Subject: [linux-audio-dev] And kernel 2.4 ? Was: Re: RFD: Audio Needs for Kernel 2.5+
From: Benno Senoner (sbenno_AT_gardena.net)
Date: Tue Jun 20 2000 - 20:22:09 EEST
On Tue, 20 Jun 2000, Paul Barton-Davis wrote:
> >i don't know too much about the internals of kernel development, and
> >it might well be that the lowlatency patches have become obsolete by
> >now, but i'd like to make sure.
>
> to a large extent this is true. its not 100%, but the extent of Ingo's
> work would be dramatically reduced for a low latency patch for 2.4
> compared to 2.2. I have used 2.3.51, for example, with 3ms latency,
> and got results almost as good as 2.2.10+LL.
I tested some of the 2.3.x on my boxes, although better than plain 2.2.x,
it's FAR away from 2.2.x + lowlatency.
I believe only a few routines were optimized in 2.3.x like making some
allocation/freeing routines (inodes?) O(1) instead O(RAM size).
But the rest like the conditional rescheduling and the various preemption
points are completely missing.
Linus and Alan do not seem interested too much in low-latency,
And that is bad IMHO.
They care only about RAW performance, but throughput isn't everything:
in some cases low-latency can improve throughput:
Consider the fact that low-latency would be even useful for a beowulf
cluster: assume you run a cluster consiting of 100 nodes, and due to
disk I/O a random machine (on average) every 50 secs stalls for 100msec.
That means since you are using MPI and barriers, the other 99 machines
have to wait these 100msec as well.
A a bit of math: assume a stall occurs every 50sec on every machine,
assume the uniform distribution: that means a stall every 50/100 = 0.5sec.
Therefore 100msec/500msec = 20% ,
YES, if you are unlucky a standard kernel could be up to 20% slower
than a low-latency kernel in the case of heavy disk I/O.
Judging from the number of visits my crappy , out-of-date page had,
and about 20-30 mails saying "this is exactly the stuff I looked for,
when will this stuff get into the mainstream kernel ?"
there SEEMS SOME people interested.
I am not a kernel programmer but looking at the patch,
I think that Ingo's modifications do not make a radical change within
the kernel.
There are still issues with SMP in the 2.2.x patch, but sometime
ago Ingo said for 2.3.x it would be much easier to solve.
I think low-latency patches included officially in 2.6 would be WAAY too late.
As Dave said we need "out-of-the-box" low-latency support,
and that means it has to go into the official kernel or all distro manufacturers
have to include it. (but the latter would be MUCH harder to achieve)
Why should we wait 1-2 years (or more) for 2.6 when the technical
issues are already solved ?
The folks involved in linux on the desktop should step in too
(GNOME, KDE folks, Corel etc),
A desktop is not only for using a wordprocessor and a spreadsheet,
but watching videos, or playing with audio software / games too.
My proposal would be that the linux-audio community (we) should write
a letter to the linux-kernel folks ( CCing Linus,Alan,Ingo,SCT,ecc)
and explain the audio issue in detail asking for official statements.
Another proposal would be to make a sort of "low-latency/audio" petition
and submit it to the leading linux news sites.
That could generate some buzz and bring us some supporters.
other ideas ?
> >
> >(would be best if someone who already hangs out on lkml take over
> >this job [pbd ?])
>
> I'm there, all the time, but until the VM guys get close to fixing up
> the VM/cache system so that streaming I/O is close to usable again,
> I'm keeping my mouth shut for now. 2.3.51 could record 24 tracks of 32
> bit audio to a single disk without breaking a sweat. the stupidly
> named 2.4.0-test1-ac* series can't do more than 6-8 without falling
> over itself.
hehe Paul, pretty explicit language, I like it. :-)
Anyway I agree, with 2.3.99+ we are below 2.0.x performance level
on certain areas. :-(
Benno.
This archive was generated by hypermail 2b28 : Tue Jun 20 2000 - 21:48:56 EEST