last steppe
Since dedmen has added some new functionality to profiling we should update the profiling page.
https://community.bistudio.com/wiki/Performance_Profiling
Let's collect some ideas here.
safe vessel
https://community.bistudio.com/wiki/diag_captureSlowFrame
new parameters
bool toFile (if true, doesn't open UI and writes straight to file, note that even if false it might still force to file if it doesn't have a UI, like a server or HC)
scalar continuousCounter (Captures N slow frames. Can be aborted by running diag_captureFrame or another captureSlowFrame)
https://community.bistudio.com/wiki/diag_logSlowFrame
Fixed. infinitely captures every slow frame, can be cancelled by running diag_captureFrame or diag_captureSlowFrame.
Somewhere, chrome export format.
Capturing to file, also creates a .trace file, which can be imported into chrome://tracing (in chromium based webbrowsers), or https://ui.perfetto.dev/
The same format can be exported by pressing the "COPY ALL" button, in the capture frame UI, with CTRL key being held down
stuck hawk
@last steppe do you mean this page https://community.bistudio.com/wiki/Performance_Profiling ?
last steppe
Yes
last steppe
@safe vessel Question:
What is the difference between doing profiling with arma3diag_x64.exe vs with arma3profiling_x64.exe?
~~I might have answered that myself. I guess certain commands are not even available on diag.exe such as https://community.bistudio.com/wiki/diag_captureFrame~~
stuck hawk
diag = development tools
profiling = performance measurement
i'd say
there is some overlap of diag commands, and some only available in either one afaik
last steppe
Yeah, now I am not wiser 😄
stuck hawk
last steppe
I know
stuck hawk
last steppe
Thank you
last steppe
why would there be one
Because we got a special .exe called profiling something something 😄
safe vessel
Yes. Because we cannot play diag exe in multiplayer on stable servers.
last steppe
So profling exe is a dumped down version of diag with profiling commands so that it works in mp?
stuck hawk
ie diag exportTerrainSVG is not in profiling, and all others that would create problems in a branch you can use for MP play
last steppe
Would you agree if I say for profiling use the profiling.exe because it is closer to the stable.exe performance wise?
stuck hawk
you have to use that exe as only that has the internal labels active for profiling (plus a few exclusive sqf commands afaik - checking that atm)
last steppe
Is has the labels? I thought that was internal.exe only?
stuck hawk
at least i would assume the engine only sets theses in certain exe - not all the time
as it creates a (probably small but still) overhead
ok diag.exe has also the profiling sqf commands in
last steppe
So what's the conclusion? Diag.exe or profiling.exe?
stuck hawk
well in general perf branch is updated more often, plus not everything goes into the dev branch
at least afaik the diag.exe is dev branch only: https://community.bistudio.com/wiki/Arma_3:_Diagnostics_Exe
stuck hawk
third point isnt necessarily true
overall the guide probably should start with motivation and target audience
there is broadly speaking:
a) to optimize sqf coding (and to some extent missions)
b) identify engine coding that may have optimization potential (aka identified by anyone but so far only addressable by Dedmen)
safe vessel
So profling exe is a dumped down version of diag with profiling commands so that...
No.
Profiling exe is performance exe but with profiling enabled.
diag is a normal exe, with diagnostics and profiling enabled
safe vessel
Would you agree if I say for profiling use the profiling.exe because it is close...
diagnostics adds alot of extra checks and.. diagnostics that are not in normal builds. So it always performs worse
safe vessel
you have to use that exe as only that has the internal labels active for profili...
What are internal labels?
I'm not aware of such a thing
safe vessel
third point isnt necessarily true
Third point is true. Diag exe is slower.
stuck hawk
well i guess the statement is OK. however it depends on the purpose / goal - for some instances you may want to have the additional diags available while profiling but probably a rare case. looking at those available https://community.bistudio.com/wiki/Arma_3:_Diagnostics_Exe potentially in combination with PX, AI, particles and a few others it could make sense
it seems diag exe also creates the debug.log - wasnt aware of that/forgot. so the additional logging makes it slower. beyond that isnt it mostly slower only if you have a particular diag command activated. at least i'd assume its mostly behind a bool check. yet if the engine has to do more work all the time, then its slower no doubt
stuck hawk
What are internal labels?
I'm not aware of such a thing
so the profiling scopes are also set for the public/regular build?
safe vessel
The profiling scopes are set for profiling builds
beyond that isnt it mostly slower only if you have a particular diag command activated
No. diag has alot more checks (asserts) that are always active, no matter of diag commands.
