CS2 Input Lag Explained - The Monolithic 64 Hz Client Simulation Tick Stalls the Render Thread

[cuaderno]

This is on pracc DM server, fps_max 360, 9800x3d, 7900XTX

Thanks for the feedback!

I have some questions:

• Is HAGS on? If not - PresentMon uses legacy kernel scheduler which is not accurate.
• Is Antilag2 on? You might have prerendered frames in the queue which cause input lag but smooth out the graph.
• Are you on vulkan or dx11?

Hey,

1. Hags OFF
2. Antilag2 ON
3. dx11


Tested again on pracc DM with Hags ON (fps_max 360), no difference


Also tested on the awful Valve DM (fps_max 360, Hags ON), which detected the spikes


And empty offline server, spiked again



The feeling of playing on 3rd DM server is noticeably better despite same fps cap value, which matches the findings. Next I can test on official Valve Premier, which never felt as bad as Valve DM server.

[vnb]

This is on pracc DM server, fps_max 360, 9800x3d, 7900XTX

Thanks for the feedback!

I have some questions:

• Is HAGS on? If not - PresentMon uses legacy kernel scheduler which is not accurate.
• Is Antilag2 on? You might have prerendered frames in the queue which cause input lag but smooth out the graph.
• Are you on vulkan or dx11?

Hey,

1. Hags OFF
2. Antilag2 ON
3. dx11


Tested again on pracc DM with Hags ON (fps_max 360), no difference


Also tested on the awful Valve DM (fps_max 360, Hags ON), which detected the spikes


And empty offline server, spiked again



The feeling of playing on 3rd DM server is noticeably better despite same fps cap value, which matches the findings. Next I can test on official Valve Premier, which never felt as bad as Valve DM server.

very interesting. So this confirm what i've been saying for months here...

Community servers have different networks settings (packets pacing, < MTU, margins, and others) which also affects the gameplay and how it feels overall.

Also your results are much cleaner than OP because the 7900 XTX has a hardware scheduler while Nvidia GPUs have heavy reliance on software drivers scheduler, which naturally create a CPU overhead and less clean results.

[Babluga]

what do you think about this? CS2 DM server 60 sec

360 FPS cap


Uncap

[cuaderno]

Ok, so I did random testing of different variables to see if anything stood out.

Using AMD Chill and RTSS async made the spikes disappear in the capture tool, but this is not fixing the root issue.

Here are other variables that I tried using and made no difference (ie, the spikes were present)
Launch options: noantilag, insecure
Servers: gamersclub dm and warmupservers dm (they had noticeably less spikes than Valve dm servers, but more spikes than pracc dm servers)
VSYNC On or Off, Freesync On or Off
High ping and low ping connections
Vulkan or dx11
Valve Premiere and Rush servers

[soul4kills]

Looks like it's a side affect of gsync+vsync enabled.

edit: added 3 separate readings playing a cs server. focused on the peak to peak distribution reading as that's what matters. I'm not getting that 64hz peak to peak interval that's being discussed.

[cuaderno]

Looks like it's a side affect of gsync+vsync enabled.

What refresh rate and what cap value did you use?

The spikes show up a lot less on the capture when I cap low, such as fps_max 324 at a 360hz display, even with sync off.

The odd behavior is that despite using fps_max 324, the in-game cap never goes higher than 315 fps. If I used at higher number, like fps_max 360, it would stay steady at 360 fps but the spikes appeared.

[unconnected]

Looks like it's a side affect of gsync+vsync enabled.

I believe that's the G-Sync + V-Sync pacing buffer absorbing alternating ticks.

As for the tests, I'd leave FPS limiters off because they just mask the problem and they make no sense with Reflex and Anti-Lag technology in place.

What do you guys think about the findings so far?
On all uncapped runs (both AMD and Nvidia), every frametime spike lines up directly with a spike in MsGPUWait, while MsGPUBusy stays flat. To me this looks like the render queue is getting blocked on the engine side. Does anyone see any other explanation for this or does this basically confirm it? Would love to hear your feedback.

[cuaderno]

Agree that limiters just hide the problem because finding a cap value that make the spikes disappear come with input lag addition.

The info we are missing is what the community servers DM are doing differently than offline / valve servers. Another consideration is that the game plots a VProf telemetry after every match that shows a P1 number that discard the effect of the spiked frames, which seems deliberate by Valve.

[soul4kills]

It did finally rear it's head. But, it's not a very consistent. My frame rate is 138. I have gsync and vsync on with reflex at ultra.

These are all tests on valve servers. I tried it on a custom surf server. It would also be hit or miss of the pattern showing up. The only consistency of the pattern is it stops/appears when the server changes. May not be a valve vs non valve server. Do servers have a syncing routine that happens during connect? Is it possible during connect the server has a bad sync to the client? But even during the times when the pattern did show, I didn't experience any input lag.

Can't you set up your own server and try different configurations and see what setup results in the pattern showing up.

[Moontrance]

what the community servers DM are doing differently than offline / valve servers

When we run offline map we're starting local server, so PC resources are shared between client/server, tick happens at nearly the same time and spikes appear more than online.
Community DMs are tricky. Cybershoke, for example, provide custom wallhack protection, it's not sending enemy information until it's in your POV or will be in your POV in a couple ticks, so you can join 24 slot server and if no one near you, you won't even hear shooting sounds and stuff, server will seem empty from you client perspective, so no drop due to heavy client simulation happens. Valve doesn't do anything similar and sends you data about all players on a server AFAIK, that's why Valve DM lags so much and that's why competitive/premier feel better, they're just less crowded, so client processing is easier.