Linux latency measurements and compositor tuning

When I switched to Linux from Windows last year I noticed I had a lot of keyboard input latency only when gaming. It was like I had ~100ms of input latency and it felt exactly the same as playing old school Quake I on a dial up connection without client side prediction turned on. It's like you're skating on ice with a delay when seeing the output vs when you performed the input. This was despite having a solid 60 FPS.

Turns out it was due to a combination of things.

I was using niri (Wayland window compositor) and this input latency was present with or without v-sync turned on. It happened when I was using a 60hz 4k monitor with an NVIDIA GPU.

Then I tried playing the same game on a laptop (same distro and dotfiles) with an AMD GPU and no external monitor. The delay disappeared.

Then I played the same game on that laptop but hooked it up to the 4k monitor and I had the same keyboard input latency only when v-sync was enabled. When I turned off v-sync and capped my FPS then the input latency was reduced by an amount that I could no longer perceive the delay.

Then I put an AMD GPU in the original desktop I was testing and reproduced the same results as the laptop.

However, when I switched to using KDE Plasma with X or Wayland, the keyboard input latency disappeared. This was with the 4k monitor and both NVIDIA and AMD GPUs.

I reported it to niri but it hasn't gotten traction, I just know I can reproduce it on 2 completely different systems with different GPU vendors and hardware when the common ground is having a 4k 60hz monitor hooked up.

> I found it hard to believe we couldn’t get a wakeup timer more granular than 1 ms, so I looked at what KWin was using. Indeed, it was passing the sleep duration in milliseconds to a QBasicTimer.

I fixed the same thing in GNOME a few years ago across GLib, GTK, and Mutter/GNOME Shell. It required getting glib onto ppoll() finally.

No X11 because 'KDE drops it' is a rather silly argument for not including it, especially given that X11 is likely to trump Wayland-based systems on this metric. I'm using X11 and will continue to do so until something better comes along. Thus far Wayland is not better, just different and in several ways less functional. Yes, I do use X11 network transparency, sometimes through X2go, other times directly.

> Originally, I had meant to compare with X11 sessions as well, but with KDE removing them soon, I dropped it.

Last time i checked X11 without a desktop compositor (very important) still provides the least latency of any environment, so even if KDE Plasma (not KDE as a whole AFAICT) drops support, there are other WMs/DEs that will work with it and have minimal latency on desktop.

> While randomly closing apps, I found the culprit: the Zed editor. Apparently, an open Zed window can add latency to all my other apps even while idle in the background.

Zed definitely does funny things to KWin. It's not the only app that does, but this point in particular would be worth more investigation. I've noticed it causing weird issues with the frame pacing as well sometimes.

> There it was, something about my desktop profile was introducing at least 3 ms of latency [as compared to creating a fresh account]! From here, I tried a bunch of things: plasma-manager to diff my existing profile against a clean one, removing all virtual desktops and disabling all KWin effects and any display scaling. While randomly closing apps, I found the culprit: the Zed editor. Apparently, an open Zed window can add latency to all my other apps even while idle in the background.

Things like this are so maddening. I don't worry too much about performance on Linux, reserving a Windows machine when I want full hardware acceleration and optimization.

It would be interesting to know if the same is true for Intel/AMD. I have one NVIDIA PC and the desktop - not even Games - feels very sluggish on GNOME - I haven't tried KDE yet. On the contrary, my AMD gaming rig subjectively feels very snappy, but I have never measured it.

I don't know the whole tricks and ways that use steamos but it gives very good quality for it's hardware.

I wonder what is Con Kolivas up to these days, he was THE og Linux ui latency guy.

Mouse pointer movement (rather than click) to display update would be amazing to get too… including X11…

duplicate, already posted: https://news.ycombinator.com/item?id=48476462

I think this is only the top of the iceberg.

I say that because I play some timing tight action games natively on elf(glibc)/linux. Let's take an example with one of the toughest: Silksong.

I was trying to beat lost lace, her timings were too tight and nearly at each try I was locked in some hardly humanly dodge-able pattern combination. I knew all of her patterns by heart after zillions of tries.

Then, I started to have strong suspicions: I closed all background apps, disconnected the network and try again... did beat her, first try, super ez, like she was transparent to me and slow, I could "read" her and react in time easily.

I am running a Zen2, 12 cores at ~4GHz... and native x11 with xorg, sooo... the main culprit seems to lie in game engine programming and then would not completely tied to wayland programming (don't worry, I am coding my own wayland compositor, so, I am going to move to wayland, well actually designing a 'binary layout' for a wayland compositor to be accurate).

I am now in a Silksong steel soul run, and you bet I'll keep this experience in mind, because when I watch video streams of other people fighting some bosses, I can cleary "read" their moves like it is "slower" and which seems much less "aggressive", but once I fight them on my system, nope, I get a much harder time at reading the boss patterns. The "closing all apps and disconnecting the network cable" did not change a thing here, because I am currently fighting "ez' bosses then I always manages to get rid of them before I get really used to their patterns again... we will see with later and harder bosses.

But this could be another [obvious] culprit: stress. I know I am very, VERY, sensitive to stress: it disrupts severely my mind and worse with very little of it. In other words, I would have "brain fog" while fighting a boss because of stress, and the time I did beat lost lace in one shot ez: I was "testing" something without the stress instead of actually trying to beat her... a abysmal difference.

The idle Zed culprit fits its GPUI renderer, which has a history of repainting continuously rather than on demand, so the compositor never reaches its idle path and other clients inherit the wakeup cost. Confirming whether the backgrounded Zed window still holds a live frame callback loop would nail it, since that alone keeps KWin out of its low power present timing.

[dead]

These vertical labels make me unreasonably mad.

GE 10.34 released March 23th, 5 days after 10.33 - any reason to test with 10.33 in June? Was there a regression?