perf_prof_branch

[weak panther]

no matter what quality setting set

[whole cloud]

probably not on profiling so you can't do a frame capture for me?

[weak panther]

You right sorry, probably I didn't wait for update to be installed, on the second launch I saw some particles and then got the status violation

[whole cloud]

If you got a crash then please send me the crash report

[whole cloud]

⚠️ v29 was replaced with 2.18.152842.
Crashfix for ui2texture @thin wyvern
Crashfix for netcode @fast hornet

[vale shoal]

Hi! I've also sent you a crash dump a couple of days ago. But I don't know if its still relevant. Just wanted to give you a heads up to it.

[whole cloud]

Ah yeah you are too far down the unread DM's stack

[foggy vine]

Particles are gone after I went AFK for a bit, died from AI, picked another slot in lobby, and spawned in

Got RPT on Gist if it's needed

GPU usage is on 100%?

Goes away after a bit when I Alt+Tab back into the game and wait a while, but comes after another Alt+Tab out of the game

Got my FPS back after another re-slotting along with my proper GPU usage and FPS more than 1

[whole cloud]

The gpu usage, or the particles?

[foggy vine]

GPU usage

With working particles and no lag spikes from GPU usage:
Around 50% GPU usage when I'm moving mouse around like a maniac with zoomed-out FoV, but I haven't checked my usual GPU usage on stable for comparison. No smokes or grenades deployed

I guess LoD taking effect on grass mostly when I'm approx 5-10 meters in front of it is intentional/known? 🤔 I've seen a fair bit of LoD popping up out of nowhere on objects I'm rather close to

[whole cloud]

probably intentional

[restive pilot]

I think I've seen this issue when you pass invalid stuff to the GPU

[foggy vine]

Such as?

[restive pilot]

not sure but I think invalid floats was one (NaN, Inf, etc)

[heavy vortex]

This is task manager GPU usage?

[foggy vine]

Yeah

[restive pilot]

at what frequency tho? is it maxed out?

[heavy vortex]

I would guess it just asks the driver these days.

[foggy vine]

https://i.imgur.com/kdFh6M7.png this is what I've been looking at instead of the Performance tab at the time of it happening unfortunately

[heavy vortex]

wallpaper32.exe 👀

[foggy vine]

Related or just something that caught your eye? :P

https://i.imgur.com/D5xOz9U.png 3D peak value is me zooming in on a hill ~1.5 km away from me with 3km view and object draw distance

The 0% usage is me alt tabbing out of Arma, then alt tabbing back and looking without zooming in

adapter description : NVIDIA GeForce RTX 3050 Ti Laptop GPU 🤷‍♀️

[lean light]

Maybe thermalthrottle?

[foggy vine]

Hmm. Would it display 3D usage like that from thermal throttling only in that case?

[lean light]

My english not so good so ->
||Видеокарта может троттлить, сбивать себе частоту до ~50%, арма утилизует всю достпную мощность и диспетчер показывает 100%. Проверь через MSI afterburner OSD. Там видно будет действительно ли игра выжирает всю мощность, или это термотроттл. @foggy vine ||

[foggy vine]

Good point, thanks. Forgot that exists

[magic elm]

DM'd my crash reports from yesterday on the previous v29

[woven loom]

Yup, it's common

[woven loom]

Pretty typical honestly, Arma 3 is limited by the CPU's performance

Look at the sky for high GPU usage 😄

Also keep in mind the GPU might downclock to save power, in which case GPU usage may jump higher. And then when it clocks up again, GPU usage will drop.

[foggy vine]

Hm, true. You learn something new everyday

[magic elm]

152842 fixed the Headless Client issue we were having, running smoothly now

Vehicle smokes arent appearing though for us

Like the smoke grenades

[heavy vortex]

What, like consistently failing to appear for everyone on the server?

[light cargo]

task manager reports % incorrectly because it bases it on power scheme and stuff like that

process explorer or process hacker or gpu-z show the real usage

[woven loom]

Or HWINFO

[light cargo]

hwinfo is primarily for temp sensors not for % usage of cpu/gpu

[swift narwhal]

A friend of mine reported he got his particle effects working again after lowering his view distance. Had him repro by upping and lowering again.

Maybe this is useful Information

[woven loom]

Bruh

[azure jewel]

I love this update

[whole cloud]

So upping it, made them disappear?

Maybe that's why I can't repro, maybe I have different view distance than the affected people

[inland dew]

looks like higher draw distance = no particles
lower draw distance = particles are present

but overall/terrain or exactly/only objects distance and what's the exact value and particles quality setting

[swift narwhal]

I just asked again to be sure,
he didn't repro like i asked.
But lowering view distance definitely fixed his missing particles.

I'll ask him to try playing around with his view distance later this evening and see if particles go missing again on higher view distance.

[potent sparrow]

what basic.cfg are you using if you don't mind - I played with the network settings years ago and stuff has worked fine since so much so I don't have a clue what the values should be for the profiling branch anymore 😄

[whole cloud]

I need to find a way to make the particles disappear.
Not a way to fix it after they've disappeared

[inland dew]

ask what's particles quality is set to and whether only objects distance and/or overall/terrain distance affect that
it's a big difference

be as specific as possible, so dedmen can finally repro, thus fix it

[whole cloud]

That issue goes back till 2.16, a year ago.
And we weren't aware of it until you reported it now

[wise sparrow]

Huhh, interesting 😄

[fading seal]

Are these client side improvements supposed to work with proton and/or older CPU tech (Ryzen 2700x (zen+))? I haven't seen any big difference compared to a year ago. As soon as the AI or I get in a battle or inside a town the fps tanks to like the 20-30 range. This is with just me playing and using a (remote) dedicated server to offload the mission and AI. Could be the DLC content/mission/mods but I guess I'd ask before trying to find a problem.

[inland dew]

ryzen 1000-3000 are very very bad in arma in general

[fading seal]

It's kinda frustrating seeing the CPU sitting at low-ish utilization and the GPU at like 30-40% 😆

[inland dew]

check per core and not whole cpu

[fading seal]

I know, none of the threads or cores are being used to their full potential

[pseudo zealot]

Just popping in to say the work in this space is amazing, can't believe the performance increase this gives 🤘

[whole cloud]

yes. I don't check what CPU you're using to then turn off all improvements 😄

[fading seal]

https://imgur.com/a/wZrTvs5

this is me just sitting at the edge of the map with some random AI patrolling

(can't upload a normal image for whatever reason 🤨)

[whole cloud]

capture frame to see what its doing

I assume you're unhappy about the high cpu usage

[patent sky]

low in this case, seems like some sort of frame limit is going on

[fading seal]

yeah low CPU usage mostly

[whole cloud]

95% on main thread? looks normal to me

[fading seal]

the 95% load of the main process isn't reflected on any of the cores though. It should show at least 1 core bar thingy near fully filled

I have a captured frame, how do you want this data?

[whole cloud]

can just drop the .trace file in here

[whole cloud]

Not if the scheduler keeps moving the threads between different cores.
The core display on top is averaged

[fading seal]

it seems I can't upload files here?

[patent sky]

gotta accept the #rules iirc

[fading seal]

ah.. been here for years but I guess new rules lol

the thread scheduling thing seems kinda probable I guess

[whole cloud]

The frame looks okey. All looks like expected

[thin wyvern]

For proton users https://www.phoronix.com/news/Proton-10.0-Beta-Released

[fading seal]

I've been using experimental so I guess that doesn't matter much (for me anyway)

[fading seal]

Alright thanks. 🙂 I guess it's just too slow to benefit from the optimizations 😢

Well I could get rid of most mods and see if that improves anything

[potent sparrow]

I had a 2700X before I upgraded - I feel your pain, it was an excellent CPU for it's time generally but it ran arma like a drunk one legged man

slap a 5700X3D in that bad boy and you should see a significant uplift, arma really likes the improved IPC but it really likes the X3D's

from a while ago (running mainline with full modset I used at the time)

see bottom two for X3D vs nonX3D (have a 7950X3D for ref so I can actuall bench X3D vs non X3D by pinning)

[fading seal]

Neat, yeah I want to upgrade the CPU at some point. Just waiting for a better deal I guess 😛

[potent sparrow]

between the non X3D and X3D CCD there is a 23ish % uplift on X3D despite the X3D CCD locking at a lower freq than the non-X3D

neither was thermal throttling (a 420mm RAD in a 7000D means throttling isn't a thing to worry about tbh)

[fast hornet]

arma just really really likes the big L3-Cache.. 🙂
(and so does Factorio btw.)

[potent sparrow]

pretty much all games do - it's almost the perfect fit between use case and use - much less of an uplift on other types of software but that's why I love the heterogeneous CCD's on the 7950X3D/9950X3D - "Do you want productivity or gaming?" and AMD just pulled a "why not both" card

[inland dew]

please -> #hardware_vs_arma

[fading seal]

no mods, in the editor, edge of Altis looking at the sky and the GPU is still sitting at 40%. I guess I'll play the Virtual Reality map at 400 fps with 99% GPU usage haha

[fast hornet]

on another note, 152842 has been running great for us for the last 24h. Seems to be a "good" version 😄

[potent sparrow]

Same, We beat on it with a hammer last night (I have a training mode on our zeus server that spawns a collosal amount of AI dynamically (and continuously as it gets killed) with a lot of automated waypoints and stuff - I use it for testing since it stresses the server more than most zeus ops would - took it like a champ

[whole cloud]

I'd have fun in seeing some bandwidth statistics with a server setting sockets/maxBandwidth to a higher value and blasting clients with data 😄
I topped out at 80MB/s but maybe a proper server can reach more?

[fading seal]

Gave it a try, realizing that experimental isn't the same thing (wine 9 vs 10 etc). The only difference I'm noticing is that the little launcher dialog thingy starts centered on screen instead of the bottom and the sun (and rays) seem to render differently (could be placebo or a little shift in time or whatever but it feels different)

No difference in performance (for my under powered system anyway)

(and mouse capturing is different, I have to click now before it captures after moving from different monitor (using dwm as my window manager))

[potent sparrow]

I mean maybe but my dedi would cap at 120MB/s under perfect conditions - you are limited by the technology of your time 😉

[inland dew]

view distance 500 m and here goes your gpu usage.
view distance higher = cpu/ram limit

[fading seal]

I wish, even at 500m I'm CPU limited. Setting object view distance to 1 bumps up the FPS a lot but GPU still just chilling.

I guess a tiny silver lining is that I can run at 200% render resolution with barely any FPS drop lol

[fast hornet]

well, your CPU was already not great when it released and by now it's ancient as well. You won't get better performance in Arma with it.

[fading seal]

Yeah and I guess the proton overhead, if any, doesn't help much either

[spiral pond]

Any meaningful way to monitor/compare the network throughput after the rework? (besides basic in and out max)

[whole cloud]

the out max, would now be able to go higher, if the server needs it

[knotty wraith]

for some reason the game on 14th generation Intel processors loads only one thread - the latest and most inefficient one, with the lowest frequency

and in general only E cores are involved

https://imgur.com/b0lcKKr

I've been looking for a solution all these days, finally found it - disabling E cores in the BIOS... But I was shocked in the first days, why the game chooses the latest and most ineffective thread and only E cores

[whole cloud]

The game doesn't choose anything, it lets the operating system choose

But maybe start parameter helps the.. the thing that sets the gaming thread flag -setThreadCharacteristics?

That tells windows, that the threads are for "Gaming"

[knotty wraith]

https://imgur.com/QscgrCh
it didn't work

[whole cloud]

@wise sparrow I tracked the issue back to 2.12, from february 2023. Its still broken all the way back then.
I don't know whats going on there.
I have also seen youtube videos from 2/5/9 months ago, where it works fine.
But I tested over a dozen builds back till 2.12 and it was broken for me in all but one, and that one I retested and it was broken again on second try.

Either QA (and everyone else?) missed this for over 2 years, or something outside of Arma has changed (and the 3 youtubers who's campaign gameplays I've watched don't have whatever that is)

[knotty wraith]

I suspect that Windows 11 will solve this, but I don't really want to check it)

https://imgur.com/UTT4lfJ
another fact is that simply disabling cores in Lasso doesn't help, terrible lags

[whole cloud]

If you disable cores, you need to use -cpuCount parameter to tell the game about how many cores it has left.
It can detect how many cores you have, but not if any are disabled

[gritty wasp]

first rule for any modern intel CPU: Disable HT and all E cores and forget.

[potent sparrow]

Process Lasso 😉

lets you lock whatever to whatever cores you want

[knotty wraith]

I didn't think that cpucount still works and is relevant) yes, thanks - I set 6 cores, E cores don't work
https://imgur.com/2KtU7tv

https://imgur.com/CXd7RMq

[potent sparrow]

or run your dedi on one of those and just use taskset -c 0-11 and lock it to one of the 6 perf cores 😉 - let the E cores run teamspeak

there is no reason to force disable E cores when you can tell windows or linux to don't use X for Y in software

[wise sparrow]

Hold on, even stable builds? And yeah didn't see it on the playthrough I watched before

[inland dew]

@whole cloud since you've implemeted this.
the related ticket can be closed
https://feedback.bistudio.com/T152279

[potent sparrow]

then process lasso (free version works - I bought it because it's awesome) it, that was how I locked arma to running on the X3D CCD before windows scheduler/drivers caught up 🙂

[whole cloud]

Yes

[wise sparrow]

Ain't that weird

[potent sparrow]

I knew about that - but it just occurred to me, does the same apply if you pin arma to specific cores via taskset/lasso, i.e. if you pin it to P cores on a P/E arch do you have to adjust cpuCount to the number of P cores?

