Subject: [linux-audio-dev] Lowish-latency test results with kernel 2.4.0-test4
From: Tommi Ilmonen (tilmonen_AT_cc.hut.fi)
Date: Fri Aug 04 2000 - 09:56:14 EEST
Hi,
Few more test results with 2.4-test4 kernel and Andrew Morton's
"lowish-latency" patch. I had bit of trouble finding the patch and
finding out whether this is the right kernel to patch...
I am using my own benchmark program. This works in a fairly usual
fashion I believe:
1) Lock memory
2) Get RT-priority
3) Open /dev/rtc and set clock frequency to 4096 Hz (or 2048)
while(no HUP signal) {
4) Get reference time
5) Waste some CPU by calculating random numbers
6) Wait for two ticks to come from the RTC
7) Get time and compare against reference time (from step 4)
8) Save the result
}
9) Save results to disk
This more or less resembles the behaviour of many RT-apps (well, at
least mine...).
Once I get working ALSA drivers I'll test with audio running at all
directions...
--Short test period results (tested with time span of a few minutes):
Application starts seem to introduce latency: Whenever I start Netscape the RT-app experiences a 14 ms hichkup -> not nice. Nearly all apps cause some problems as they are started, but Nutscape (not too surprisingly) is the worst. BTW: 2.2.16 with Ingo's ll-patches does not seem to be any better (to be tested thoroughly tough).
This does not seem to be a resource problem - I can run "make -j 10" with maximum latency around 6 ms. And I am not running out of memory either. Simply starting a big binary seems to be enough to cause the problems.
--
Long period resuls (left the machine do benchmarking overnight in an unloaded machine):
Single latency peak at 66 ms, another at 15ms, otherwise below 7 ms. I have no idea what caused these problems.
--
All in all the results are either fairly encouraging or really bad (depending on your view).
These results seem to be a bit different from the ones people have been getting.
Hardware: Uniprocessor K7_AT_600MHz, 128Mb, IDE-HD (is this the source of trouble ?). I made the kernel as small as possible (minimal number of modules/features).
Tommi.
This archive was generated by hypermail 2b28 : Fri Aug 04 2000 - 10:32:57 EEST