Subject: Re: [linux-audio-dev] low-latency kernel
From: Christopher Lee (clee_AT_gnwy100.wuh.wustl.edu)
Date: Mon Feb 12 2001 - 16:44:35 EET
Date: Mon, 12 Feb 2001 08:39:14 -0600 (CST)
>>>>> "Andrew" == Andrew Morton <andrewm_AT_uow.edu.au> writes:
>> I recently read about Andrew Morton's talk at linuxconf.au and
>> downloaded his low-latency kernel patches. (I'm currently running
>> 2.4.1-lowlatency SMP w/ reiserfs on ATA-33 drives and still see
>> occasional spikes in "latencytest-0.42" crossing the 2.9ms
>> line.... I guess have some questions and could offer to do smp
>> tests.)
Andrew> Was that observed while using reiserfs?
Yes, in the disk-related stress tests.
Andrew> reiserfs doesn't work as well as ext2 - 3 milliseconds isn't a
Andrew> big surprise, but it's top of the range. (But it's better than
Andrew> the normal 100 mSecs...) The reiserfs guys were very helpful
Andrew> in the preparation of that part of the patch and I did back out
Andrew> a few parts because of performance concerns.
Andrew> If you're really after best latencies, stick with ext2 for the
Andrew> while. (Isn't 2.9 mS good enough??)
I think you've done a great thing---the current occcasional spike to 3.5ms
is definitely much better than without your patches :)
I was concerned because my results were different than the notes on your
page led me to expect. I did suspect reiserfs because I had seen the
thread on the kernel mailing list which mentioned that reiserfs hasn't been
optimized for low latency, only for throughput, in their benchmarks.
For now, I'll create a big ext2 partition and see if there's a way for me
to set the audio buffer size so that it's 3.5ms instead of 2.9ms. I'm a
newbie for this sort of thing, but I wonder if you've looked at SGI's XFS?
I like using journaled file systems on these big partitions (an fsck of a
30GB ext2 partition takes a *long* time).
Finally, let me again offer to test out your patches if you're looking for
someone with an smp box w/ reiserfs.
Many thanks for your work,
-chris
This archive was generated by hypermail 2b28 : Mon Feb 12 2001 - 17:08:53 EET