[whole cloud]

You have to adjust the CPUcount, to how many cores are available

[gritty wasp]

they are not visible to system if disabled in bios

[whole cloud]

We weren't talking about bios

[potent sparrow]

ah..interesting, I've used both Lasso and taskset to lock arma onto specific cores for ages but never occurred to me to chang the cpuCount 😄 - brb breaking everything 😉

[whole cloud]

With the new multithreading stuff, it behaves quite badly if it has fewer cores than it thinks it has

[potent sparrow]

yeah that I've run into (or related), I got smart and tried to lock HC's to one set of cores per HC and the server to another - thinking that might be better....it very very much was not

[potent sparrow]

cores or threads, I.e. 7950X3D, 2 CCD's, each is 8C/16T, if I lock to CCD0 (the X3D one) should cpucount be 8 or 16?

[whole cloud]

Threads

[knotty wraith]

https://imgur.com/X7AAJkb
@gritty wasp
thanks for the tip - this turned out to be the best solution, disabling HT and E
the processor started working...

[potent sparrow]

🙂 ty, I put 16 then thought..hmm should ask

[whole cloud]

Well. CpuCount says how many threads Arma will spawn.
The real cores vs fake threads thing.. is a thing you'd have to test

[heavy vortex]

Well, 8 might still be better, but 16 will be fine.

[fast hornet]

16 is the correct value for one CCD on a 7950X3D. 8 cores, 16 threads.

[potent sparrow]

yeah I'll YAAB the options and see which plays nicest, for the server I'll just match taskset to cpuCount anyway which is 0-11

[gritty wasp]

There is no issues with core detection if you disable and forget. Think about it 😏

Extra heat from e cores can be converted in Hz for performance cores. For me there is noticable difference in YAAB

[potent sparrow]

off topic for here - but that can depend and only applies if the CPU is actually unlocked (7950X3D aren't they'll only boost to their locked freq limit) and they are thermally capped - in my case its never thermally capped - if they are actually unlocked (9950X3D's are and will boost to their TJMAX) then its different, arma is CPU heavy but it's not stress-ng on 32 threads CPU heavy

@whole cloud well...damn 🙂 - that's a shiny little boost for 16 in the box ~10% (everything else was identical)

[inland dew]

#hardware_vs_arma

[spiral pond]

sframe cheat to get a capture of that 0 fps / mini freeze could be of interest.
the drop is normal but shouldnt be that extreme

[inland dew]

might also be obsolete and thus can be closed?
https://feedback.bistudio.com/T123355

[potent sparrow]

ran it again identical to run with the frame drop, same result +/- 3% (69.9) second time, never dropped below 43 - so yeah curious but can't reproduce I'm an idiot, I ran the wrong params 👀

[whole cloud]

I don't remember implementing that xD

[whole cloud]

No I like that one

[heavy vortex]

Wait, is that still an issue?

[whole cloud]

Probably not

[inland dew]

maybe not you, but a separate pip distance slider is now there and doesn't affect pip quality, as before it was only select quality and thus also automatically a preset distance, depending on quality, with no separate slider

[whole cloud]

Oh right now I remember

[inland dew]

so the ticket is about adding that pip distance slider and separate pip distance from pip quality.
thus ticket implemented and can be closed

i guess so...

[knotty wraith]

https://discord.com/channels/105462288051380224/108187245529268224

[naive osprey]

Profiling server and non-profiling clients may have a potential issue with.... animations or something? Still have no idea what happened at the end of the video, never seen it before
https://outplayed.tv/arma-3/daRg5n

Not home so I can't replicate right now

[fast hornet]

"he's hacking"

[fair swan]

AI in static weapons were appearing about 10 meters in front of their seated positions, then snapping back to normal location every few seconds. Issue was resolved by switching the server back to main branch.

The weapon they were sitting in would also rotate when they appeared back at normal position for a very short amount of time, almost unnoticeable but can be seen in the pic above

[magic elm]

What Dedmen said about socket bandwidth had me curious so I upped mine from 25 to 128 and ran smothly. Wish my rcon had been working so I could get hard stats will try that again tomorrow but looking at the logs even trying to put the server into desync with 50 people we just couldn't. There was some small local desync but theres nothing as far as pending messages between the server and HCs, no transport layer issues and no backlogs I could find

[whole cloud]

The particles disappearing being related to view distance doesn't make too much sense though, as people reported getting the bug after respawning, where they wouldn't be changing the view distance 🤔
Animations could be that one animation message sending I changed, but I think that was v28 or earlier

[spiral pond]

Are particle creations sent as network message, or rather created locally from other type of network messages?
If the latter, and always related to player position changes, could the camera source/source to determine what particles to draw not get updated?
Finally a debug of sorts to show amount of particles drawn, may help to understand better whats going on?

[whole cloud]

particle sources are network created, particles themselves are local

Huh interesting, just learned that object view distance, doesn't apply to particles. They still show when everything else is invisible

Messing with view distance yielded me no results

[swift narwhal]

Yeah we tested yesterday night, this time his particles were present throughout the whole session. We tried messing with his view distance but nothing happened this time around...

[inland dew]

was just a coincidence then...

[wise sparrow]

Noticed missing particles yesterday when I loaded into a saved campaign mission, went away after loading up editor messing around throwing a smoke and loading back into the mission. So no clue what might've triggered it 🤷‍♂️

[fading seal]

I've had particles disappearing as well a few times over the last couple of weeks. It seemed kinda random but the most memorable moment was me being a commander in a spearhead sherman tank

Could it be related to VRAM? Sometimes when VRAM is full and it seemingly is trying to load more stuff into VRAM it'll turn into a 3 FPS dog until I run the flush and/or superflush command. I feel like the 3 FPS stutter fest and particles disappearing happened at somewhat the same-ish time

(unrelated question; anyone know why this channel doesn't want to stay in the sidebar? It's toggled on in the channel overview thingy but as soon as I change channels it's gone) Never mind... apparently it was muted 😂

[opal hound]

What could cause somebody to become stuck in the lobby without being able to pick a slot or press OK?
We've upped MaxMsgSend as suggested before and it's still happening occasionally

[whole cloud]

Cannot press OK until slot is selected.
Clicking on a slot, just sends a request to the server for that slot. If the server refuses it, then you won't see yourself pop into the slot.
Could also be that the server is lagging (like when mission is being loaded at start) that it doesn't send response

[fair ore]

...significant performance opportunity?

[opal hound]

What do you mean by mission loading, Is there something separate to the mission file download?
We haven't added anything to refuse picking a slot and it's only started since we switched to profiling server. Is there any data we can grab to troubleshoot?

[whole cloud]

no

[whole cloud]

after mission file download, the server loads the mission. loads terrain, spawns the objects, runs init scripts.
During that its essentially frozen

[opal hound]

I believe this is before mission file download

[whole cloud]

Other players can slot, just one player can't?

[opal hound]

Yeah

[fair ore]

Insignificant performance opportunity?

[whole cloud]

Maybe its a missmatch because of non-profiling branch client?
I think server RPT should log if a role selection is refused, but it happens so rarely that might also be broken.
If client reconnects, they can slot?

[whole cloud]

About as significant as joey

[opal hound]

Yeah they're likely stable clients, I can try to get an RPT from them
They've been verifying/reinstalling and having the same issue so no, rejoining doesn't fix it

Yes

Interesting, I'll see what I can find out

maybe lol, but they're complaining in the support channel so seems unlikely

[fair ore]

so very significant

[whole cloud]

probably not

[craggy matrix]

anyone heard of the parachute extesion mod ?

"extension"

[quaint flame]

wrong channel

[foggy vine]

Same shit different day, just had it happen from the usual alt-tabbing in and out. GPU 3D cache usage remains at 100% when this happens. Sometimes taking a few seconds to get back to normal, sometimes minutes.

[jagged hemlock]

I my self got a long time problem that by using my mouse button which emulates pressing ";" for PTT, after some time i also can't choose slot which is fixed by lobby reentry and choosing slot without pressing it.

[whole cloud]

Everytime you change the mods you're launching, Arma has to recreate the AnimDataCache.
That takes 10+ seconds. But luckily it caches it and if its cached it takes only about 1 second to load the cache.

That's of course annoying if you change mods often, having to endure that slow loading.

But, if we throw multithreading at the creation, the non-cached start, with my modset goes from 15 seconds down to 5.
8 lines of code, 10 seconds faster game start

And I can feel better about CPU utilization.

[restive pilot]

did you also multithread the config loading?

[whole cloud]

No, can't do that because it contains scripts and thats a mess

[restive pilot]

what about caching it?

[whole cloud]

but I want to try caching it, that should cut off 9 seconds but that might not work well

Configs with #if could enable/disable config entries based on date.
Which doesn't go well with a cache

[whole cloud]

I planned to do config cache since 2021.

But then in january this year I made a improvement that shortened the config merge time from 17 seconds down to 2.

Cache to save 17 seconds very worth it, cache to save 2, eh. That is not config parse though, that is the merge step post parse (so after the #if troubles above are already resolved)

I could also cache stringtables, but then I discovered binarized stringtables, which load alot faster. And improved our stringtable parser too so that is now also very minor 😄
(Now I just need the CUP mods to use binarized stringtables 😢 )

[whole cloud]

Found another second, with a single-character code change.
Every filename in every pbo is stored as a deduplicated string. That would save memory if the same string appears in multiple places.
But, filenames in PBO's should be unique, there should never be the same file in multiple PBO's, so duplicates don't happen and we pay time for that deduplication.

After fixing that.
~60% of pbo loading time, is spent setting up a lookup table so we can find which file is at a specific offset in the PBO.
We only use that table for debug logging. Setting up the table is also quite inefficient.
We check for every inserted file, that there are no duplicates. But its not possible to have duplicate files inside one pbo (atleast with a properly made pbo)
And we re-sort the table after every insert. Instead of sorting once at the end, there's even a todo comment on it saying to sort it only once 😰

If we first insert all and then sort, the code is alot faster and its also just 1/3rd the lines of code to do it.
It is so much faster, that it doesn't even show up in profiling anymore.

PBO loading from 3 seconds down to 0.5.
Saving all the precious seconds.

Another 1.5 seconds to be had if mods would use script bytecode 😢
Another 2 seconds if mods would binarize their configs.
Another second if mods would binarize their stringtables.

[restive pilot]

you can binarize stringtables?

[whole cloud]

😄

Indeed.
Addon builder doesn't do it, afaik mikero also doesn't, HEMTT does now

[quaint flame]

Are Vanilla stringtables binned?

[whole cloud]

in 2.20 yea

[restive pilot]

speaking of stringtables, is this doable?
https://feedback.bistudio.com/T190844
I would love to be able to localize to a language independant of game language

[whole cloud]

no

We only load the strings that are needed

Well okey technically yes.
By re-loading all stringtables for that lookup

[analog acorn]

Does that mean they'll no longer be readable?

[whole cloud]

We still include the unbinarized xml's in the pbo's if you want to read them

[autumn timber]

at least with a properly made pbo
Custom obfuscated pbos breaking in 3... 2... 1... 😆

[silk pewter]

https://tenor.com/view/ares-god-of-war-xwp-xena-warrior-princess-kevin-tod-smith-gif-20224131

[autumn timber]

Last time I checked (must have been over half a decade ago), I think that the game just took the first available file with a name, but unpackers would overwrite it on disk with a second one, if you had them unpack the whole file to a directory

Or was it the other way around? I don't remember. Not that I care. Throwing it here as trivia 😀

[whole cloud]

Well if obfuscated pbo's break.. Fine by me

I just broke some last year

[whole cloud]

Now that I got my game start down to 35 seconds.
And to 33 by setting process priority to highest, during the loading 😄

Do I want to save another 8 by caching the config and stringtable 🤔

[patent sky]

preload the game on windows startup, instant start up times 😄

[heavy galleon]

[quaint flame]

Click to run for Arma, why not 😄

[fast hornet]

https://tenor.com/view/no-nope-cat-cute-gif-4544032

[scarlet jolt]

sounds like windows is the problem. Bare metal arma 3 2025?

[fast hornet]

complete assembler rewrite in 2026

[spice dirge]

FPGA clients in 2027

[fast hornet]

Arma 3 ASICs in 2028

[silk pewter]

inb4 ArmaOS® 2035

[heavy galleon]

Which Linux will it fork?

If it's not Hannah Montana Linux, I'll riot

[whole cloud]

My v28 with empty animcache takes 70 seconds till main menu
the new one takes 47 seconds 😮

[restive pilot]

is a different Montana ok?

[whole cloud]

oh wow. Stringtables can actually run through preprocessor? Why

[restive pilot]

splitting and #including them?

[whole cloud]

Aw caching configs doesn't work, or would be too much effort to make work.
I could do stringtables, for about 2 seconds saved

[fiery crescent]

I keep following this channel, and wanted to say that it is amazing how much you guys do for us!👌👍😊🥂

[swift narwhal]

I keep following this channel and wonder why it hasn't been renamed to #dedmens-dev-diary

[worldly badge]

"guys" yeah indeed we have bunch of Dedman

[autumn timber]

The contents of this channel would unironically be a great source for a lecture at a gamedev conference.

The title should have something about skeletons in the closet 😛

[quaint flame]

Thanks for the ear worm

[silk pewter]

"spooky scary skeletons"?

[quaint flame]

I got some skeletons in my closet and I don't know of someone knows it so before they throw me inside my coffin a close it...

[silk pewter]

glad I didn't know that 😎

[potent sparrow]

Eminem

[silk pewter]

oh, I found "liltoe" xD

[fair ore]

tech debt: a saga

[light cargo]

a saga of treadmill work

[real creek]

dedmen documentary when

[obsidian condor]

Would they have to wear a bandana face mask in it though?

[verbal shell]

Why the CUP ones specifically?

[whole cloud]

They are big

[verbal shell]

Ah, well, fair

[potent sparrow]

Arma 3 Brownfield: Always code as if the guy who ends up maintaining your code will be a violent psychopath who knows where you live.

