Debian 10 - Vanilla & PREEMPT_RT Kernel Benchmarks

0 Members and 1 Guest are viewing this topic. Read 315 times.

emailtim

Debian 10 - Vanilla & PREEMPT_RT Kernel Benchmarks
« on: 17 Feb 2021, 09:15 am »
It has been a while since I tried music on Linux so I decided to give it another try.

I installed Debian 10 64-bit with vanilla and PREEMPT_RT version 5.9 kernels from buster-backports on an old Intel Atom D525 single board computer and an IvyBridge i7-3770K machine to experiment with interrupt latencies compared to Win7-64.  I also installed the Redhat "tuned-adm" profile package and JRMC 26.

Here are some cyclictest latency result plots that may (or may not) be of interest.

Intel ION D525 PREEMPT_RT kernel, 2 channel 192kHz content playback:
Max latency spikes at 121 microseconds.



Intel i7-3770K Vanilla 5.9 kernel, 2 channel 192kHz content playback.
Note some fliers out to 506 microseconds not present on the D525 RT:



PREEMPT_RT on i7-3770K running 2 channel 192kHz content playback.
Note max latency spikes are down to 11 microseconds from 506.



PREEMPT_RT on i7-3770K running 6 channels of convolution with 256K taps/channel, steep 384kHz XO's with 192kHz content playback.
Note max latency spikes @ 21 microseconds with heavy convolution load.



All of these have lower latencies than Win7-64 (in the millisecond range) on the same i7-3770K machine.

Also discovered that the PREEMPT_RT kernel and NVDIA drivers DO NOT MIX.

garyalex

  • Jr. Member
  • Posts: 132
Re: Debian 10 - Vanilla & PREEMPT_RT Kernel Benchmarks
« Reply #1 on: 17 Feb 2021, 03:44 pm »
Hi Tim,

Thanks for this.  Do you think latency at this level makes an audible difference?

emailtim

Re: Debian 10 - Vanilla & PREEMPT_RT Kernel Benchmarks
« Reply #2 on: 17 Feb 2021, 04:48 pm »
Hi Tim,

Thanks for this.  Do you think latency at this level makes an audible difference?

I will be running those A/B tests next after just completing the configurations and plots.  On the weak Intel ATOM D525, full screen YouTube music videos would lock up the machine on Win7-64 where it actually works on Linux.  That delta maybe just the deltas between Windows and Linux in general.

I am waiting for an XMOS flash for my main system's 8 channel DAC before I can A/B test on my main system.  Initial tests with other DACs on a different system are promising, but I don't have any specifics yet that I can soley attribute to just the lower latencies due to other system changes.

Here is a writeup from Redhat Linux and a different type plot view of their data.

https://access.redhat.com/documentation/en-us/red_hat_enterprise_linux_for_real_time/7/html/installation_guide/chap-why_use_rt_to_optimize_latency




Here is a plot of a Windows system from AudioLinux.  My Win7-64 install is similar to this plot which is significantly higher than 21 microseconds on the same hardware.

http://www.audio-linux.com/


whell

Re: Debian 10 - Vanilla & PREEMPT_RT Kernel Benchmarks
« Reply #3 on: 17 Feb 2021, 11:23 pm »
Just my $0.02:

I don’t think the RT kernel makes much difference, especially in a system where I’m just loading and playing FLAC files.  Stipulated this is not with doing measurements or A / B testing.  Just using Daphile Linux, my system and my ears.  Your mileage may vary, etc. 

RT kernels were targeted at the music production environment, where mixing audio from various sources is where latency issues might matter.  Playing one music file at a time, probably not gonna matter as much. 

emailtim

Re: Debian 10 - Vanilla & PREEMPT_RT Kernel Benchmarks
« Reply #4 on: 18 Feb 2021, 01:18 am »
Just my $0.02:
....  Just using Daphile Linux, my system and my ears.  ...

Never used Daphile before, but see it's 64-bit ISOs are released in both vanilla and RT kernels.  Maybe you should give the realtime version a spin since you are accustomed to the vanilla version and report back.  It appears some people like one over the other on other threads.

I may make a Daphile thumb drive to try without having to load it onto a SSD.

https://www.daphile.com/#download

I read one thread where one person was running the realtime version of Daphile in VirutalBox and didn't like the results.  I think he didn't realize that VirtualBox on a nonrealtime host pretty much defeats the realtime aspects of the client OS inside of VirtualBox. =)

FWIW, this guy's review is a few years old, but he seems to prefer the RT version of Daphile over the stock version.  Similar feedback on a few other forums.

http://www.modelpromo.nl/Audio-Dandy_streamer_ROON2.htm

whell

Re: Debian 10 - Vanilla & PREEMPT_RT Kernel Benchmarks
« Reply #5 on: 18 Feb 2021, 04:19 am »
In fact, comparing the two versions of Daphile is what informs the post above. 

emailtim

Re: Debian 10 - Vanilla & PREEMPT_RT Kernel Benchmarks
« Reply #6 on: 18 Feb 2021, 04:44 am »
In fact, comparing the two versions of Daphile is what informs the post above.

Could you clarify your last post or expand upon it? 

I don't know what your response means, especially in light of your original post, implying you have not tried the RT version of Daphile.

"... Stipulated this is not with doing measurements or A / B testing.  ..."

whell

Re: Debian 10 - Vanilla & PREEMPT_RT Kernel Benchmarks
« Reply #7 on: 18 Feb 2021, 05:48 pm »
Probably wasn't clear in my post.   Didn’t say I hadn’t used it.  I did say that I didn’t hear a difference Daphile RT versus non RT kernel.  That's running the system with RT kernel Daphile on a USB stick, then rebooting and running Daphile standard kernel via USB stick.

emailtim

Re: Debian 10 - Vanilla & PREEMPT_RT Kernel Benchmarks
« Reply #8 on: 18 Feb 2021, 07:33 pm »
Thanks for the clarification.