Subject: Re: [linux-audio-dev] Re: Low Latency Kernel Combos
From: Josh Green (jgreen_AT_users.sourceforge.net)
Date: Sat Jun 02 2001 - 02:53:23 EEST
Benno Senoner wrote:
>
> Indeed the unpatched 2.4.5 is performing very well.
> As for the 2.4.4-lowlatency: have you tried 3x128 buffers ?
> If you have time can you post them too ?
Yes. I will try 3x128 buffers and post results (its Linux 2.4.5
actually, the 2.4.4 low latency patch was used though).
> 3.8ms is good but it would be better being able to deliver rock-solid sub-3ms
> latencies.
> See my old 2.2.x test .. 2.1ms rock solid ... 2.4 should be able to do this
> too.
>
Cool, I'm assuming trying 3x128 buffers with latencytest just modifies
the buffer and won't increase performance?
> For example the phpnuke portal would be ideal to announch test results like
> yours ecc.
Yes, I was also thinking that it would be neat to have some sort of
Linux lowlatency database that people can submit their results to or
query. An html form with things like Kernel version, patches that were
applied, hardware info (CPU, Chipset, Hard disk type IDE/SCSI, video
card, sound card, etc), along with the results of latencytest (perhaps
have an easy way to plug in the important numerical results rather than
the entire graphs?). This would help those who have dissapointing
results figure out what the problem could be (a specific driver,
hardware, etc), and also would help those who are looking for nice Linux
hardware/software combinations.
> (My friend Udo has already agreed that he will produce RPM packages of new
> lowlat kernels as soon as they will be available and that he will publish them
> on the phpnuke portal. This is one of the reason why I'm advocating this
> "community driven" portal)
>
Yeah, that would make the lowlatency patched kernels more accessible to
people. Perhaps someday a lowlatency patch won't be necessary :)
>
> Yes in order to set SCHED_FIFO you do need root privileges, but perhaps future
> kernels (the famous capabilities concept) will allow some users/apps to set
> this without the need of being root.
Seems like it should be some user accessable service (something in
/dev?) that allows higher priority scheduling of programs. It would also
be nice if it didn't have the side effects of SCHED_FIFO (100% CPU with
SCHED_FIFO thread == locked up box).
> Non SCHED_FIFO programs take advantage of a lowlat kernel as well (much lower
> latency peaks) but the peaks will still be much higher than using SCHED_FIFO
> scheduling especially when the machine is heavily loaded.
>
Thanks for clearing that up :) It would be nice to test the latency of
non SCHED_FIFO programs too. I'll have a look at your latencytest to see
if I can figure this out. Do you already have a switch for this or
should I just comment out the sched_setscheduler lines? Lates..
Josh Green
This archive was generated by hypermail 2b28 : Sat Jun 02 2001 - 05:13:55 EEST