[whole cloud]

https://github.com/ValveSoftware/GameNetworkingSockets This is basically what our netcode does, but in alot worse and more amateurish 😄
Maybe we could just switch to that 🤔

The cool thing is that it handles NAT traversal. No more port forwarding just to invite a few friends to your local hosted server.
Just host server, and steam invite friend and it just works? that'd be neat

Their ACK is alot nicer.
We just have a bitfield of 32/64 acks.
They do it with RLE encoding 🤔 Very schmart

[fast hornet]

imagine a world with only ipv6 and without NAT.. the dream

[ripe anchor]

that would be amazing specially since my isp decided to force implement cgnat

[woven loom]

That would be game changing, literally even

[potent sparrow]

You can have that world any day now* (* any day now is not contractually binding and may be 50 years from now - if depression over the fact that we are still using 32 bit unsigned numbers as the global defacto routing system strikes - consult you doctors, side effects of IP6 may include happiness, not having to deal with CGNAT ever again)

[restive turtle]

dedmen doing gods work out here

haha

[opal hound]

Getting a lot of

22:14:17 Sound: Error: File:  not found !!!
22:14:17 Sound: Error: File:  not found !!!
22:14:17 Sound: Error: Multi-stream: Load failed, stream  not found !!!
22:14:17 Sound: Error: Multi-stream: Load failed, stream  not found !!!
22:14:17 Sound: Error: Multi-stream: Load failed, stream  not found !!!
22:14:17 Sound: Error: Multi-stream: Load failed, stream  not found !!!
22:14:17 Sound: Error: File:  not found !!!
22:14:17 Sound: Error: Multi-stream: Load failed, stream  not found !!!
22:14:17 Sound: Error: File:  not found !!!
22:14:17 Sound: Error: Multi-stream: Load failed, stream  not found !!!

is this a profiling thing or a mod acting up? I haven't seen it before

[void badger]

Probably a bad mod

[opal hound]

I can't grep it out of my mods folder, and the mod set hasn't changed for a month or so

[silk pewter]

https://tenor.com/view/sinbad-dreamworks-give-that-guy-a-raise-wet-gif-16580209

[whole cloud]

There were no changes to file or sound

[scarlet jolt]

another staurday operation, immediately server CPU started periodically spiking to 100% and server fps would drop to 0 for a few seconds. Turning off all the other services we have running except for Teamspeak minimized it but I dont know if it resolved it

ope this might not be relevant information anymore since I think we are on the V28 server

Still gonna try a slow frame capture and see what that reveals

[pale carbon]

Was this branch updated recently? today i started getting terrible stutters when in gunfights, making the game unplayable. Disabling all mods and DLC didnt fix anything either. When switching to the main branch the stuttering is fixed.

[opal hound]

The last update is in the pins, 5 days ago

[restive pilot]

in multiplayer?

[elder path]

Since the last update to profiling, my dedicated server has had serious desync issues with no obvious cause. This does not seem to affect every player, either. Is anyone else experiencing this?

[scarlet jolt]

Is there any guidance anywhere how to read one of these trace files?

I've got it open in perfetto and I am realizing I have got no clue what I'm actually looking at

[violet elbow]

Changes to networking were made, might wanna double-check your basic.cfg?

[elder path]

Where are these changes detailed?

[crystal haven]

Just a little teaser of an in-game admin panel using ctrlWebBrowser for a framework I'm working on

And another teaser of an in-game garage ui using ctrlWebBrowser for the framework I'm working on (still a wip)

[heavy galleon]

Wouldn't the worst case scenario with "old" basic.cfg be the same result?

[pale carbon]

singleplayer

[scarlet jolt]

if your arma server is on the perf profiling branch but you aren't running the profiling.exe, do you still just get the perf enhancements or are you just running the base arma executable?

[woven loom]

The former

[scarlet jolt]

Ahhh... interesting. Thanks for the info

[whole cloud]

Check the affected player in the player list on the map screen.
Do they have low bandwidth? Desync? High ping?

[elder path]

It does show up as desync in some cases, yes. I mentioned more info from further troubleshooting in #server_windows

It seems to be related to the newest version of the profiling branch primarily

I have reverted the server to the previous one and made changes to my basic.cfg to bring it in line with what others have suggested here after a lot of searching through old threads

I'm going to keep poking at the issue and see if it continues or is fixed- Unfortunately I can't do much testing because it's such a random issue and doesn't affect everyone, meaning I need a larger group of players- Not easy on short notice

[whole cloud]

I think the desync number on map screen. Would mean the player has low bandwidth, such that the server isn't even managing to send out the messages that would be needed

[elder path]

That makes sense, yes

[whole cloud]

Keep an eye on the bandwidth numbers.
If they really are low, it really can only be bad ping or packet loss

[elder path]

I'm hoping the adjustments I made to my basic.cfg settings help mitigate that. I have very little knowledge of the network side (I'm a hardware guy primarily) so this is pretty arcane to me

Also, will do

[scarlet jolt]

How's the V29 server for large scale stuff now, is it stable enough to put my group through 4 hours of arma on it now?

[heavy galleon]

Try it, let me know, thx

[inland dew]

still on v5, because of smoke and other stuff...

[heavy vortex]

Got a six-second bout of network message pending with DS + local client. Not sure if working as intended.

[naive osprey]

I've had v29 on a server ranging from 90 to 160 players since it came out, only notable issue is that packet loss and desync are much more noticeable than before, though the frequency of said desync has dropped significantly.

[whole cloud]

probably because your socket maxBandwidth is limiting it

[heavy vortex]

There's no basic.cfg, for what that's worth.

[whole cloud]

yeah default is 2MiB, might be tight for a local client

[deft oak]

Any recommendations on what to bump it to now on v29? Or do the recommendations for basic.cfg from v27 still hold true for the v29 changes.

[whole cloud]

It will only use as much as it needs and has

so can't really be too high

The problem with local clients is that application layer ignores bandwidth, but transport layer still obeys it. That causes the pending messages missmatch

if you raise transport layer high enough that it can handle all that should go through (which local client will be able to) then that problem won't happen

[crystal haven]

When using the ctrlWebBrowser I'm getting multiple lines of the following in the rpt while the UI is open:

12:08:40 UpdateTexture 7027
12:08:40 UpdateTexture 7028
12:08:40 UpdateTexture 7031
12:08:40 UpdateTexture 7037
...

Is this due to using it as a display an not a dialog?

[whole cloud]

That sounds like I forgot some logging

[crystal haven]

What are some more ideas/plans that you have for ctrlWebBrowser?

[whole cloud]

None at the moment

[silk pewter]

can it emulate IE6?

[fast hornet]

https://tenor.com/view/begone-demon-exorcism-garlic-begone-demons-gif-15627834

[heavy galleon]

Is there a chance we would ever get arm64 server binary?
You know, with cloud providers offering more and more arm64 servers.

Guess it wouldn't be as simple as switching the compiler (but perf prof would be awesome for testing it )

[woven loom]

wouldn't raw performance be much worse anyway?

[restive pilot]

0 chance because many things in Arma are written for x86 x64 specifically

[heavy galleon]

ay caramba

[grizzled summit]

Pray to Eden that Arma 4 does 🙏

[heavy galleon]

Don't think so apple silicone showed that arm is competitive. And companies are investing heavily into it now too.

[woven loom]

yeah, it's workload dependent though from what i've seen, and depends on what you're after (efficiency vs raw performance)

[silk pewter]

you can wait for Dedmen's absolute final confirmation on the topic, but I would not hold my breath: as Leopard20 said many things are written for a different architecture and ARM64 support is not a priority at all

[woven loom]

i guess if ARM servers are much cheaper to rent than x86 and performance is similar for the game server, then it has a lot of advantages?

i don't think i've seen game server specific benchmarks though

[heavy galleon]

Yeah, will wait a week
Priority surely not, or even on a list anywhere, but would be cool to add it to the list!

[grizzled summit]

I scrolled up and found nothing, is profiling fixed with the issue it was having with missing ID's?

[heavy vortex]

What issue with missing IDs?

[silk pewter]

please don't hold your breath for a week

[grizzled summit]

Sorry - Didn't see your response, let me grab the error code

bin\config.bin/RscDisplayMain/controls/TitleSingleplayer.textureNoShortcut

The little black box buttons dissappear

On the main menu and in servers, meaning you can't join a multiplayer server

I believe it ONLY happens when loading a mod with missing cfg's or something? I just moved to stable branch, and was already 2 hours late to an operation so I didn't dig much deeper

[storm crag]

The debug corner is so fun

[median belfry]

I learned yesterday how to use the web browser and was thinking to make some UI with it, it looks pretty good, gonna try to revamp all the ui now, thank you for that Dedmen the only thing was I notice is that the fading in effect sutters some times, probably because of arma? idk

[potent sparrow]

arm servers can have comparable performance to x86 but it is somewhat nuanced when performance matters the software often has to be tailored to the hardware and even then x86 crushes arm for raw perf (perf per watt is more contentious) - proper servers running EPYC's are hard to beat

[fast hornet]

arm servers also typically have lots and lots of (slow-ish) cores, so not something arma greatly benefits from

[restive turtle]

i run an EPYC rn

its pretty good

[potent sparrow]

yep, game servers are an exception in that they like higher freqs/IPC/X3D (mostly because the engines where written with those processors in desktop use as the target)

they like lots of RAM as well (but lots in the desktop sense, so 64GB) not the server sense (where 256GB is kinda standard these days and 512/768GB isn't exceptional)

last place I worked if you counted the RAM in the DB cluster is was ~1.5TB (and then there was the slave clusters..) - kinda blows my mind since my first computer had 16KB of RAM 😈

[fast hornet]

2mb for me 😉
But yeah, 1TB and more memory is pretty normal now, especially for hypervisor systems. 256/512GB has been normal for so long that companies are throwing away systems with that amount.

[smoky ridge]

does anyone know what exactly a GIF pre stack size violation error is?

[quaint flame]

Nvm

[shell lion]

@whole cloud
I send you a pm, of some crashes from users or server over the weekend with perf 2.18.152803/2.18.152842

[potent sparrow]

Still waiting for the 48GB DDR5 to come down so I can upgrade from 64 to 96 - meanwhile servers are into the TB's - 10 year old me wouldn't believe the kind of computing power I have access to these days at home

[fast hornet]

my past self would not have believed i'd casually have 16c/32t at home. I'm still at only 32GB though, bigger modules just were sooooo slow last time i looked

oh yeah, also sent you a crash memdump Dedmen 🙂

[potent sparrow]

I'm on AM5 so no real point going over DDR5-6400 so I just went with 2x32GB of that - wanted the 48's but I was already way over the line on the "missus acceptance factor"

[fast hornet]

also on AM5 .. but i'm running DDR5-6000 at 28-35-35-59. Won the chip lottery a bit by getting modules with really good hynix chips.

but looks like you can even get 64gb and 96gb cl26 kits by now.. 2 years of ddr5 production doing wonders 😄

[potent sparrow]

aye, I built this when the 7950X3D just came out - world moves on fast still 🙈

[fast hornet]

haha.. also on 7950X3D. Basically had the whole system ready for 1-2 weeks already, just waiting for the CPU delivery

[elder cradle]

This seems cool, like tarkov hideout. What is this contraption you’re creating?

[median belfry]

its from my survival mod a3s, for the last year I'm working on a huge update, its gonna be crazy when its done
https://steamcommunity.com/sharedfiles/filedetails/?id=1278609853

[inland dew]

please -> #hardware_vs_arma or #offtopic_hardware

[potent sparrow]

small world, was looking at that the other day 😄

[heavy vortex]

Means you confused the SQF parser enough to break it. Usually a missing bracket somewhere.

[restive pilot]

GIF?

[potent sparrow]

suspect he meant GIAR

[heavy vortex]

there's GIAS as well I think?

I don't know what the letters mean. There could be 20 different ones for all I know.

[silk pewter]

Gröss Inkompetence About Skripting!!

[potent sparrow]

De' mughchu'be'bogh DayajlaHbe'chugh, bIlughbe' 'oHbe' QaghwIj'e'.

[smoky ridge]

nope, GIF

Also apologies, i thought i had put it in #arma3_scripting

[heavy galleon]

I'm starting to turn slightly purple 🟣

[quartz rampart]

stop drawing on your self in class while you are bored.... that does not count

[whole cloud]

Has anyone noticed a speed difference with non-http mission downloads since v29?
With the better netcode, the server should be able to send alot more data faster

[patent sky]

Ruining my hard work

[whole cloud]

oopsie, that should've been a #if
Thank you

There are actually quite a few more of these, like ExecuteJavascript which actually logs the scripts you throw in

[whole cloud]

Mac already has a arm binary 🤣
But its not worth the time, so no.

[whole cloud]

Its a broken mod, something with camofaces I think that breaks all the UI configs

[heavy galleon]

Worth it for me!
(Literally in the process of unpacking a Pi)

[light cargo]

pi running arma 3? screams oh no

[median belfry]

Dedmen if I use SendConfirm via js how is the best way to send the string back ? since JSDialog only uses boolean

[light cargo]

there are better arm systems and how about risc? :)

[whole cloud]

There should be no stutter. Like youtube videos and twitch streams ran just fine.
But there could be some lag when the browser starts up.

Is that interaction menu, a new browser instance spawned every time? that's pretty fast

[median belfry]

I ended up just using SendAlert -> JSDialog -> ExecJS back with the data, but I thought Sendconfirm would get back the string automatically with JSDialog

[whole cloud]

#arma3_scripting message

[whole cloud]

no crossposting 😠 I'll get to your message sometime

[median belfry]

Yeah, it's a new browser instance every time. But is there any other way to do it? The browser instances close as soon as the dialog closes. The only alternative I can think of is using a hud, but then you can't interact with it.