Every interaction with an array for example, is slower.
last steppe
The same format can be exported by pressing the "COPY ALL" button, in the capture frame UI, with CTRL key being held down
@safe vessel Is this already available?
safe vessel
no
last steppe
Ok
stuck hawk
Since dedmen has added some new functionality to profiling we should update the ...
aside from #1299347444580225055 message (motivation and target audience), as best as i can tell right now, first we need to determine what can be done in what ways. this will change to some extent with the next perf build, and potentially going forward based on Arma internal capabilities available, as well as third party tools
for example profiling sqf code (schedule and unscheduled) you cant really with the current internal tools - mainly lacking is the ability to define scope labels to get picked up for visualization or data logging
what you have right now (from what i understand) is:
for unscheduled there is https://forums.bohemia.net/forums/topic/211626-arma-script-profiler/ - it can define scopes/labels, and using external viewers is quite powerful. the interface of external viewers is fairly complex, thus would also need a guide/use case descriptions
Capturing to file, also creates a .trace file, which can be imported into chrome://tracing (in chromium based webbrowsers), or https://ui.perfetto.dev/
these will need exploration on that
as for engine (non sqf) profiling - the internal viewer is usable, if you understand the general principle of engine computation cycles, and some idea of the parts and it subparts (this is where a table with important scopes/labels will be quite important)
external viewers probably still better - functionality wise for sure; UI/interface wise depends/to be seen
last steppe
Yeah. I want to just quickly explain the intern viewer, but the external one is probably the one everyone should use.
stuck hawk
in terms of profiling approach for that - i think there is two approaches basically:
Enable server performance logging with the #monitords <seconds> command,
Load the resulting server log file using the button below.
cant find a sample image atm
Server load: FPS DDD, memory used: DDD MB, out: DDD Kbps, in: DDD Kbps, NG:0, G:DDD, BE-NG:0, BE-G:0, RQ:0, Players: 1 (L:0, R:0, B:0, G:0, D:1)
monitorDS data is a lot more simple, yet graph visualization is quite important to understand data
ie to hide/filter irrelevant data is quite essential to easily see the relevant parts
would be good to get some other community members involved that either used the internal tools, or external profiling tools - there may easily be different approaches or little but important things to know (in terms of approach/workflow/features/tools/etc)
ie on the bottom left you can switch to averages and sums (default tree view)
the visualization isnt working for the other two, but the sorted display can be useful
another likely impactful change will be the increased MT Dedmen has been working on. this will add a different layer to profiling analysis (but may not be really relevant/understandable without engine access)
stuck hawk
@last steppe should i start posting some stuff, or better do the meet-up first?
last steppe
Sure Go ahead
stuck hawk
important: https://ui.perfetto.dev/ doesnt accept .trace/.any files with UTF-8 BOM - must be Windows-1252 encoding (may happen if you paste the data into a new text file)
last steppe
I tried with a new .txt and it wored. Weird.
stuck hawk
depends on your text/code editor whats default. UTF-8 is for international character set
stuck hawk
so lets start:
stuck hawk
Options to capture:
ASP is currently not working yet with latest profiling branch
ASP is the only to profiling unscheduled code
Visualization options: [WIP]
How to open:
Data import:
captureFrame-ID-DATE.NUMBER.log created by chat commands on the server or toFile logging of diag_captureSlowFrame [or saved/shared logs])Data import: (.trace)
safe vessel
but not (yet?) continuousCounter available (capture multiple cases)
Starting with prof v6
last steppe
https://community.bistudio.com/wiki/User:R3vo/Sandbox#Scopes
@stuck hawk Do you think this is usefull?
stuck hawk
checking
not sure - depends on the goals i guess
like i'd say almost everyone only needs to know the important ones (that take relevant computation time) - if you look for say minifreezes, you would need to know others
the other part is the subscope layering is quite relevant
in general a basic list is still useful - if to be done alphabetically, or grouped in some way (per "main scope"/etc) to be determined
last steppe
My idea was to fill the important one with some information so ppl can lookup what they mean.
stuck hawk
better to have than not to have for sure. just the goal shouldnt be to fill all/at much as possible
last steppe
Well, the list is there now so we at least have a place to put things if we find something interesting
stuck hawk
maybe group at least by main scopes (via separate tables, or in order, or having a colum for "main scope")
stuck hawk
well you can take the main scope from the UI [network, sim, sound, ai, render, vis] (and probably have non main thread ones also separate)
or otherwise grab the main ones from the log
well also a little tricky
can you grab from UI when sorted by duration?
last steppe
I can grab everything thats in the tree view
stuck hawk
also just to make sure - not all scopes are always shown, plus the UI shows less than in the logs
so if you switch to averages and the other via the button it grabs them correctly?
last steppe
Maybe
safe vessel
I had something in mind last night that I wanted to write here. Sadly I have forgotten.
Ah! The. Thing. Eh.
Yes, how to run profiling exe. Because you need to swap the exe files, or you need to rename to profiling one to the 32bit one, so when you launch 32 bit in launcher, it'll run the profiling one.
Basically somewhere on wiki should describe how to actually run the profiling exe with mods
last steppe
Yeah good point.
stuck hawk
alternative (especially if you want different startup params)
safe vessel
only works if RPT has the full startup line.
With many mods, it most likely doesn't
stuck hawk
true. however profiling with many mods..