Generic 24.04 Calc performance ticket

[caolanm]

Lets track 24.04 separately than earlier releases wrt flamegraph results

[caolanm]

Todays calc 24.04 flamegraph

[caolanm]

Todays's calc 24.04 flamegraphs

[caolanm]

part of todays 24.04 calc test

[grandinj]

Well hello collectAutoStyles(), my old friend - I remember having a bash at this one, with not much success.

However, I can do something about ScDocument::GetColDefault, over here: https://gerrit.libreoffice.org/c/core/+/165346

[caolanm]

todays 24.04 calc session

[caolanm]

todays scheduled 24.04 calc profile

[mmeeks]

Surprisingly unbalanced re: this UpdateCursorOverlay thing ... GetRectsAnyFor should not take that sort of time clearly - for what is ultimately only a few rectangles; looks like a substantial 2x+ perf. regression in the making.

[grandinj]

If the spreadsheet involved had a lot of overlapping/hidden/merged cells, that would explain the time spent in GetRectsAnyFor. Unfortunately, the code involved in overlapping/hidden/merged does not look easily improved.

[caolanm]

Todays ad hoc spreadsheet test

[caolanm]

Todays scheduled call with 24.04

[caolanm]

^^^calc 24.04 session with background save enabled^^^

^^^calc 24.04 session with background save disabled^^^

[caolanm]

^^^calc 24.04 session with background save enabled^^^

[caolanm]

calc 24.04 session

[caolanm]

todays calc 24.04 session, background saved are merged under a single kitbgsav

[caolanm]

Tuesday calc 24.04 profiling session, with mostly automated users

[caolanm]

I did a retest of https://github.com/CollaboraOnline/online/issues/8570#issuecomment-2031853407 with just the watchdog profiler active and did some full column copy and pastes from out typical calc testing document

The amount of break iterator loading looks odd, GetScriptTypeOfLanguage etc.

I wonder if we really need to figure out the area affected in UpdateAutoFillOverlay to send to online via LOK_CALLBACK_CELL_AUTO_FILL_AREA of it we could send an address there instead f.e

[mmeeks]

Surely we can have a one item cache, or somesuch for: SvtScriptType GetScriptTypeOfLanguage( LanguageType nLang ) =)

[grandinj]

That flame graph looks suspiciously like a debug build.....

[caolanm]

objdump -Wi /opt/collaboraoffice/program/libuno_sal.so.3|grep DW_AT_producer gives....

GNU C++20 12.2.1 20221121 (Red Hat 12.2.1-4) -mtune=generic -march=x86-64 -ggdb2 -O2 -std=c++20 -fvisibility=hidden -finput-charset=UTF-8 -fmessage-length=0 -fno-common -fstack-clash-protection -fcf-protection=full -fstack-protector-strong -fvisibility-inlines-hidden -fPIC -fexceptions -fno-enforce-eh-specs

and similar for coolwsd itself

[grandinj]

Then I am guessing that o3tl::strong_int:anyOf is not being compiled properly on that platform, for some reason, because it should definitely turn into just a handful of immediate instructions.

In which case we should change we should change MsLangId::getScriptType to use explicit if/then statements instead of o3tl::strong_int::anyOf

[caolanm]

Todays calc weekly session, online commit: https://github.com/CollaboraOnline/online/commits/cd9875ec core commit: https://git.libreoffice.org/core/+log/07c93a6

coolwsd.xml contains:

true

I wonder why I don't see kitbgsav in this session

[caolanm]

background save as expected

https://git.libreoffice.org/core/+log/421ef05 https://github.com/CollaboraOnline/online/commits/1b586a5

[mmeeks]

Hmm - that trace has 40% of the save time in the foreground - which is unexpected; we should have one non-background save at exit but ... =) Still no sign of: "WRN Failed to ensure we have just one, we have: 2| kit/Kit.cpp:1388" in the logs since the 17th which is good; prolly nailed that then.

[caolanm]

todays scheduled calc testing profile

[mmeeks]

Feedback was very positive on copy/paste today thanks @grandinj =) Digging into the 'GetOptimalHeight' code-path there I see that (oddly) there seems to be an explicit call to OutputDevice::GetTextWidth as a child of GetOptimalHeight which looks like something we could loose easily ? =) And I wonder if there are other height-not-width optimizations we could do there somehow. Anyhow - glad its feeling better :-)

[grandinj]

there seems to be an explicit call to OutputDevice::GetTextWidth as a child of GetOptimalHeight which looks like something we could loose easily ?

Fix (I think) here: https://gerrit.libreoffice.org/c/core/+/167988

[caolanm]

Todays devel version

[caolanm]

Todays stable version

[caolanm]

profile taken during calc testing while I wasn't present, so I'm not sure about this one, but I'll put it here anyway

[caolanm]

Tue test of devel version

[caolanm]

Tue test of stable 24.04 version

[mmeeks]

There is concerningly little threading of delta compression in the last two profiles; it seems we do almost all the work in-line in the main thread there; that points to something bad; I wonder if we (I) have screwed up the rendering coalescing of tiles so we render one tile at a time somehow; if so that would be not good(TM).

[caolanm]

Todays staging-perf calc profile

[caolanm]

24.04.4.3snapshot git hash: 8628721 Collabora Office 24.04.4.20240629 git hash: 2657979

[caolanm]

24.04.4.2 git hash: https://github.com/CollaboraOnline/online/commit/20b6b94a32ca8cb6bc4e7a7950755673e3ab87b0 Collabora Office 24.04.4.2 git hash: 2d2c24d (confusingly this one is master I think and the above is 24.04 branch)

[caolanm]

todays regular scheduled cool meeting calc test

[caolanm]

todays co-24.04 calc collaborative testing session

[caolanm]

weekly collaborative calc testing session

[caolanm]

todays 24.04.5.2 collaborative calc testing session

[caolanm]

todays 24.04.4.5 collaborative calc testing session

[caolanm]

Todays regular scheduled cool meeting calc profile

[caolanm]

calc collaborative testing

[caolanm]

Todays calc profile. I see a non-async SvxCharacterMap::run call in there FWIW

[caolanm]

Todays, share calc profile

[caolanm]

Todays staging calc profile, transliteration wrapper seems hot

[mmeeks]

@grandinj ooh - that's interesting - 8% of our time saving (and prolly lots of memory wasteage) is from ScXMLExport::WriteTable to ScMyDefaultStyles::FillDefaultStyles to ScDocument::CreateColumnIfNotExists - which seems like a bug; are we creating all columns on save/export again somehow ? =)

[caolanm]

Today's calc collaborative session

[caolanm]

Today's calc session on stable version

[caolanm]

Today's calc session in master version with a large copy and paste. If we have background saves, maybe we should have some form of background copy to clip.

[caolanm]

Today's brief staging-perf calc session