[median belfry]

my bad

[whole cloud]

I don't know if you can maybe keep a dialog open, but hide/disable it so it doesn't capture click.
Probably not

[restive pilot]

make it possible to transfer the ctrl
ctrl ctrlSetParent newDisp
that way you can move it to display 0 or 46 and disable it without closing it (and bring it to your own dialog later)

[heavy galleon]

For a "test server" where you pretty much only check mods. Great

[whole cloud]

Mh yeah I think that would work.
But for 2.22 if someone remembers

[restive pilot]

well if you can allow displays to be deactivated that's also good too
whichever's easier

[whole cloud]

moving a control seems easier

[median belfry]

can steam influence the webbrowser ? because for like 5minutes everything was lagging, delayed and transitions didnt feel smooth at all. even an restart of arma didnt help

https://streamable.com/znkibk

[light cargo]

checked steamwebhelper cpu or similar? iirc this inconsistency was partially talked about and the feature is running under the webhelper instead of its own cef

[median belfry]

cpu mem gpu everything looked fine. running 32gb ram 7800x3d rtx 3080

[magic elm]

Significant improvement. At least 50% faster if not more

We’re also getting a loss less infinite loading screens or people with higher pings failing to connect

[charred pagoda]

I noticed that downloading speed on non-http is limited only by socket max speed now

[patent sky]

Sidenote, you might wanna replace those item icons/images with your own 😄

Don't want Nikita knocking on your door

[heavy vortex]

My connection had a bad packet loss day yesterday (DSL training) and I was getting kicked off a server for signature verification failures literally every minute, despite everything else working ok. While you're messing with the netcode, you might want to make that code a bit more... fault tolerant.

[whole cloud]

Steam IS the webbrowser

[whole cloud]

Increasing socket speed might be an interesting experiment then

[whole cloud]

Maybe next month 🫂

[whole cloud]

2.18.152860 152863 152866 new PROFILING branch with PERFORMANCE binaries, v30, server and client, windows 64-bit, linux server 64-bit
- Tweaked: Improved game start speed
- Fixed: Multithreading related crash (Thanks @chilly geyser )
- Fixed: Crash while ADS and switching to spectrum device while remote controlling a unit (Thanks @uneven bluff )
- Fixed: Animation hitching (Thanks @wise sparrow )

If you don't want to use the Steam branch, the files are also available for alternative download here:
https://drive.google.com/drive/folders/15p9j7C2nHUt6NoVfChX4YFuqzFXzblJh
Note: There are separate Dll files that also need to be placed into Game folder.

[humble zenith]

Are there different server config tweaks for perf now? What are the optimal settings?

[opal hound]

Highly dependent on your server specs but the general recommendation is here #perf_prof_branch message

[vale shoal]

Hm, is anybody seeing some unusual server behaviour with the latest prof? The complete machine hang itself away suddenly. The only thing I saw on the Arma 3 instance (Server) was "Parsing configs", then nothing. No response at all to Windows. Had to reset the machine. It could be coincidence but I don't know honestly.

Additional to that, the .rpt was basically empty after loosing connection to RDP. This are the last lines:

Allocator: D:\Servers\arma3\altis\Dll\mimalloc_v219_20250213.dll [] [] PhysMem: 64 GiB, VirtMem : 131072 GiB, AvailPhys : 58 GiB, AvailVirt : 131068 GiB, AvailPage : 67 GiB, PageSize : 4.0 KiB/2.0 MiB/HasLockMemory, CPUCount : 14 =====================================================================

[fast hornet]

have v30 running on our testbox for ~5 hours now, no crashes or other problems so far.. but apart from me connecting once as a test it usually has no clients.

[vale shoal]

Was there changed something in the config parsing process by any chance? If yes, if it helps, I use obfuSQF. Maybe there is a conflict somewhere.

[analog acorn]

#perf_prof_branch message 🤔

[potent sparrow]

in the context of that post what is sockets.maxBandWidth referring to, I'm familiar with the other settings in cfg but that isn't one I remember seeing

ahh it's its own config class (found it on github in someone elses file)

[weak panther]

1. fly any plane e.g. wipeout
2. during fly, eject
3. while you fly in seat exec player setPos [5298.72,5613.71,-0.000198364]
4. my UI stuck, cannot press buttons, cannot close pause menu via Esc key

given coordinates for example, it seems that any execution of setPos will give such behavior. after move out from seat setPos work ok

[silk pewter]

some disableUserInput at work here? 😓

[fading shuttle]

is there any good way to time how long it takes to initialize mods to compare to before?

[whole cloud]

There was a change with pbo loading, but only with profiling binary.
And I don't see how that would fail 🤔

[whole cloud]

Is that new?
Why would anyone teleport themselves while ejecting

[whole cloud]

a stopwatch 😄

From launch button press, to main menu

[fading shuttle]

do rpt timestamps work bc it showed there i got 25 seconds without profiling and 28 with 😭

loading about 18 gigs of mods

[analog acorn]

The file size of the mods isn't really the problem. One mod with terrains, high-detail models, high-res textures, and lots of sounds could add a lot of file size, but none of that will be actually loaded until it's used in the game. The mod loading time is affected by the number of mods, and the size of the configs within those mods.

[fading shuttle]

operation trebuchet and first contact

it's a lot of configs and pbos

at least 364 config.cpp files

[whole cloud]

It also depends on how often you start the game.
On first start (when mods or game binary changed) stuff is cached, second start only loads the cache.
Creating the cache is now much faster too though.

But second start is still always faster than first start

[fading shuttle]

ok ill go try that

yeah

after two runs it went down to 18 seconds total for the rpt

[whole cloud]

if its obfuscated, and you can narrow it down to it (like load vanilla, but add that obfuscated pbo). Then you can send me that pbo and I can check why

[vale shoal]

The client seems to work fine with obfuscated pbos, I joined on server with v29 without problems. Just played for like 5 min but nothing happened.

[steady cairn]

performance improved since new update?

[heavy galleon]

Please word it differently.
The loading time is affected by the number of pbos
You then have people who will go around and confidently say that they mega repack because "one mod loads faster than 70 "

[ember owl]

Just had this issue a few times over the course of the past two days. I used to not have this so figured I'd echo it. I play in windowed fullscreen.

[restive pilot]

are you guys sure it's not a driver issue? (specifically nvidia ones which are terrible recently)

[foggy vine]

Do you use Discord overlay?

[ember owl]

Yeah that is probably it- they also just had an update recently for it as well.

Cheers

[red thorn]

someone of my clan had the same issue as client
joining the server resulted in complete PC freez
had him revert to v29 with no problems joining

I myself had no problems with v30

[whole cloud]

Like worst case, if there really is a problem with obfuscated pbo's. The worst that could happen is 100% CPU load and the game freezing.
And that is at game start during the splash screen, not mid game.

[red thorn]

I don't thinks its related to obfuscation of pbo's since the mission was not obfuscated and the all the mods (at least the first pbo of each can't check all) wasn't either
I checked by opening the pbo's with pbo manager -> error shows when pbo is obfuscated

[azure roost]

I’ve got an issue where my PC seizes up on profiling at the splash screen. Only way out is to power button out of it.
It’s only occurring on profiling.
I also tried turning off discord overlay.
Even with no mods it does the same.
Tried validating files, even full reinstall.

Only thing I may need to check is drivers as I’m fairly certain they’re up to date but not 100%. Only started happening with this prof update and it’s only happening to me of the bunch from my group

[uneven kernel]

I had that a wee while ago, funnily enough spamming escape fixed it. It went away eventually, never found anything in the .rpt or what the cause might've been

[whole cloud]

Well we increase the main threads priority now, during start up it runs as a high priority.
But windows shouldn't let any process kill/freeze the system, even if they are a high priority thread that runs at 100%.
But maybe we're triggering a windows bug then?

Is that not similar with what happened back then with the -setThreadCharacteristics on old windows server versions? Was that not the same, some system freezes? Don't remember

@azure roost @vale shoal @red thorn which Windows version (as precise as possible) on the affected machine?

[vale shoal]

Windows Server 2022 (21H2, Build 20348.3453)

[fast hornet]

what hardware and/or hypervisor is running it? No problems here with two arma v30 servers on different systems, also on Win2022 10.0.20348.3453

[whole cloud]

I don't know what would peg cpu load too 100%.
When I start a local server, it sits at 0% usage after initial loading.

Maybe it would run if you manually lower process priority? And then could analyze with vTune to see what its doing?

[red thorn]

ok got a correction: the freez didn't happen at server join but at game start like the others
he has Windows 11 Home (24H2, Build 26100.3915)

[whole cloud]

If I complain about it to microsoft, Raymond Chen will make a blogpost about something I'm doing wrong and that its actually my fault 😄

[silk pewter]

then learn from it and fix it! 😆

isn't CPU 100% load a good thing? people complain when the game uses 10%, when it uses 100%, IDK how to please them :p

[whole cloud]

⚠️ We replaced prof v30 with 152863
Removed: increased thread priority during startup
Fixed: Crash in ctrlSetModel (Thanks @subtle vapor )

[knotty wraith]

@subtle vapor - the experimenter https://www.youtube.com/@andreigud

[knotty wraith]

I kindly ask you to add a description of the innovations to Google because the site is not working

[silk pewter]

aren't they pinned?

[knotty wraith]

my apologies, i thought only the last message was pinned

[fast hornet]

is that xkcd?

[potent sparrow]

Yeah, 1172

[foggy vine]

Swapped to profiling branch literally less than an hour ago, kept alt-tabbing around from Arma to stuff like Windows Explorer and VSCode while I was loading the VR terrain...

... and Arma crashed without a memory dump.

RPT

[whole cloud]

fullscreen-windowed mode?

[foggy vine]

Yup

[whole cloud]

Mh. Well maybe a one time fluke 😄

[foggy vine]

Hopefully :P I'll swap to windowed and back just to be sure

Yeah one time fluke

[heavy vortex]

It usually hangs on a black screen rather than crashing, right. It's not expected to survive unlimited alt-tabbing.

[whole cloud]

yes, it tries to open a message box from rendering code but fails and gets stuck

[heavy vortex]

IME it usually survives unless I get bored and alt-tab during a loading screen.

[foggy vine]

It usually survives my endless alt-tabbing even on a black screen, but that's good to know

[heavy vortex]

ACE death screen doesn't count :P

[whole cloud]

I found a new issue with our friend linked-lists again.

Game freezes (sometimes 10+ seconds) when opening server browser.
The game is preloading all the signature files, but that itself is not what is slow.

The game doesn't want to keep opening and closing files, so it keeps a cache of open files.
But for that cache not to grow too far, it only caches 64 files (besides files that are locked open, like pbo's)

So when a new file is to be opened, but the cache is too full. We need to throw out one file.
So far so good.

Our file system has multiple levels of caches.
The next level is a request cache. Like reading 4096 bytes of file X at offset Y.
That is cached, so when the same file section is read again, it can pull that from cache instead of having to re-read it from disk.

That request cache is stored in two ways. One is a hashmap and another is a linked list.
The hashmap to quickly find requests to see if a file is already cached.
The linked list is used to store the order of how "recently" a file was accessed. Such that when the game is running out of memory, it can purge the oldest and thus most likely not needed anymore files first.

Now when we want to close a file, we first need to remove all references to that file from the request cache.

To do that, for every file we want to close, we iterate that linked list containing all cached requests and find the ones referring to the same file, and delete them.

Now, iterating a linked list of all file chunks that have been read before, quickly gets into the 100k items range, more with more mods.

If we want to close one file, we iterate 100k items in a linked list.

The signature loading code, loads all signatures, one or multiple per pbo. Even vanilla that's a thousand files.
A thousand times, iterating through 100k+ items in a linked list..

Brute force fix, we have the table which is a bunch of arrays, so I can multithread that. (10 second freeze, down to 2 seconds)
Better fix, have another hashmap that is indexed by just the file so we can quickly find all open requests belonging to a file.

// opening file can be very slow on XBox
// it is therefore a good idea to keep cache of recently opened files in case they are used again
Thank you xbox

[heavy vortex]

Arma 3 doesn't actually support xbox at the moment, does it?

[light cargo]

it's from a earlier arma that was planned to be on xbox most likely

[whole cloud]

Neither does it read files from DVD (which is also a big thing in this code)

[light cargo]

why is that there?

[heavy vortex]

Feels like on any proper OS you could just let filesystem cache deal with that and comment out the whole thing.

[inland dew]

probably comes from OFP Elite on Xbox (2005)...

[light cargo]

ya there we go

[whole cloud]

Actually instead of another hashmap, I can store it directly on the file that is opened (it already stores a count of how many requests are open on it, just not links to the requests themselves).
That quickly hacked in, 10second freeze down to 0.4s

good enough 😄
Before, like 90% iterating a linked list, and a tiny bit spent actually closing the file.
To no linked list at all. A bit freeing memory of closed requests, closing the file, and the mutex now makes an appearance because its so fast.

That should also affect performance when using filePatching.
It basically affects all file closing, and filePatching opens and closes alot of files, potentially

[opal hound]

ohhh that's what's going on with my server browser
I thought it was just that I had too many non-responsive servers in history

[whole cloud]

It gets worse the more files are in cache. And as I increased the maximum file cache size because people wanted more ram utilization... 😄

[celest sparrow]

this is the way, give me some keep busy loops and stuff so it looks like its all working hard

[potent sparrow]

digging the HAL9000 profile image 😄

I used that as the status indicator in some employee performance tracking once (went green if scanned in job was under time, red if over time - no one got the reference, I was sad)

[celest sparrow]

im def not a bot, I am real human boi

feel safe an comforted

[vivid rune]

Could this be also the reason for that YAAB has the tendency to be a little bit slower in the second run?

[azure roost]

Windows 22H2 OS Build 19045.5737 for my PC, playing not hosting.

[runic sigil]

Found someone with what appeared to be this thread priority on startup bug
Message link is here, don't know if the rpts are useful anymore
#arma3_troubleshooting message

[patent sky]

yeah was looking at that one, its the v30.1 prof build so that should no longer have high thread prio, perhaps something else is causing the issue

[whole cloud]

Don't think so, yaab is generally faster for me in second run

Looks to be the multithreaded animation cache initialization.
But only failing for a handful of people is weird

[whole cloud]

@vale shoal are you doing some stuff like controlling cpu affinity, -cpuCount parameter, how many cores the game can use?

