Messages in this thread | | | From | Vincent Guittot <> | Date | Wed, 21 Sep 2022 18:11:41 +0200 | Subject | Re: [PATCH v4 4/8] sched/core: Add permission checks for setting the latency_nice value |
| |
On Tue, 20 Sept 2022 at 16:56, Vincent Guittot <vincent.guittot@linaro.org> wrote: > > On Tue, 20 Sept 2022 at 12:18, Tim Janik <timj@gnu.org> wrote: > > > > Hi. > > > > On 19.09.22 14:41, Vincent Guittot wrote: > > > Hi, > > > > > > Thanks you for describing in detail your use case. > > > > > Ok, Your explanation makes sense to me especially because we want to > > > ensure to not provide more cpu time with this latency prio. I'm > > > curious to see the feedback from others about the reason we want > > > CAP_SYS_NICE other than following nice priority. > > > > > > Side question, Have you tried this patchset (minus this patch) with > > > your use case ? > > > > I have now tested a modified version of the ALSA Test_latency.c program > > that acquires latency nice as non-root: > > https://gist.github.com/tim-janik/88f9df5456b879ecc59da93dc6ce6be1 > > > > With a busy but not overloaded CPU, the short time latency tests are > > often better, measured with: ./lnice-latency -p -s 1 > > > > But the results aren't very reliable with this test. I.e. requesting a > > latency nice value of -20 reduces the chance for underruns somewhat but > > doesn't eliminate them (and lnice-latency.c gives up on the first XRUN > > It's expected that latency nice can't fix all scheduling latency > problems. The hard real time constraint can only be ensured with FIFO > or deadline scheduler > > > in the given time period). It might be better to instead count the XRUN > > occurances over a given time pertiod. > > Thanks. I'm going to have a look the test
I have done some tests with your modified ALSA Test_latency.c on my dev system. I have been able to run lnice-latency -p -s 8 -e -m 128 simultaneously with hackbench -l 20000 -g 2 on my 8 cores arm64 system. The same test with a default latency nice 0 triggers a lot of XRUN.
Side note, my system doesn't have many RT threads, IRQ or softirq running . But as explained, latency nice can't do much with those
> > > > > > > -- > > Anklang Free Software DAW > > https://anklang.testbit.eu/
| |