The Current Windows 10 vs. Linux Browser Performance For Google Chrome + Mozilla Firefox

[lucasbekker]

this is one of the threads that make me ask myself:

1. I have better battery life in windows
2. I have games, ms office, photoshop and sony vegas on windows
3. browsers are faster on windows
4. My nvidia card works better, no CUDA, gcc and lower level compatibility issues

then I ask myself why am I here?

then I answer myself
1. In linux I can customize my desktop environment with convenient shortcuts
2. code compiles faster
3. I have more free RAM available
4. bash

Hope I don't ask myself for much longer..

I found that CUDA development is less painfull on linux than on windows, would you care to elaborate?

[carewolf]

Did you use official Google Chrome, or distro shipped Chromium?

[Michael]

it's so sad that i am on openSUSE Tumbleweed and it is the slowest...
Would be interesting to know how the openSUSE's Leap version compares with the rest...

Is there a script that will run automatically and make all the things necessary to benchmark my current system?

Thanks.

sudo zypper install php-cli php-xml
git clone https://github.com/phoronix-test-sui...test-suite.git
cd phoronix-test-suite
./phoronix-test-suite benchmark selenium

[xorbe]

I tried to bring Tumbleweed's perf to the attention of the community, but everyone turns a blind eye, and even rbrown of opensuse calls this site trash. I don't know what's going on. I have no problems with my installation (though I run ext4 and turn off mitigations), but I'd simply like to know the root cause of the perf delta. Default file system? Kernel? Boot options? Compiler flags? Cpu governor? Firefox options?

Also, it would help to know if this was on an Intel or AMD system.

[bobbie424242]

sudo zypper install php-cli php-xml
git clone https://github.com/phoronix-test-sui...test-suite.git
cd phoronix-test-suite
./phoronix-test-suite benchmark selenium

package php-zip is also needed.

[bobbie424242]

I tried to bring Tumbleweed's perf to the attention of the community, but everyone turns a blind eye, and even rbrown of opensuse calls this site trash. I don't know what's going on. I have no problems with my installation (though I run ext4 and turn off mitigations), but I'd simply like to know the root cause of the perf delta. Default file system? Kernel? Boot options? Compiler flags? Cpu governor? Firefox options?

Also, it would help to know if this was on an Intel or AMD system.

It's intel. From the first page

The same system was used for testing throughout (obviously) and included the Intel Core i9 9900K, ASUS PRIME Z390-A, 2 x 8GB DDR4-3000 memory, Samsung 970 EVO 250GB NVMe SSD, and Radeon RX Vega 56 graphics

[JeansenVaars]

I found that CUDA development is less painfull on linux than on windows, would you care to elaborate?

that is a good point of attention actually. The thing is, when you bring it down to what xorbe said, try combining it to Tumbleweed / hybrid hardware.

in Fedora you have RMPFusion, Negativ0 and Nvidia self's packages, which is a trial and error use case until you find the one. Then depending on the implementation, tensorflow execution, Blender or whatever, libraries are looked for in different places, such as /usr/lib/ /usr/lib64/ /usr/local/lib etc so you start softlinking stuff everywhere, and then when upgrading... good luck with that. Not considering if you want to rollback a kernel version, and see if akmod is doing its job or not (although again, I am on Fedora)

You also have the mess of cuda repos, xorg-cuda, xorg-nvidia, and combined with tumbleweed you have another trial and error process across using optirun, primusrun, or a bridge of optirun to primus with serious performance loss. Each programs needs its own LD_LIBRARY_PATH linking so you need to figure out options, plus the NVIDIA GPU options to make it right.

Its a joy of a ride, but not an easy one. Documentation tends to be outdated and does not match a universal distribution default.

Tensorflow particularly is a good ride to try on windows, as support is usually very limited, buggy and non-supported properly. I can go on, but its off topic here.

[mlau]

Aside from SeleniumBenchmark: MotionMark, Windows wipes the floor with Linux.

It means the browser binaries have been best optimized for Windows.

[AndyChow]

that is a good point of attention actually. The thing is, when you bring it down to what xorbe said, try combining it to Tumbleweed / hybrid hardware.

in Fedora you have RMPFusion, Negativ0 and Nvidia self's packages, which is a trial and error use case until you find the one. Then depending on the implementation, tensorflow execution, Blender or whatever, libraries are looked for in different places, such as /usr/lib/ /usr/lib64/ /usr/local/lib etc so you start softlinking stuff everywhere, and then when upgrading... good luck with that. Not considering if you want to rollback a kernel version, and see if akmod is doing its job or not (although again, I am on Fedora)

You also have the mess of cuda repos, xorg-cuda, xorg-nvidia, and combined with tumbleweed you have another trial and error process across using optirun, primusrun, or a bridge of optirun to primus with serious performance loss. Each programs needs its own LD_LIBRARY_PATH linking so you need to figure out options, plus the NVIDIA GPU options to make it right.

Its a joy of a ride, but not an easy one. Documentation tends to be outdated and does not match a universal distribution default.

Tensorflow particularly is a good ride to try on windows, as support is usually very limited, buggy and non-supported properly. I can go on, but its off topic here.

I had infinite headaches until I discovered conda (I use the one from anaconda). You can set an environment with python x.x.x, and then the cudnn and pytorch, tensorflow, all of a version you want, for each project you are working on. You also have the nvidia-docker thing if you want to deploy it to other people.

TensorRT is a linux exclusive, also. And all of those things are supposed to be eventually consolidated in CUDA-X. For example, with RAPIDS, it's as simple as, in a new conda environment,

conda install -c nvidia/label/cuda10.0 -c rapidsai/label/cuda10.0 -c pytorch -c numba -c conda-forge cudf=0.6 cuml=0.6 python=3.7

Them if you realise "Oh no, this project only works on cuda 9, python 3.6, and I don't have time to debug why", then you deactivate that environment, make a new one, and

conda install -c nvidia -c rapidsai -c pytorch -c numba -c conda-forge cudf=0.6 cuml=0.6 python=3.6

and it's as simple as that. No softlinking, no mess on your base environment, no rollback headaches.

[Vistaus]

Michael
Nice tests. I wonder how much of the performance difference is due to Spectre/Meltdown/etc. mitigation.
Do you also have power consumption/battery life figures?
Also I think ChromeOS results would be interesting here. Do you still own the Acer Chromebook C720?

But how would he be able to install PTS on ChromeOS when the C720 doesn't have support for Linux apps?

https://www.howtogeek.com/394049/some-chromebooks-wont-get-linux-apps-but-heres-what-you-can-do-instead/