[vale shoal]

only in the bat I do start /high

I removed -cpuCount and -enableHT because you said it would give better performance

[whole cloud]

I can reproduce the freeze, but only if I tell the game to use more cpu cores than I actually have 🤔

[vale shoal]

start /high /wait D:\Servers\arma3\altis\arma3server_x64.exe -mod=@Exile;@RwG;@CBA;@EM;@EMR;@CUPW;@CUPU;@CUPV;@RHSUSAF;@RHSAFRF;@RHSGREF;@RHSSAF;@BWM;@USAFM;@USAFF;@USAFC130;@3CBE;@3CBV;@3CBVRETEX;@3CBVS;@AVMG; -servermod=@obfusqf;@ExileServer;@a3_infiSTAR_Exile;@a3_infiSTAR_vision;@infiSTAR_TM;@a3_infiSTAR_Exile_customApps;@SM_LGC;@ESM; -connect=127.0.0.1 -port=2302 -config=@ExileServer\config.cfg -cfg=@ExileServer\network.cfg -name=AVMG -profiles=D:\Servers\arma3\altis\BEC -autoinit -hugePages -bandwidthAlg=2 -limitFPS=120 -loadMissionToMemory -pid=altis.pid -malloc=mimalloc_v219_20250213

That's the full command

[whole cloud]

So guy in #arma3_troubleshooting had the freeze because he had 6 cpu cores, but told Arma it had 12 to use.
I can reproduce it when I tell Arma I have 16 but set affinity so it can only use 8.

And its not frozen, it just takes like.. atleast 10x maybe 100x as long to load the game, while also spinning the CPU at constant 100% during that

[whole cloud]

I see cpuCount 14.
I spotted an issue in the code if cpuCount is larger than 16. Then the remaining threads would just spin at 100% load doing nothing.
But if you're only at 14 then that wouldn't be it

Jup I found it.

When we do a bunch of work in parallel (like calculating thousands of animation actions)
We first spin up all the worker threads, because they take a while to come out of their sleep. Even though they don't have any work submitted for them yet, they start hot spinning waiting for us to submit the work.
Then we set up the tasks, and submit them. Because the threads are hot they'll grab them immediately.

Then we wait for the tasks to be done, which most likely also causes main thread to grab one of them.
Then we disable the flag telling the threads to stay hot, so they can go back to sleep.
And we are done with the parallel work.

Problem with that, if there are too many threads and they cannot work.
Threads may grab work to be done, but because the operating system doesn't have enough cores, it will suspend the thread that was doing real work, and give cpu time to another thread.
That other thread won't find any jobs to do, because other threads grabbed them even though they're not working on them. And it see's the hot flag and keeps spinning at 100% cpu load doing nothing. Until the OS time's out, force suspends the thread.
And then we have a basically random chance on what thread gets woken up next.
With luck, its one of the ones that was actually doing work, which will finish its task, and then also go into a hot spin, waiting for more work which isn't going to come.

So we are playing random chance ping-pong, hoping to hit the threads that had actual work, until all threads that had work finished it.

And the way this work specifically is designed, is extra bad.
Each job processes 16 actions, and they have a config lookup which has a mutex at the top.
The mutex is likely to call to the OS, which the OS likes to use as a chance to suspend the threads.

So the threads that do nothing, just hot spin at 100%, making the OS think they are doing important work, and also not giving the OS any opportunities to suspend them, until the timeout expires.
The threads that are actually doing work, are constantly spamming the OS with opportunities to suspend them, so are much more likely to be suspended, letting the OS use the cpu core for one of the useless threads.

The fix, swap two lines of code. Remove the hot flag, before we wait for the work to finish.
That will cause the threads that can't find any work, to go back to sleep.

So simple This multithreading stuff is very sensitive stuff

I assume this happens on the servers, because there are just other processes running that are hogging some cores. Even if the game can't use just one of its threads, this would send all the others into a hot spin, until that one thread gets its go at a cpu core

[cloud sky]

But hey, you made the "why's da game not using more CPU" peeps happy!

[whole cloud]

[drifting karma]

[whole cloud]

⚠️ We replaced prof v30 with 152866
Fixed: Extreme slowdown (might look like a freeze) when the game has less cpu cores than it thinks it should have

[quartz rampart]

just mine some cryptop while they wait for work 🤣 that way BI could refinance your time 😛

[charred pagoda]

Is it related only for servers/clients that started with specified -cpuCount option?

[whole cloud]

Also servers without cpuCount, if something else was blocking cpu cores for some reason

[weak radish]

@whole cloud How does the MT deal with the 12th, 13th and 14th gen P/E core architecture.

If one has a i7-13700 for example, should one use 8 for the 8 physical P cores, 16 for the 8+8 total physical cores or 24 for the total logical cores 8*2+8?
Can the game differentiate between the P and E cores when it automatically chooses an amount?

If the multi-threading is waiting for all the threads to complete then having jobs running on the E cores might be detrimental in some cases

[fast hornet]

you should restrict the server CPU affinity to the p-cores/threads and use the appropriate -cpuCount. So 16 for p-cores with 16 threads.

just letting it run on all cores is even worse performance-wise than using the e-cores only.

[knotty wraith]

Windows 11 will solve the problem with the correct load on all cores - personally tested, Windows 10 does not cope with this

[whole cloud]

it doesn't. Your operating system hopefully deals with it

[fast hornet]

Server 2022 also does not have the newer scheduler.. 2025 might, didn't test it yet. Might also be different between client (sets threadcharacteristics as "game") and server (doesn't set threadcharacteristics)

[weak radish]

So task manager set affinity it is then. Thanks 🙏

[fast hornet]

nah, you can just set it with powershell on launch

[weak radish]

Client via steam

[fast hornet]

ah ok.. would also work though .. you can just task schedule a script that runs every few minutes or something

Like that for example. The affinity number depends on your cpu though, might not be the correct value for yours.

Get-Process -name "arma3_x64" | ForEach-Object {
    $_.ProcessorAffinity = 65535
    $_.PriorityClass='AboveNormal'
}

[potent sparrow]

for anyone on linux you can just do (say it's a 13500 which is a 6P core) taskset -c 0-11 <arma3commandasusual

[cloud sky]

Or just disable the E-cores via UEFI with e.g. KVM Console

[potent sparrow]

disabling the e-cores is kinda silly when they are ideal for running stuff like teamspeak/voip solutions - linux has mechanisms built in to just say "you run there, and you run over there"

[fast hornet]

the powershell value is binary, read from the right.. so 0011 would be core/thread 0 and 1 enabled, 2 and 3 disabled.
But for some reason the command doesn't accept the binary value, you have to provide it as decimal or hex. Don't ask me why, microsoft. 🤷

[cloud sky]

Dunno, maybe it's optimization overkill but I try to keep all the CPU caches in Arma server's use as much as possible 😅

[potent sparrow]

right...and saying "arma you run on the p-cores, everything else that's using CPU that matters you run on the E-cores" does that

the "just disable E-cores" thing needs to go away - it's silly

[cloud sky]

Do they have separate L3 caches?

[potent sparrow]

@fast hornet just slap 0b on the start - with 0b you can use binary 0b1111111111111111 (called a binary literal)

[fast hornet]

not sure powershell accepts that, but i'll try. Thanks 🙂

[potent sparrow]

it does, from v7 - https://learn.microsoft.com/en-us/powershell/module/microsoft.powershell.core/about/about_numeric_literals?view=powershell-7.5

[fast hornet]

ah, not an option then.. i try to keep everything ps1 compatible so that it runs on every system.. not only those that have p7 installed

[potent sparrow]

fair, I've used zsh for well over a decade, I still write scripts in pure bash format because I don't want to have to worry if zsh is installed on the other end 🙂

[fast hornet]

but just providing the binary and then doing the conversion in the script would be possible

[gritty wasp]

can arma detect HT threads? If so, maybe startup parameter should not override detected value if it is unreasonable. max(detected, userSpecified)

[warm sedge]

man i'm so glad, i don't know what you guys are Talking about 🙂

[patent sky]

startup parameter is exactly for when autodetect isn't working/correct (or when u wanna run with less cores). so relying on the detected amount to determine something is "unreasonable" doesn't make much sense

[knotty wraith]

@whole cloud
windows 11
Why does it work like this, the game just doesn't start (very rarely), with vpn it always starts
https://youtu.be/uaf8mTaZBPY

[whole cloud]

Well it looks to be BattlEye, I don't know BattlEye

[patent sky]

maybe try a different DNS (1.1.1.1 or 8.8.8.8)

[whole cloud]

or 9.9.9.9 😄

[knotty wraith]

thanks - it works)

[azure roost]

Yeah can confirm mine had 8 cores set in parameters when I only have 6 🤦
Working fine for me now though, thank you for looking into

[whole cloud]

Still the game will perform worse if the count is not correct

Instead of big lag's, you'll just have small ones

[gritty wasp]

if it can not detect correctly and you ask more you will suffer. What is the point to allow that?

[patent sky]

doesn't detect correctly != can't use all cores

[slender lagoon]

should we include the number of logical cores in cpucount or only the physical ones?

[whole cloud]

cpuCount is total threads

[gritty wasp]

DO NOT TOUCH IT. Let it autodetect

[fast hornet]

For scenarious like the intel CPUs or running multiple servers on one system. E.g. when you set affinity to the P-Cores only on a 13900, Arma can use 16 threads. But it will detect the total count which is 32 on that CPU. So you restrict it to 16 with -cpuCount=16

[whole cloud]

In my tests, including hyper threads performed worse in YAAB

[gritty wasp]

having HT enabled almost always worse in almost all software. Very limited scenarios have benefits from it

[gritty wasp]

My max example was wrong. I'm talking about reasonable values. 16 out of 32 detected - reasonable. 16 out of 8 detected - hmmm.

[whole cloud]

Users could have valid cases for intentionally wanting to harm their game. Like I had today to test this bug 😄

[slender lagoon]

I have HT enabled for years, it helped in the past

[gritty wasp]

In Pentium 4 HT

[whole cloud]

Closing a file:
Before: Even if there are 0 references in cache, still iterate through 150 thousand cache entries to see if there are any
After: We have an array of cache references, usually just a handful, iterate over array and delete them. Done. Bye!

Coming in v31

[silk pewter]

FPS + 1000 ?

[gritty wasp]

in server browser

[indigo anvil]

More like: previously dormant code triggered resulting in crash

[vale shoal]

Works flawless, thanks for fixing it!

[woven loom]

would still suggest seeing if Intel is willing to test and add Arma 3 to APO whitelist https://www.intel.com/content/www/us/en/support/articles/000095419/processors.html

Also @weak radish you may want to try APO in advanced mode in the meanwhile https://www.intel.com/content/www/us/en/support/articles/000098240/processors.html

Support is much wider now

Supported Intel® Core™ Desktop and Mobile Processors with at least 6 p-cores or more.

[whole cloud]

Should've gone with AMD

[weak radish]

Sadly my OEM is a bit of a pain, disabled overclocking on i7-13700HX’s etc so I’m not sure if it’ll work. I’ll get it all set up tomorrow and see what works best for me

[dense light]

getOrDefaultCall crashes the game when called with a nil value for the code argument on profiling and dev. Not on stable.

private _map = createHashMap;
// _map getOrDefaultCall ['thing']; // no crash
_map getOrDefaultCall ['thing',nil];

[weak panther]

Wanted report no particles in game again but noticed that:
Game in pause menu says game ver is 2.18.152866
But launcher says Game version: 2.18.152405, Branch: profiling
And no particles on mission

[patent sky]

game and launcher version mismatch on profiling is normal, we dont build a new launcher for every prof build since there is no need

[weak panther]

On mission at one point particles gone, but after respawning they appeared again, I think it happened twice. I couldn't find a 100% way to get it

[turbid vortex]

I can say I have encountered issue with some particles missing on stable, most notorious was smoke grenade, with smoke sometimes not being shown for some people. Maybe issue isn't exclusive to profiling, or maybe it got worse.

[heavy galleon]

Also had issues with particles yesterday, as I didn't see the impacts of my bullets when shooting si 1.1km. some shots I saw, some not. Almost full random

[weak panther]

build 152866, just joined on koth - no particles (bullets, smoke, explosions...)

[heady echo]

having an issue where I can't join a server when on profiling, deleting mission files and restarting game etc still 50/50 chance of getting in. tested on normal version and have no issue.

[heavy galleon]

Has anyone had any crashes on v30, especially on dedicated servers?

[ember owl]

I've had two segfaults but I was also running JSRS at the time which is known to cause crashes so would take my experiences with a grain of salt.

[heavy galleon]

The new one, or old one?

[edgy chasm]

same build, also had this issue a few times after destroying a bunch of vehicles, respawning, and teleporting back - no tracers, no explosions, no smoke stacks despite wrecks glowing orange, and no particles on bullet impacts

seemed temporary though, getting away from the area resolved it

[azure roost]

Since the newest update, got an issue where if I load arma on perf branch with mods it'll get to splash screen and just do not responding and crash. Repaired, tested with normal branch and it works fine.

Different to what I had before, have got the CPU Count set properly this time. Any mod at all, doesn't matter how many.

Any ideas?

[azure roost]

Fixed, NVidia Graphics driver was out of date. Worth noting that it didn't cause a crash before this latest update/fix.

[heavy vortex]

Sometimes that's just timing. I don't think any actual display logic was changed recently.

[ember owl]

On new but all of my crashes were on client not server.

[heavy galleon]

Ah, we just played, no crash on dedi. JSRS was force loaded, but old one

[woven loom]

No client side issues for me while playing today, but low player count (10-ish), mix of profiling and stable, server running stable. Very few small mods + CDLC

[whole cloud]

[whole cloud]

what does getting in mean? where and how are you failing

[whole cloud]

Send me the crash reports

[latent sigil]

Hello I cannot start the game while using profiling.

[whole cloud]

Provide more details please

[latent sigil]

While using profiling branch from the steam the games get stuck at games logo splash screen. Tried starting with no parameters as well gets stuck at splash screen.

[whole cloud]

Send RPT

!rpt

[frozen tundra]

Arma RPT

Arma generates a .rpt log file each time it's run, which contains a lot of information like the loaded mods, or any errors that appear, this log file can be very useful for troubleshooting problems.

