Hi Tim, without hackbench and frieds (idle system) I get a max. latency of about 4ms. Running hackbench (default arguments, -l 1000), it still increases up to 5ms: ... # Total: 000000193 # Min Latencies: 00024 # Avg Latencies: 00205 # Max Latencies: 05058 ... The obvious PM stuff is disabled. No dynamic frequency scaling, etc. Really weired stuff. I already mentioned the 3.2 system with the same CPU in the OSADL lab. It comes with max. latency of 118us. So I really expect something similiar. Matthias On 01/07/2014 09:38 AM, Tim Sander wrote: > Hi >> after some chatting with Nicholas and being sure that I didn't do anything >> totally stupid I would like to provide my current kernel config and a trace >> generated by "cyclictest ... -b500". >> >> kernel config -> https://www.dropbox.com/s/zsud0b77xoosjm1/config >> trace -> >> https://www.dropbox.com/s/5uf58oqo0qwcn4x/trace-cyclictest-2508us.txt >> >> I still get max latencies of about 4000us on this TI AM335x system running >> 3.12.6-rt9 with timers-do-not-raise-softirq-unconditionally.patch reverse >> applied. >> >> sched_rt_runtime_us and sched_rt_period_us are identical. The system is idle >> during the test. Rootfilesystem comes via NFS. >> >> root@generic-armv7a:~# cat /proc/sys/kernel/sched_rt_runtime_us >> 1000000 >> root@generic-armv7a:~# cat /proc/sys/kernel/sched_rt_period_us >> 1000000 >> >> root@generic-armv7a:~# uname -a >> Linux generic-armv7a 3.12.6-rt9-dirty #3 SMP PREEMPT RT Sat Jan 4 23:01:19 >> CET 2014 armv7l GNU/Linux >> >> root@generic-armv7a:~# ~/cyclictest -p99 -m -n -i200 -h400 -q -l100000 -b >> 500 # /dev/cpu_dma_latency set to 0us >> INFO: debugfs mountpoint: /sys/kernel/debug/tracing/ >> could not set ftrace_enabled to 0 >> could not set tracing_max_latency to 0 >> # Histogram >> ... >> # Total: 000018604 >> # Min Latencies: 00024 >> # Avg Latencies: 00158 >> # Max Latencies: 02508 > Wohaa, thats bad for a latency patched kernel. Have you disabled > powermanagement? I have an older rt patched kernel and it was not too good. > If i remember correctly it was between 100-200µs latency. But thats > far better than this. You could also load the system with hackbench, if its > the idle pm than latency would get better. > > Best regards > Tim > > -- > To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in > the body of a message to majordomo@xxxxxxxxxxxxxxx > More majordomo info at http://vger.kernel.org/majordomo-info.html > -- To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
