On Mon, 2004-08-02 at 19:57, Florian Schmidt wrote: > I think that this is the right mooment to grab a vanilla 2.6.8-rc2 and > just try one of Ingos new volountary preemption patches. You might be > positively surprised.. If not, send a report to Ingo. Lee Revel is also > busy testing Ingos patches and reporting all latencies he found. But > more testers are needed. Hey, I would have commented earlier but was out of town all week. As Ingo's patches stabilize, it is very important for more people to test them. With 2.6.8+voluntary-preempt-O3, using voluntary_preempt=3, kernel_preemption=1, and the soundcard and RTC irqs non-threaded, it works for me (tm) - I have been unable to produce a latency over 50 usecs, even while punishing the machine with parallel builds and iozone -a. We need people to test with different hardware configurations and workloads. Especially needed are testers with ACPI, APIC, SMP, HT machines. There are still some ACPI related bugs, more testers are needed to nail these down. There are some reports of lockups which I cannot reproduce. I would not recommend any patches other than the stock kernel or -mm kernel + the latest voluntary-preempt patches, just to simplify things. Here are some of my test results using 2.6.8-mm2 + voluntary-preempt O3. I have found the mm series provide about 10-20% better latency, but I have had a few lockups that did not happen with the stock kernel; this is the expected behavior for a development vs. stable kernel. I should point out that we are operating very close to the limits of the hardware - the difference between 45 and 55 usecs. This is at least an order of magnitude BETTER than 2.4 + ll and closer to RTLinux than anything else. http://www.members.dca.net/rlrevell/testresults/test1.png http://www.members.dca.net/rlrevell/testresults/test2.png http://www.members.dca.net/rlrevell/testresults/test3.png My test relies on some patches to jackd that I posted to the jackit-devel mailing list. If anyone would like help getting a test environment up and running please let me know. Some real world testing with actual audio apps would be useful, since my tests measure scheduling latency using a jack setup with no clients. Lee