To get to your RPT files press Windows+R and enter %localappdata%/Arma 3

Additionally see the wiki page for more info: https://community.bistudio.com/wiki/Crash_Files

To share an rpt log here, please use a website like https://pastebin.com/ (Set expiry time to 1 week or less) to upload the full log, that way the people helping you can take a look at it and try to figure out the problem you're having together with you.
Note: RPT logs can hold personal information relevant to your system, the game or others.

[latent sigil]

https://pastebin.com/yKhLHqzV

[whole cloud]

-exThreads=7 -enableHT
That is not "with no parameters"
that exThreads does nothing, and the enableHT more likely hurts you.
-checkSignatures
Why? Do you actually need that?

Ah its multiple RPT's smushed together

[gritty wasp]

For me it crashes too.

[whole cloud]

Yes it crashes

[gritty wasp]

Ah. I thought you got error message instead of crash

[latent sigil]

I saw in some guide to enable the HT, but dont see any impact if its enable or not

[whole cloud]

Your game freezes, what does CPU usage show in task manager?
0%? 100%

[latent sigil]

one min i need to check it, I rolled back to normal

its around 30-40%, pc stays responsive only game freezes

[whole cloud]

Please get it to frozen, then ..

in task manager, under details.
Right click the game and create a "dump" "memory dump" something like that I don't know the english word

[patent sky]

"Create dump file"

[latent sigil]

ok

[whole cloud]

And send me that, probably pack it in a zip and upload somewhere like uh.
wetransfer or any of these file transfer websites.

That particle issue seems to be related to respawn right?
I heard it happen in koth, antistasi. Generally game modes that have respawn.
And I never had it, because I never respawn? 🤔

Maybe combined respawn with the far teleportation

[whole cloud]

If you use process lasso or any way to limit the game's CPU cores, you need to use the -cpuCount parameter to tell the game that.

It'll just start very slowly if you don't.
Even after the fix its still quite bad because there are many small tasks in this.

But generally people who don't mess with things won't have problems with this

[latent sigil]

Thank you for your help 🙂

[foggy vine]

Had it happen on a custom mission without respawns

[patent sky]

did it start happening randomly in the middle of the mission or right away from the start?

[gritty wasp]

@heavy galleon did have it after respawn?

[heavy galleon]

Nope

Nope

[foggy vine]

Hmm. Not quite sure honestly, didn't pay much attention, but it may have happened either at the very start of the mission or a little after the start before mid-length time period of the OP

[autumn timber]

Any way to detect being limited by the OS and maybe add a warning to the RPT in case someone is forcing an unreasonable number? (yeah, I know, the chances someone will read that warning are quite slim)

[heavy galleon]

Deddy daddy, what about overall particle limit. Lately we've had smokes disappearing, like the emittors were "reset" and all particles deleted.
Seems to be tied to some limit.

The particles from bullets use the same feature I would assume

^^ but that is global for everyone

[whole cloud]

The particle limit being a thing is old, and I'm not fixing that for now.
If new particles are spawned, exceeding the limit, the oldest ones are force-killed. But the old ones will recover because the emitters will still be running

[heavy galleon]

For us it looks like it gets triggered immediately

I'll try to find a video, I think I sent it to liquid few months back

Also weirdly if you alt tab as I did here, the smokes "reappear" and immediately get deleted. Almost like fi the limit didn't apply while in background

[whole cloud]

But the particles are still there, just forced down to very few.
That's not the issue where particles are completely gone

[heavy galleon]

Ok.
But for the respawn/teleport.

On Friday I think I encountered the missing particles. Didn't see impact of few of my shots, or shots of the sniper next to me, while he told me that he's hitting the bmp (and he was)
Neither of us were Respawned nor teleported.

[whole cloud]

Not seeing particles of sniper impacts a km away is also not new.
I've had that for years, it always felt unreliable

[heavy galleon]

Saw some, some I didn't

I'm gonna go schizo from this game one day

[whole cloud]

I have another idea, we'll try that in v31

[heavy galleon]

I'll keep an eye on it on Friday

[whole cloud]

With the allocator changes, the particle rendering was also moved around, to start a bit sooner.
But if it breaks something that should always break it, not work fine 95% of time but rarely break.

I'll try to move it back down and then we'll try and see 🤷

[analog acorn]

When this performance stuff comes to stable, it might be an idea to put a note in the SPOTREP that's like "hey, if you were trying one of these weird hacks for better performance, maybe stop, the default handling is better now and your hack probably wasn't helping anyway".
It's really nice all the information you give us about how this stuff works, but it's still tucked away in a corner of the Discord, which isn't great for info accessibility. It'd be helpful to have a permanent official source to try to combat all those years of dodgy YouTube videos giving counterproductive advice.

[fast hornet]

it's probably too late by now, people will still find one of those useless "guides" on steam community or elsewhere from 10+ years ago. But can't hurt as well 😄

[spiral pond]

i would make the two(?) parameters obsolete/deprecate and add new ones

this way you drop all the faulty use and encourage actual usage from now on/a reset by those who know what they are about/invest some time to get educated about them

[tacit frost]

I returned to Arma after a long break and noticed that the performance in the profiling build has dropped significantly, in the sense that it is now about the same as in the stable version.
Before that, I last played around 02/20/2025 - then the difference between the versions was impressive, with an increase of literally tens of FPS.
I checked the performance in the same scenario.

[opal hound]

What's the scenario you're comparing?

[tacit frost]

This:

https://steamcommunity.com/sharedfiles/filedetails/?id=2132357035

[heavy vortex]

Yeah, there's https://learn.microsoft.com/en-us/windows/win32/api/winbase/nf-winbase-getprocessaffinitymask. Could use that to cap the cpuCount calculation.

I think the assumption was that setting affinity is quite rare and people who use it know what they're doing :P

[whole cloud]

But that wouldn't help with anything that sets the affinity after game start. Like task manager would be

[heavy vortex]

You mean actively using a UI to set affinity after starting the game?

[autumn timber]

quite rare
To paraphrase: in 60% of the time they set it every time 😉

[heavy vortex]

Those people know what they did. The issue is people who just set something six months ago and forgot about it.

But I dunno. Maybe utilities like process lasso actually set affinity too late to be detectable at startup.

[swift narwhal]

For the first time I also just got no particles/effetcs bug, happened after I respawned in Antistasi need me to do something? (create dump or send rpt)

[patent sky]

@whole cloud do you still need/want a prof framecapture when someone has the no particles bug?
I assume he's running perf tho

[swift narwhal]

i am

ok effects came back randomly now (I checked my task manager after it happened and created a memory dump just in case, my RAM was on full usage, I closed my browser to free up some RAM, maybe it helped?)

[heady echo]

When I join for example EU#5 on koth, I click on join the server from the in-game server browser and I Just get a black screen. It hangs and you can see the load bar go upto around 5%. Its only been happening for 2 weeks. I've been on profiling for about 3 months now.

(I have tried joining different missions / servers too, its not just on the one server)

I leave and rejoin a few times sometimes It will work. sometimes I have to restart the game to get it to join. and I have tried switching back to the normal branch and it works normally.

[heady echo]

I've found to fix bullet impacts/smoke etc, that putting game into windowed mode and changing back to your regular settings fixes it too. just a temp workaround

[inland dew]

shouldn't have to do any of that

[woven loom]

I didn't really see MMG impacts at a distance too, few hundred meters, but i feel like that's normal possibly? No other particles were "missing" that I could tell

I have a feeling it's an interaction with grass rendering workaround that might be hiding impact particles (e.g. sinking them like units are sunk), or maybe they're not rendered that far out?

Or people just need more zoom

[whole cloud]

yes

[whole cloud]

Can you try manually going back to v28 using the google drive download, and seeing if that fixes it too?

[whole cloud]

2.18.152881 new PROFILING branch with PERFORMANCE binaries, v31, server and client, windows 64-bit, linux server 64-bit
- Tweaked: Improved performance of engine closing files (could in some cases cause lag spikes) (Thanks @edgy chasm )
- Fixed: Microfreeze when pressing "Next Target" key - https://feedback.bistudio.com/T191741
- Fixed: Potentially fixed sometimes missing particle effects (Please report if it happens again)

If you don't want to use the Steam branch, the files are also available for alternative download here:
https://drive.google.com/drive/folders/15p9j7C2nHUt6NoVfChX4YFuqzFXzblJh
Note: There are separate Dll files that also need to be placed into Game folder.

[inland dew]

2.18.152881 new PROFILING branch with PERFORMANCE binaries, v31, server and client, windows 64-bit, linux server 64-bit

• Tweaked: Improved performance of engine closing files (coulFixed: Microfreeze when pressing "Next Target" key - https://feedback.bistudio.com/T191741
  ❤️

[inland dew]

if particles fix/reverting works, finally possible to upgrade from v5 😱

[heady echo]

I am in work but reading some stuff about cpu cores count up in the chat etc I might have messed with that I will try reverting all changes before Incase it’s a me problem lol , completely forgot about it

I’ll test later anyway

[opal hound]

Is it possible for me to check if a server is running profiling?

[whole cloud]

I think the server browser shows you the build number.
But whether perf or prof, no

[opal hound]

that's fine, that's what I needed

[slender lagoon]

what is the difference between perf and prof?

[whole cloud]

One does profiling, the other doesn't

[nocturne cairn]

anyone know how to fix 'Files are not signed by a key accepted by this server'

[opal hound]

which files

[nocturne cairn]

literally just put the html in the html modlist loader on the startup and loaded the list into the files list

[nocturne cairn]

mods

[quaint flame]

#arma3_troubleshooting

[nocturne cairn]

rog thanks

[opal hound]

Anyway, still getting issues with both players stuck in the lobby screen and Wrong signature for file shortly after join

This only seems to happen on the higher traffic server of the two running profiling, players who cannot get past lobby on server 1 (high traffic) seem to be able to connect to server 2 just fine with the same session
Network settings have been tweaked but neither seems particularly network or performance stressed

[fast hornet]

stumbled upon https://arma3-servers.net/version/2_18_152881/ a while ago.. no idea who runs the site etc. - but their gui version filtering is kinda neat 🙂

[opal hound]

okay thank you, that actually helps a lot. they are running different versions 🙃

1 is 152803 and 2 is 152881

[whole cloud]

Trying to figure out the "last" netcode bug.

Split messages need to be re-assembled.
For that the messages need to go into reassembly step in correct order.

There is a bug where, while a re-assembly of a message is active, a part of a different message comes in. And two different messages get spliced together and cause corrupted data to be processed.

So, the mess.
The server only accepts messages in a certain range, to filter out way too old messages.
That range is regularly moved up, when new messages come in.
But it doesn't consider when a old message was still missing, it can just move the range beyond it.

So when the message that we were missing, finally comes in. We just drop it because its too old.
Meanwhile the other code is still waiting for that message to hopefully come in, with thousands of messages depending on it, queued up.

Usually, when a guaranteed, ordered message was lost. The game cannot continue, it should just trigger a disconnect but it doesn't.

The ordering works such that each message contains the serial number of its predecessor.
The receiver then checks if it had already received the predecessor, if not then its queued up until it arrives.

So you'd think if one message in that long chain goes missing, everything would be held up until we disconnect.

But, we do not send the predecessor to the last ordered message, we set the predecessor to just the last message.

So, server sends a unordered message, that does not get queued up because it doesn't have any predecessor that could be missing.
Then the next ordered message comes in, we did get its predecessor so its fine, we let it through.
And.. We just successfully skipped over thousands of queued messages.
We missed thousands of guaranteed messages that should never go missing, and we have a big section of messages in the queue, that will never be resolved, because the missing part will never come in.

TCP does this by just refusing to accept new messages when there are too many ones queued up. But we just don't do that :/

[gritty wasp]

Meanwhile the other code is still waiting for that message to hopefully come in, with thousands of messages depending on it, queued up.
Hachikō will wait.

[whole cloud]

But, we do not send the predecessor to the last ordered message, we set the predecessor to just the last message.
Fixing that "works".
Instead of the ordering chain breaking and invalid messges being spliced together.
Messages just forever get pushed into the queue.. up to16gb memory usage and 11 billion messages in the queue 🤣

[opal hound]

Anecdotally, startup seems better after todays update, less (or less severe) CPU spiking

[quaint flame]

Did you look into that valve netcode thing you posted a few days ago? Anything usable from that?

[whole cloud]

Lots of usable things in there surely, but no.

[median belfry]

if you fix that issue that would probably also fix all the inventory issues

🤞

[whole cloud]

No

[median belfry]

isn't the issue with items vanishing that guaranteed messages are getting lost?

[whole cloud]

No

[ripe anchor]

Usable things that will be implemented later on or nope?

[heady echo]

yes v28 works, I load in instantly to the server. The only difference i notice is I have a map loading screen pic on v28 where as its a black screen on newest version

just before it I reverted my cpu affinity etc and it still would hang on the loadscreen for ever, it was frozen in taskmanager.

[heavy vortex]

As a dumb test, try setting -cpuCount=4 (or some other suitably low value) and see if that fixes it with v31.

[vivid rune]

Would this fix maybe also the connection error of clients to the server that caused by to long mod string? Workaround for that was to increase "steamProtocolMaxDataSize" size in the server config.

[whole cloud]

Potentially

[whole cloud]

Then it's something netcode, and not CPU related.
But why only you 🤔

Frozen in task manager?
Does the game generate freeze dumps to RPT folder?
If not then you can manually create a memory dump by right clicking it in the task manager details panel, with that I should be able to see where it's hung

[whole cloud]

No

[gritty wasp]

Let's continue with guessing. My turn. This will fix signature check kicks?

