Share your Benchmarks!

raptor · May 4, 2024, 10:46am

Use this Topic to share your benchmarks of different kernel schedulers and optimizations.

Please include the following information in addition to your results:

Specs
Kernel (uname -a) & personal kernel modifications if any
Scheduler
Benchmark type (game or benchmark program e.g mini-benchmarker)
Display system (X11 or Wayland) and specific WM/DE
BIOS modifications
Power governor

Wyn · May 4, 2024, 9:35pm

Any recommended benchmarking utilities?

ptr1337 · May 5, 2024, 6:31am

–mini mode of mini-benchmarker:
https://aur.archlinux.org/packages/mini-benchmarker

to visulize the results:

github.com

julmajustus/mini-benchmark-scraper/blob/main/benchmark_scraper.py

import os
import re
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.colors as mcolors
from collections import defaultdict

# Function to parse log files and extract test data, system information, and kernel versions
def parse_log_files():
    test_data = defaultdict(list)
    kernel_info = defaultdict(dict)
    kernel_versions = defaultdict(dict)

    for file in os.listdir('.'):
        if file.endswith('.log') and file.startswith('benchie_'):
            with open(file, 'r') as f:
                data_text = f.read()

            kernel_version_match = re.search(r'Kernel: (\S+)', data_text)
            if kernel_version_match:

This file has been truncated. show original

ptr1337 · May 5, 2024, 7:32am

Lets start.
Here you can find scx_rusty against EEVDF-BORE on the 6.9 Kernel.
scx_rusty was build from commit Merge pull request #261 from sched-ext/rusty_interactive · sched-ext/scx@efb97de · GitHub

tldr:
Rusty does really compete with EEVDF-BORE. The perf sched regression is the only “bigger” one, but all in all it is looking really good.
dvernet is working to fix the perf sched regression.

Rusty is also since the above commits very responsive under heavy load and rusty gets slowly to the path for a very good scheduler.

YUBY · May 5, 2024, 12:28pm

I made a little comparison of rusty, lavd and bore in Rocket League and spotted two differences.

AMD Ryzen 5700X + Nvidia RTX 3060Ti + 16GB RAM 3200MHz
6.9.0-rc6-3-cachyos-rc + scx-scheds-git 0.1.9.r40.g07b521b-1
amd_pstate=active mitigations=off
X11 + KDE
PBO and c-states disabled
performance governor

FPS stability.

BORE can maintain 359.XX FPS (with cap to 360 FPS in game.) but lavd and rusty constantly droping to 355 on rusty and even 340 on lavd.

Latency.

Game Thread latency on BORE is around 0.60, on rusty around 0.80 and lavd around 0.90.

Moreover Game Thread MAX latency is spiking in moment when screenshot is taken so this value is not credible on screenshots.

Link to explanation what this statistics exactly mean.

https://www.epicgames.com/help/en-US/c-Category_RocketLeague/c-RocketLeague_TechnicalSupport/rocket-league-performance-and-network-graphs-pc-only-a000084302

The_Grongle · May 5, 2024, 1:40pm

Kernel: 6.9.0-rc6-4-cachyos-rc
CPU: AMD Ryzen 7 5800X
SCX: 0.1.9.r40.g07b521b-1

From top to bottom:
eevdf+bore
scx_lavd
scx_rusty

YUBY · May 6, 2024, 5:49am

Game:
Cyberpunk 2077 v2.12
Game settings:
Preset Ultra with DLSS OFF and RT OFF
Hardware:
AMD Ryzen 7 5700X + Nvidia RTX 3060Ti + 16GB RAM 3200MHz
Kernels:
linux-cachyos-rc 6.9.rc6-4 + scx-scheds-git 0.1.9.r40.g07b521b-1
linux-cachyos-rc-lto 6.9.rc6-4
linux-cachyos-echo 6.8.9-1
linux-cachyos 6.8.9-3
amd_pstate=active mitigations=off
DE:
X11 / Wayland (once) + KDE
BIOS:
PBO and C-States disabled
Governor:
performance

Long story short - no significant differences. With each configuration I did two runs but I add only one screenshot with better result. Additionally I tested SMT ON/OFF/Auto Reflex ON/OFF and VSync ON/OFF but there was no differences in scores. Number after scheduler name is benchmark run order. Originally I planned add all screenshots.

The_Grongle · May 6, 2024, 9:21am

The important question is how cpu limited are you?

YUBY · May 6, 2024, 10:06am

Game:
Cyberpunk 2077 v2.12
Game settings:
Graphic settings lowest possible with DLSS Ultra Performance
Hardware:
AMD Ryzen 7 5700X + Nvidia RTX 3060Ti + 16GB RAM 3200MHz
Kernels:
linux-cachyos-rc 6.9.rc6-4 + scx-scheds-git 0.1.9.r40.g07b521b-1
linux-cachyos-rc-lto 6.9.rc6-4
linux-cachyos-echo 6.8.9-1
linux-cachyos 6.8.9-3
amd_pstate=active mitigations=off
DE:
X11 / Wayland (once) + KDE
BIOS:
PBO and C-States disabled
Governor:
performance

This time I tried to reduce GPU bottleneck. Now we have some differences so I add all screenshots. I did all benchmarks at one fell swoop so I run one more BORE extra to see how performance can be degraded when CPU is heated. Exactly order is BORE 1&2 next Rusty 1&2 next Lavd 1&2 then BORE 3. The remaining tests required restarts so heat was lower.

YUBY · May 6, 2024, 10:07am

YUBY · May 6, 2024, 3:34pm

New kernel linux-cachyos-rc 6.9.rc7-1 and mitigations ON / OFF

The_Grongle · May 6, 2024, 7:30pm

i think a lot of space could be saved if perhaps we just posted the fps results in text lol

bigvnd · May 9, 2024, 6:40pm

Hi, I’m new to cachyos.

I tried cachyos-rc kernel, the results are more details on graphics, less stutter, better brightness but less fps (1-2 fps) than cachyos non rc kenel.

The_Grongle · May 10, 2024, 12:17pm

None of that sounds right. Im sorry to say this but if youre as new as you sound. Best you dont post here. go ask the discord for help

The_Grongle · May 10, 2024, 12:18pm

Anyway. Looks like having SMT off on cyberpunk for my 5800x very much improves fps

bigvnd · May 11, 2024, 1:21am

I have no problem on fps. everything else is great even better than any other distros. Cachy Os is the best right now

Visentinel · June 22, 2024, 3:58pm

I was interested in how well scx_rusty handles background heavy loads so i ran some memory latency benchmarks while peazip is smashing away on 8 threads on my 5800x3d and to my pleasent surprise the latency was only slightly higher at the 262144kb page size from 56-58ns to 66-68ns very impressive.

nipnipj · July 13, 2024, 1:35pm

Specs: i5 10600k + rtx 2060 + 16GB
Kernel: Linux 6.9.9-2-cachyos
Scheduler: some of the currently available in scx-scheds-git
Game: Isonzo
Display system: KDE + x11 + compositor enabled
BIOS modifications: none
Power governor: default (game-performance doesn’t work for me).

FPS distribution:

FT distribution:

Note: in-game bots might have behaved a bit different on each test.

Conclusion: Bpfland seems to win in location (not so far from bore) and dispersion in both FPS and FT.