSelector#applies).
>
> The problem is that BitSet (used for storing and comparing CSS style
> classes) has special logic to do fast containsAll checks if it detects it
> is compared against another BitSet. However, due to a read only wrapper on
> one of the sets this detection fails, and it will fall back to standard
> containsAll logic which is a bit slower.
>
> The fix should be relatively trivial, and I'll create a PR soon.
>
> I did note a few other things in my investigation:
>
> - JFXCentral has around 1000 style classes, which means a BitSet as
> currently implemented, could use 125 bytes of memory for each StyleClassSet
> if a high bit was set
> - Most selectors involve only 1 style class, sometimes 2 and rarely 3 or
> more
> - Most nodes have 1 style class, sometimes 2 and rarely 3 or more
>
> Now, if I were to implement this kind of logic knowing that the number of
> possible style classes can be quite high, and knowing that most selectors
> and nodes will rarely have 3 or more style classes assigned to them, I'd
> probably not use a non-sparse bit set as it is currently implemented.
>
> I've played around a bit with a sparse bitset, and although I did not see
> amazing (overall) performance gains, the change (after applying the
> regression fix) did shave off another 10-20 ms for the "loadTimeFX"
> measurement in the JFXCentral application.
>
> --John
>
>
> On 31/12/2023 15:24, Christopher Schnick wrote:
>
> Hello,
>
> I just tested this with our JavaFX application and can confirm that there
> are massive differences. It takes around 1-2 seconds to completely apply
> all application stylesheets in JavaFX 20 but takes around 6-7 seconds in
> JavaFX 21.
> On 12/31/2023 3:00 PM, Florian Kirmaier wrote:
>
> Hi Everyone,
>
> Sorry for the delay - but I couldn’t find the time to extract the
> TestApplication for this bug.
> Luckily, I found another application, which is also open source, which is
> affected by the application.
>
> I'm speaking about https://www.jfx-central.com/ - both the desktop and
> web versions are affected.
> I’ve seen a performance deterioration of 10x when switching pages when
> using JavaFX21 compared to JavaFX20.
>
> I’ve created a ticket with further instructions on how to test it:
> https://bugs.openjdk.org/browse/JDK-8322795
>
> Greetings
>
> Florian Kirmaier
>
> On Fri, 27 Oct 2023 at 21:31, Andy Goryachev <andy.goryachev at oracle.com>
> wrote:
>
>> Please create a ticket, Florian.  Would it be possible to profile the
>> application when scrolling?
>>
>>
>>
>> Thank you
>>
>> -andy
>>
>>
>>
>>
>>
>>
>>
>> *From: *openjfx-dev <openjfx-dev-retn at openjdk.org> on behalf of Florian
>> Kirmaier <florian.kirmaier at gmail.com>
>> *Date: *Friday, October 27, 2023 at 04:20
>> *To: *openjfx-dev at openjdk.java.net <openjfx-dev at openjdk.java.net>
>> *Subject: *Performance Regression in 21 - CSS
>>
>> Hi everyone,
>>
>> I've noticed that some parts of one of my applications is significantly
>> slower with 21. It's fast with 20.
>> The application heavily uses (and reuses) TextFlow with a Cell pattern.
>> When I scroll, it is smooth with 20, but has big freezes with 21.
>>
>> I've tried all the commits that happened in between, and pin-pointed it
>> down to the following:
>> ticket: https://bugs.openjdk.org/browse/JDK-8304959
>> PR: https://github.com/openjdk/jfx/pull/1070
>> commit:
>> https://github.com/openjdk/jfx21u/commit/3fa02ee96a6dadbc20cacbf399a2d65df708eee1
>>
>>
>> According to the description and the discussion in the PR - this wasn't
>> supposed to change any performance.
>> Is this regression known?
>> Otherwise, I should create a ticket for it.
>>
>> Greetings Florian
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/openjfx-dev/attachments/20240102/55cf18ad/attachment.htm>