[whole cloud]

It would fix the game crashing when spamming several MB/s of publicVariable/remoteExec data to it

[inland dew]

have tested my 14900ks (32 threads, so 8 p + 8 ht +16 e), but with ht disabled in the bios, so 24 threads total (8 p + 16 e)

8 p + 16 e (24 threads ht off) and not using -cpuCount=24 - fps is similar to -cpuCount=8
-cpuCount=8 - fps is lower than with -cpuCount=24

avg of min fps and avg of avg fps after 7 runs per setting, not counting the 1st run

[heavy vortex]

So, server sends a unordered message, that does not get queued up because it doesn't have any predecessor that could be missing.
Then the next ordered message comes in, we did get its predecessor so its fine, we let it through.
Could this have caused loss of guaranteed messages even in the case where there wasn't a split message waiting for reassembly?

[whole cloud]

Yeah.
The split reassembly stuff is on a different layer than this problem

[heavy vortex]

Excellent

[inland dew]

somehow, fps is always lower, if one doesn't use -cpuCount...
all 32 threads (ht on) and not using cpuCount, results in lower fps than when using -cpuCount=32

same when 24 threads (ht off) and no cpuCount vs. -cpuCount=24...

but yeah, with ht on, fps is the lowest, cpuCount or not

[inland dew]

ok, p+e cores (ht off) and -cpuCount=24 it is then

[whole cloud]

The number of actually used threads is listed in the top of RPT.
Afaik the current limit is 16

All that enableHT/cpuCount does, is change that number.

[inland dew]

limitFPS works in a strange way...
if i do -limitFPS=500, it results in 330-333-337 fps max displayed in video settings right upper corner...

so i have to set it to a way higher number so it tops out at 500 fps max for real

[fast hornet]

server fps was always about ~10% below the set limit for some reason.. at least according to diag_fps

[inland dew]

default is 50. but it's never 50. max is like 46 or 47, i think

and that only occasionally

most of the time less than that

[heavy galleon]

and mine is set to 75 and I usually get 76 xD

[inland dew]

oh! i see!
so setting cpuCount to >16, boosts fps, provided one really has more than 16, since auto detects only max 16.
now i understand 👍

[heavy galleon]

Do you want the game to run on E-cores though?

[inland dew]

fps is higher with e

having p+e on, but doing -cpuCount=8, results in lower fps
disabling e cores in the bios, also results in lower fps

tested that today and many times in past few years...

[fast hornet]

at least when i tested the server on a 13900 it was massively faster when i set affinity and cpuCount to the 16 p-core threads.. e-cores just messed things up. I'm just using the e-cores for other stuff like the system and database and such

[patent sky]

AMD has 16 core cpu's without Big/little architecture

[heavy galleon]

do you mean big.LITTLE?
(literally passed an course today that touched that topic )
Anyway, which ones are those?

[inland dew]

that's not how it works thought...
when you have ht and e enabled in the bios, both at the same time, fps is the worst.

but if you have only p + ht or only p+e, fps is the best

[patent sky]

r9 9900X(3D) has 12 cores
9950X(3D) has 16
7000 series also has some iirc

[inland dew]

so fps was lower, not because of e, but because of having ht and e enabled at the same time

[fast hornet]

that doesn't correlate with my findings at all.

[heavy galleon]

do those really have ecores?

[patent sky]

no, thats why I said "without Big/little architecture"

[heavy galleon]

ah, can't read

[fast hornet]

Current AMD series have chiplets with a number of cores on each of them. With X3D models one of those chiplets has the bigger cache while the other doesn't. Also communication/bandwidth is better within the same CCD compared to cross-CCD.

[inland dew]

hyper-threading hurts only if e-cores are also enabled.
if e-cores are disabled in the bios, so one has only p-cores, hyper-threading boosts fps, of course.
tested dozens of time in a3, with different patches, with way too many runs

since amd isn't hybrid, smt on amd doesn't hurt perf at all

[heavy galleon]

but, the chance of a long running task being put on ecore and you havaing to wait for it

[inland dew]

so either disable ht in the bios or you disable e-cores

p+ht (no e) or p+e (no ht), fps is ~same

[heavy galleon]

or make arma not being ran on e-cores what so ever and keep them taking care of ts, windows, browser the way they are intended to be used

[inland dew]

well, like i already said - fps is higher than just running on p-cores

i have ht disabled in the bios, since i have e-cores and ht off = way lower cpu voltage and p-cores temperature or higher p-cores frequency at same voltage/temperature

so the choice is obvious - ht off (why split each core in 2), since you have 16 more real cores that are way more powerful than just threads and there are twice as many of them
(offtopic off...)

[gusty viper]

The reason for FPS limiter being off is that the Arma process does not set a fine grained timer resolution. See also: #perf_prof_branch message

[whole cloud]

No you don't understand. cpuCount is also limited to 16

[inland dew]

if so, why is fps consistently higher then, when using -cpuCount=24 (ht off) or -cpuCount=32 (ht on) vs. not using cpuCount with ht off or on, thus defaulting to auto 16?

not by much, but reproducible

min fps only marginally, but max fps one can easily see the difference, thus also slightly different avg, over multiple runs, 5-7+ runs (not counting the 1st run)

[light cargo]

huh strange

maybe something could be done about that

what's puzzling is how you get higher fps with e cores

[fast hornet]

he doesn't get higher fps with arma being on e-cores as well, but e-cores being enabled in general. Since the system scheduler can shift non-arma threads to the e-cores and thus leave more room for arma on the p-cores.
At least that would be more consistent with my own tests.

[light cargo]

oh okay

[woven loom]

One again I shall suggest that Arma 3 get an Intel APO profile 😄

Otherwise feels like BI has to fix an Intel/OS level issue

[inland dew]

you can use Intel APO though, even if it doesn't have an arma specific profile.
helps in many games that don't have a profile

i see that min and max fps fluctuate really a lot, for my liking (not within margin of error)

so will redo tests, just comparing avg fps that is somewhat ok

min fps difference from run to run up to 10 fps sometimes

max fps difference from run to run can be even higher

[woven loom]

You can, but I think that's somewhat experimental per my understanding. Different from if Intel specifically tunes a profile for the game.

[inland dew]

sure. but one can try and see if it still helps

[slender lagoon]

placebo effect?

[inland dew]

rerunning everything, killing all unnecessary background stuff and pc reboot on each parameter change, for max consistency

[inland dew]

ok, between 8p+8ht+16e, 8p+16e (no ht no cpuCount) and 8p+16e (no ht with cpuCount) there is less than 2 fps min difference and less than 1 avg fps difference...

so whether one limits to 8 threads only or auto 16 max (8p+8e?) - there is no difference...

no affinity used!

will now try using affinity and cpuCount

[heavy vortex]

win10 or win11?

[inland dew]

10

[inland dew]

done affinity from 0 to 15 in task manager, with ht enabled in the bios + cpu count 16 and fps is bad...

[heavy vortex]

Try setting affinity a bit higher than cpuCount.

[empty goblet]

check if your your primary and active threads aren't on the worst possible e-cores / non physical cores

[woven loom]

i think you need a spreadsheet 😄

[heavy vortex]

Apparently the listing order for these is P-cores first, with 0/1 as the SMT cores for one P-core.

So if you were doing 8-core affinity with HT enabled you'd probably want 0/2/4/6/8/10/12/14.

[cold vale]

TIL: -debug doesn't appear to be detected properly if included in a -par file

[whole cloud]

Magic or placebo.

[potent sparrow]

mine as well - 13500, taskset -c 0-11 was the best setup I found, I tried to be smart and taskset the cores more finely to number of processes (since 1 server/3HC's divided nicely into 12) it did not go well :D, in the end I just landed on "stay on the p cores"

[knotty wraith]

I checked Intel APO on ARMA - no effect

to get better fps on win 10 i played with lasso and cpucount, but after installing win 11 - it was not necessary: ​​win 11 loads P cores correctly

as a result, I got the same fps on win 11 that I got on win 10 with the help of dancing and tambourines

but I couldn't get more fps on win 11, no matter how I twisted the cores up and down

[inland dew]

i retested for several hours, with reboots for each setting change and yes, there is no difference between autodetect, -cpuCount=16 or 24.

the worst performer was -cpuCount=16 with affinity 0 to 15, whith hyper-threading on.

-cpuCount=8 with affinity 0 to 7, with hyper-threading off, was 2nd worst

-cpuCount=8, but no affinity, with or without hyper-threading, was 3rd worst

then -cpuCount= 16, 24 or autodetect, with hyper-threading on, were all kinda same.

the best was with hyper-threading off in the bios and autodetect or -cpuCount=16

basically, when one has hyper-threading on and/or locks arma to p-cores only via affinity, fps is the worst.

when no affinity, even if -cpuCount=8, it's limited to 8 threads total, but not specifically 8 p-cores, but also e-cores.
fps is higher this way, because although e-cores are slower and don't do much, it results in having access to more cache, since e-cores have separate/additional L1 and especially L2, per cluster of 4 e-cores

[knotty wraith]

as Dedman noted, this is not a game problem, but an OS problem...

[slender lagoon]

Thank god i switched to amd

[knotty wraith]

I'm sure Groove is doing this for the sake of the test, he has the opportunity.

out of interest 14900ks is the best processor for a server? is there an analog for amd? sorry for offtopic, but just since the topic about 14900 ks came up

Sorry it's not in English, but you'll understand the gpt rating

[slender lagoon]

he kept testing for days he probably had a brain melt down 😂

[plain trout]

#hardware_vs_arma should be the better source

[fading seal]

So, in theory, isolating L1 and L2 to a single thread could give a little boost? In my case L2 and L1 are shared by a "real" core and a "ht" core for example

I'll probably still be bottleneck-ed by my old ass CPU but I remember this boosts CPU mining as well

[whole cloud]

correct, there are a few parameters that don't work in -par file.
Some need to be, because they need to be applied before files can be loaded.
I can probably put debug into both though. You might just miss it with early logging

The par file needs the config parser.
The config parser has the ability to store context, and context is only used if -debug is enabled. So -debug must've been read before reading the par file, otherwise the context's in that config parsing would be lost.
But in practice thats probably not an issue

Also there is a check for exThreads in par file, that tells you that it didn't take effect because it doesn't work there.
But the other parameters that also don't work there don't report it

-exThreads can be specified via command line onle, not in the .par file
including typo

Our code actually lets you set cpuCount in par file.
But it has no effect because the job system was already initialized before that.

Actually the par file is loaded before the job system loads.
But the cpuCount is only checked before the par file and a second time after the job system was initialized 😄

[light cargo]

odd

[cold vale]

Yeah I figured most people don’t use the par file so some behavior may have fallen through the cracks but this is for a test env so not really an issue I can’t work around. I only noticed because the web browser commands threw an error/warning about the missing arg.

[whole cloud]

btw argo had a -cpuAffinity parameter. Which theoretically I could enable.
Not sure if that actually works well

[whole cloud]

And thats because you then noticed that the context was missing?
Not the commands throwing the error because -debug wasn't set right?

[cold vale]

Yep, checked the RPT and it was significantly smaller than usual lol

[light cargo]

with some adjustments that could work for excluding e cores without going into the bios

[whole cloud]

Also taking that chance to add some warnings.

17:38:12 WARNING! Process affinity was attempted to be set to more cpu cores than the system has (36 out of 32). Wanted mask: 0xFFFFFFFFF Available mask: 0xFFFFFFFF
17:38:14 Process affinity forced to 0xFFFFFFFF (requested 0xFFFFFFFFF)
17:38:15 WARNING! specified cpu count, and cpu affinity, do not add up to the same number. cpuCount: 12 affinity:32. Remove -cpuCount parameter, the affinity parameter will set the count.

17:42:32 WARNING! CPU Count is higher than available CPU cores according to affinity mask. Performance will be inhibited! Wanted: 16 Available: 2

This also detects when the affinity is set by task manager or process lasso.
But only if it happens very early before the job system is initalized

17:56:46 -exThreads= can only be specified via command line, not in the .par file
17:56:46 -minimize can only be specified via command line, not in the .par file
17:56:46 -maxFileCacheSize= can only be specified via command line, not in the .par file
17:56:46 WARNING! CPU Count is higher than available CPU cores according to affinity mask. Performance will be inhibited! Wanted: 16 Available: 3

[inland dew]

well, if one can't specify specific threads as affinity, but only total number of them, it doesn't help, imho

[whole cloud]

You can set the affinity mask for the whole process.
I thought about maybe doing thread pinning too, but nah

[inland dew]

if you don't want to disable ht/smt in the bios, but want to avoid it, using affinity, then you have to be able to select specific threads, not just their total number

[heavy vortex]

Might be a misunderstanding. The affinity mask gives full control over which CPU cores the process will use.

What's not defined is which Arma threads end up on which cores.

[inland dew]

cpuCount no, but affinity should.
if no, then it's not any better than cpuCount

[whole cloud]

task manager cpu 0 and 1 == -cpuAffinity=0x3

[whole cloud]

Where did you get the idea that the cpu affinity parameter, would not set the cpu affinity?

[light cargo]

and wouldn't it be useful to have thread count parameter or that's scope creep?

[heavy vortex]

That's what -cpuCount does.

[inland dew]

in your warning messages, affinity:32

[whole cloud]

That is the count, by counting how many bits are set in the affinity mask

[light cargo]

also i saw #community_wiki why 2.22 despite it being on profiling ? means it'll supposedly land to stable after that much testing?

[inland dew]

i think that doing affinity as only "good" for sever owners

[whole cloud]

"why"
What else would it be if not that?

[light cargo]

is there a reason for the parameter to be in hex instead of decimal?

[heavy vortex]

lol

[whole cloud]

it supports both. You might even be able to do octal if you wanted

[light cargo]

oh it could land on 2.20 because why not

[whole cloud]

Because 2.20 is already closed

[light cargo]

Ahhhh got it!

[heavy vortex]

It's not gonna be very readable in decimal :P

[light cargo]

feature freeze when rc opens?

[inland dew]

because if you do affinity, as player (not server owner with 1 machine and more than 1 server on it), you cut off access e-cores that have additional cache, thus higher fps
so better to use cpuCount or autodetect, so it uses like 8 threads, but not specific, rather any, so arma can access e-cores additional cache

[fast hornet]

For an Arma 3 Server the absolute best would be the X3D-CCD of a 9950X3D if you wanna go all in.. nothing will come close.

