A New BFS "Smoking" Scheduler For Linux 3.3

[Paradox Uncreated]

Good for you. That does not mean that all workloads will have a max latency of 0.33ms.

By the way, I thought the kernel timer could be set to 1000Hz at most, so I don't understand sub 1ms latencies.

Most workloads will.
Btw, some OS`s boost threads. Threads that can occupy some time, but seconds sounds quite odd though.
Check if there are any threads set to realtime etc.

I am just going to ignore your comment on 1000hz.

Peace

[jwilliams]

Most workloads will.

Which is precisely the problem with your metric. Most is not all. Average is not worst case.

[RealNC]

Good for you. That does not mean that all workloads will have a max latency of 0.33ms.

By the way, I thought the kernel timer could be set to 1000Hz at most, so I don't understand sub 1ms latencies.

0.33ms doesn't make sense to begin with. You can configure JACK in a way that is says "0.33 ms". But that's just a calculation arising from the chosen FP, PB and sample rate you set in the settings. Just because the settings would want 0.33ms doesn't mean the kernel is gonna deliver.

[RealNC]

I'm not sure you intended it, but this post sounds very racist to me. You're just making yourself look bad right now.

It doesn't look racist to me. Just pretentious, which was the point.

[Paradox Uncreated]

Lol. Not Average. FYI the kernel can go to ca 4000hz. And ofcourse it can easily be extended. Con has also done that. But when you want to do that nowadays I am not sure. Turn on tickless and be happy.

This is a kernelconfig for 0.33 ms latency. http://www.paradoxuncreated.com/tmp/.config39

Tested to work with intel dual-core, decent mobo, and firewire audio-card. Professional sound in other words.
And Nvidia gfx.. All the nice h/w one would want.

If you have lame-ass mobochip sound, that doesn`t seem to work with that low latencies. Maybe lazy policies in the driverwriting department, since it has been reduced from almost working at 1ms to, to almost working at 0.5, where before it did not work at all at 0.5ms.

Peace.

[RealNC]

And how do you know you're getting 0.33ms? I can setup JACK for 0.167ms with BFS (16 FP, 192kHz, 2 P/B) and it works without a glitch. But it does not mean that this is the latency you actually get!

(And this would be ridiculous anyway; 8ms is more than enough for any audio work.)

Edit:
You remind me of this guy: http://www.youtube.com/watch?v=foS84bEZSO4

[Paradox Uncreated]

And how do you know you're getting 0.33ms? I can setup JACK for 0.167ms with BFS (16 FP, 192kHz, 2 P/B) and it works without a glitch. But it does not mean that this is the latency you actually get!

(And this would be ridiculous anyway; 8ms is more than enough for any audio work.)

Edit:
You remind me of this guy: http://www.youtube.com/watch?v=foS84bEZSO4

You obviously have an attiude problem RealNC. Ali G is you, not me.
I doubt pursuing explaining 0.33 ms to you, is possible.

[RealNC]

You obviously have an attiude problem RealNC. Ali G is you, not me.

Peace.

I doubt pursuing explaining 0.33 ms to you, is possible.

Well, not only to me; others also wonder what you mean with "0.33ms latency".

[Paradox Uncreated]

Peace.


Well, not only to me; others also wonder what you mean with "0.33ms latency".

Others? Well I just edited the code, and hey presto there it said 0.033 metres of snow. If you thought about buffer that was 32x37 matrixes of evil nerdwannabes, doing gymnastics in series.

I am sure your demented mind will understand.

[ninez]

This is a kernelconfig for 0.33 ms latency. http://www.paradoxuncreated.com/tmp/.config39

Tested to work with intel dual-core, decent mobo, and firewire audio-card. Professional sound in other words.
And Nvidia gfx.. All the nice h/w one would want.

You've posted a config for a 2.6.39 generic kernel;

# Automatically generated make config: don't edit
# Linux/x86_64 2.6.39 Kernel Configuration
# Sat May 21 00:45:30 2011

This generic kernel would not perform as well as the same kernel using the BFS patch set (not on either of my Core2duo's, not on my AMD Phenom II 965 x4, and i doubt it would on my i7 ~ although that machine is newer, and i haven't/won't test it). 2.6.39 was just before linux-rt-3.0 came out. For the entire time between 2.6.33-rt and 3.0.1-rt -> i was using BFS kernels, as i had some hardware that didn't play nice with 2.6.33-rt - and subsequently ended up using BFS on all my systems, because it worked noticably betterthan CFS on the same kernel, during that period. On all machines (using 2.6.39, when it came out) BFS ran smoother / more stable & reliably for proaudio then CFS, and that's with using the 'threaded irq' boot parameter, that had been merged into mainline 2.6.39, from the -rt patches ... I tested this quite a bit at the time.

I currently use RT-kernels, but BFS is definitely a great alternative to CFS for certain workloads. I would likely be using it right now, if RT wasn't an option, or if I thought it was overkill for my usage.

To say, Con is wasting his time - as you've said before, is silly. Lots of people have been using BFS for years, in my own experiences, it was great and in situations where i was getting XRUNS with CFS (on any kernel 2.6.33+) - with BFS, they didn't exist (on several machines). I think both CFS and BFS have value, and it's healthy to have them both around.

my 2 cents