[inland dew]

linux -> amd 3d, windows -> intel
just as player -> amd 3d

[restive pilot]

is binary with 0b supported too?!

[whole cloud]

-cpuAffinity only works on windows. Because.

[whole cloud]

https://learn.microsoft.com/en-us/cpp/c-runtime-library/reference/strtoui64-wcstoui64-strtoui64-l-wcstoui64-l?view=msvc-170
no

[inland dew]

but yeah, please, implement/port it
might be useful -> no lasso

[fast hornet]

is -cpuAffinity in v31 already? Or will it come in the 2.20 perf versions? 🙂

[inland dew]

practical for server owners

[whole cloud]

I did that stuff today

[woven loom]

Is there an advantage of the 9950X3D vs 9800X3D? Can A3 server actually use the extra cores?

[fast hornet]

you'd put the rest of the system threads onto the other CCD, so A3 would have the X3D exclusively.. besides that, no, no advantage over the 9800X3D.

[woven loom]

Ah okay that way

[scarlet jolt]

well, I've managed to break something.

Working on a mission, right click play as unit on a guy, spawned in and instantly froze, RPT spamming something about freeing memory and had to force quit the game.

[opal hound]

I had a similar thing yesterday but I think it was just me running out of memory

21:46:35 Longest free VM region: 4161123216 B
21:46:35 VM busy 116 GiB (reserved 94 GiB, committed 22 GiB, mapped 339 MiB), free 130956 GiB
21:46:35 Small mapped regions: 52, size 245760 B
21:46:35 Emergency memory free, freed 8 KB
21:46:35 Virtual memory total 131072 GiB (140737488224256 B)
21:46:35 Virtual memory free 130956 GiB (140612833492992 B)
21:46:35 Physical memory free 22 GiB (23948861440 B)
21:46:35 Page file free 11 GiB (11805655040 B)
21:46:35 Process working set 8.7 GiB (9319452672 B)
21:46:35 Process page file used 21 GiB (22974926848 B)
``` x about 50
Some odd numbers in there

[scarlet jolt]

Yup same as mine

14:32:54 Emergency memory free, freed 8 KB
14:32:54 Virtual memory total 131072 GiB (140737488224256 B)
14:32:54 Virtual memory free 131053 GiB (140716983119872 B)
14:32:54 Physical memory free 14 GiB (15080443904 B)
14:32:54 Page file free 24 GiB (26225020928 B)
14:32:54 Process working set 5.1 GiB (5522493440 B)
14:32:54 Process page file used 13 GiB (14373502976 B)
14:32:54 Longest free VM region: 134607568 B
14:32:54 VM busy 19 GiB (reserved 5.2 GiB, committed 14 GiB, mapped 336 MiB), free 131053 GiB
14:32:54 Small mapped regions: 32, size 135168 B

[heavy vortex]

@scarlet jolt Can you replicate?

[empty goblet]

ok, but ... what could you use to tell arma3(server) process which CPU use as the primary cpu for the primary thread ? ... meaning , using out command-line parameters
ie let say the core 6 is the best core of all the cores 😁 hence i want that core to be the the primary core for the primary thread (for now the most intensive/important thread)

[plain trout]

You limit the top frequency of the other cores. That will make Windows choose the one that has the highest clock speed

[empty goblet]

and no i don't trust Microsoft kernel scheduler, because in 101 of 100 cases it got the answer wrong

[plain trout]

That’s an older workaround for the 7950x3d when their drivers still had issues properly utilising the cached cores

[empty goblet]

sorry to burst your hope bubble there but not even that work

[plain trout]

It worked with the 7950x3D.. at least for gaming

[empty goblet]

worked is the correct definition, past sense, not anymore (and it was only with the certain AMD driver/AGESA versions)

for intel it was worse, always

[scarlet jolt]

nope, I've been trying but no luck

[heavy vortex]

5.1 GB is kinda small as a working set. Do you use -maxmem to constrain usage?

[scarlet jolt]

Nope

[woven loom]

fwiw i just tried in a blank mission on stratis, no mods, and it was fine. So definitely more going on

[heady echo]

I can't recreate it on v31. Its now working as intended 🙂

If it happens again I'll create a dump, its strange because I reinstalled twice etc, All I've done since then is go back to v28, then I verfied files on steam and its fine now.

[thin wyvern]

For client set affinity to some cores 0,2,4,6 game still see 8 cores...from .rpt w/ affinity=155 in batch file CPUCount : 8 😕

and get fps drop

I get best min fps on my 5800X3D w/ ht or cpucount

[woven loom]

For a 5800X3D you probably don't need to do anything. On my 7800X3D I just let the game auto-detect

[thin wyvern]

And I have set priority to Above normal in registry for arma3_x64

[thin wyvern]

Look in .rpt file how many cores game see

8 right?

[woven loom]

Can't check right now but I think it's 8 yeah

[thin wyvern]

Test YAAB w/ cpuCount=16

[woven loom]

Why though, it's unnecessary

[thin wyvern]

Look in my pic and you will see that I have best min fps with that patrameter or ht on

SMT is on in bios

[woven loom]

Your FPS results look like they're always within 2-3 fps. That's equivalent. YAAB has some run to run variation in fps anyway.

Min fps is especially going to vary a bit because the benchmark involves lots of AI and they can behave somewhat randomly

[thin wyvern]

w/ cpu or ht I get >60 but if I set affiniry to some cores I get lower or some drop to 2 fps 🤔

[woven loom]

From what I can see, CPU count 16 result is the same as HT enabled CPU count 8 is the same result as affinity FFFF

System affinity is 1-2 fps lower on average

[thin wyvern]

With cpu and ht game report 16 cores in .rpt

[woven loom]

Yeah because HT just doubles CPU count

[thin wyvern]

But fps drop?????

[woven loom]

Yes, so you can turn on HT because you get slightly higher fps when you enable it.

Or change CPU count to 16. For the game it's the same effect.

Don't bother with affinity

[thin wyvern]

W/ affinity set, game always report CPU count 8 🤷‍♂️

And have fps drop if affinity not set to 0,1,2,3 or 4,5,6,7 order 🧐

[woven loom]

So the issue is that you want to set affinity while enabling the HT parameter?

[thin wyvern]

No

[heavy vortex]

555 affinity is only six cores, so I sure hope you set cpuCount 6 for that one.

[thin wyvern]

Yes I miss one 5 👴

5555

And I only set affinity in batch and don't use cpucount or ht

[heavy vortex]

Hmm. My conclusion with most of these tests is that 881 is still extremely intolerant to having even slightly less cpu resources than it's expecting.

[thin wyvern]

Like this START "ArmA 3" /affinity 155 "C:\Igre\steamapps\common\Arma 3\arma3_x64.exe" -setThreadCharacteristics -hugePages -malloc=mimalloc_v219 -cfg=test.cfg -name=test -world=empty -ShowScriptErrors -noPause -noPauseAudio -debug "-mod="

[heavy vortex]

Trouble with 8-core affinity is that other processes might still hit those cores a bit.

[woven loom]

I am just confused why you want to set affinity when setting ht or CPU count give you the same performance benefit as setting affinity, and with less headache

[thin wyvern]

That why I try 5555 to set affinity to "real" cores

[woven loom]

On the newer Intel CPUs it makes sense, on AMD I think it's mostly pointless given the tiny benefit

[heavy vortex]

It's a reasonable lack of trust in the OS putting threads in the right place. But yeah, for A3 perf branch it doesn't seem to work well.

[thin wyvern]

I don't try w/ SMT off in bios..maybe in next few days

[empty goblet]

completely different problem ... i care about what the primary thread uses as it's home-core first and then about the rest of available cores / affinities

[thin wyvern]

For server that is 💯

[heavy vortex]

It should follow that SMT off will have similarly bad performance. Unless there's something else going on.

[empty goblet]

i know how 'trick' it as workaround (not via game engine) but it's not perfect and has some flaws

[heavy vortex]

In that case Arma's still expecting a whole 8 cores to itself.

[woven loom]

There is no reason for normally configured processors to turn off SMT

[empty goblet]

there is no 100% ... , imagine server CPU has 16 (or 32) cores ... i want primary thread to be at 6th core and use ALL cores for the rest

[thin wyvern]

Yes but for test I will try that

[thin wyvern]

That is best core on that cpu?

[woven loom]

I guess, I just think that operating systems do a mostly fine job unless it's with something funky like Arrow Lake etc. Especially with X3D chips, most of your performance comes from the cache, the threads are usually being fed pretty well.

[empty goblet]

the flaws i know is that

• A. windows kernel assigns random core to the primary thread
• B. windows kernel assigns last core to the primary thread
• C. sometimes the A or B cores are HT sometimes not
  plus
• D. if running multiple processes sharing same core, theirs primary threads ends on the same core from A or B and C case even after the load is already distributed (delayed start)
• E. cases A to D happens even if using manual forced affinity across multiple cores

[woven loom]

If the frame times are fine and you can't tell then 1-2 fps difference is academic

[empty goblet]

so while i have makeshift workaround solution how mitigate this, it's flawed

[empty goblet]

also this ends badly if you have like 4 or dozen server processes running on the same system and the windows kernel scheduler decides all the primary threads will run on last core 🤣

and no the (in example) last core isn't the fastest or gold or least used etc.

hence i would prefer way to designate desired primary core for primary thread of the process w/o being forced to do it manually with messy workaround (which come with other problems)

[thin wyvern]

Until AMD and Intel along with M$ fix this, we're screwed.

[empty goblet]

hence why i would like to know if this is doable or not, i know there is way to define target core for (new)threads

[thin wyvern]

I'm ignorant about Linux but maybe you can try there

[thin wyvern]

5fps difference is not the problem but fps drop is

[plain trout]

In windows 10 I was able to manually set which cores a process was allowed to use. I’m not sure if 11 still allows that access

[runic sigil]

It does, in basically the same way (task manager in details tab is the easiest)
(Called core affinity)
It's less needed most of the time, since 11 has a better scheduler for the most part
Process Lasso works for setting it persistently on a process, other ways as well (Process Lasso has to be running for it to do that though)

[woven loom]

okay but then don't use affinity?

here are my results

https://imgur.com/zW4aQqp

https://imgur.com/9RAmK2Y

HT enabled does measurably worse by 3-5 fps

https://imgur.com/SgVJh4Z

If you look at absolute minimum fps, it is higher for HT. But there will always be some random frames that drop really low in YAAB. It is meaningless, because overall in my case the game does better without manual affinity settings and without HT parameter.

You want to see averages along with 1% slowest or 0.1% slowest fps/frame times. And the autodetected parameters (i.e. cpu count 8) does better than HT.

(BTW when i say CPUcount=8 or CPUcount=16, i am just reporting what i see in the RPT file. I am not setting cpucount parameter manually ever).

I did not change anything with affinity.

Using an alternative memory allocator like mimalloc will gain you much more benefits than trying to control the affinity and stuff in my opinion. see: https://github.com/GoldJohnKing/mimalloc/releases

[whole cloud]

Vanilla? Mods? Reproducible?

[whole cloud]

Switching between CPU's is not supported by us.
The affinity affects the "primary cpu" so if you use some external tool, to move the process to a different cpu before it starts, then the affinity would affect there

Oh you mean which core.
We don't do thread pinning as I wrote above

[whole cloud]

The game cannot see externally applied affinity and the native Parameter doesn't exist till next update

I can add you a parameter to pin main thread though.
We at some point used to pin the main thread to the core that it started on. The code is there

[scarlet jolt]

modded, tried to follow my steps but no repro

[thin wyvern]

#perf_prof_branch message

[empty goblet]

that one would be useful

[whole cloud]

-cpuMainThreadAffinity=0xFF
done

[rustic swallow]

I can't log into the game with profiling. I click start the game but it doesn't start what's wrong?

[whole cloud]

More details, what exactly happens

checked RPT files yet?

[rustic swallow]

yes

nothing happen

at the same time, the stable version logs in

[whole cloud]

how did you install profiling? steam branch?

[rustic swallow]

yes

[whole cloud]

Is any RPT file being created?
Starting via Arma launcher, it doesn't open a popup saying it crashed?

[rustic swallow]

Let me try to record on video what's going on

I think it will be clearer and easier this way.

https://youtu.be/YfRpJJ4wE8Y

[restive pilot]

try unchecking cpuCount and exThreads

[rustic swallow]

It didn't help 😦

[rustic swallow]

IT HELPED! Thanks, I just had to wait.

[rustic swallow]

my fps in YAAB is 51.5 without mods and 23 with mods. nice👍

[quaint flame]

Wth are these mods

[silk pewter]

20FPS mod

[quaint flame]

I mean. What mod runs so intensively during yaab. We should find that mod and bann it

[fading seal]

Well, I gave it a try and virtually no difference 😢 (on a 2700x anyway)

[real creek]

Webknight's FPS reducer™️ (ACE3 compat)

[fading seal]

also, profiling does give a significant-ish boost in FPS even on older hardware (about 9-ish FPS more), in case anyone was wondering 😛

[fast hornet]

btw. that board you're running the 2700X on most likely also supports a 5700X3D. That would be a massive upgrade for relatively low cost 🙂

[fading seal]

Yeah I know but it's not the right time 😜

[woven loom]

Also use mimalloc if you aren't already

[woven loom]

This thing

[fading seal]

I'll give it a try some time. Worth a shot I guess

[weak radish]

Does minalloc still provide as big a boost compared to the current Intel allocator?

[heavy vortex]

Probably not, because perf doesn't do as many allocations as it used to.

[woven loom]

I didn't try with the latest binaries but up until two months ago it still did

It used to be that the lock_pages variant provided the most benefit but now that's default behaviour

Like for me it was a good 10% (~10 fps)

[heavy vortex]

IIRC there was also a flaw in some mimalloc versions that was fixed recently, so that would compensate.