PROCESSOR-SDK-AM62X: power management framework

Lisandro Bravo

Hi:

according to this post https://e2e.ti.com/support/processors-group/processors/f/processors-forum/1180549/processor-sdk-am62x-what-is-the-qos-of-the-linux-rt-system-and-what-is-its-function?tisearch=e2e-sitesearch&keymatch=%22cpu_dma_latency%22#, " the power management framework is not configured on", I would like to know if in the latest release this was implemented already and which specific kernel config parameter to set to activate it.

over 1 year ago

0 Pekka Varis over 1 year ago

TI__Mastermind 27050 points

In RT Linux (PREEMPT_RT) you would have power management off, for real-time/determinism you generally want to run constant performance point. With changes in clock speed and low power modes, worst case interrupt latency will suffer.

0 Lisandro Bravo over 1 year ago in reply to Pekka Varis

Prodigy 91 points

Hi Pekka, thanks for such a fast reply.

You are correct, the reason for my question is to be able to guarantee the CPU to be in C0, given that my current cyclictest results are not ideal (tested on am62 evk with latest rt image from TI)

cyclictest --smp --mlockall --interval=125 --priority=99 --clock=1

WARN: stat /dev/cpu_dma_latency failed: No such file or directory

policy: fifo: loadavg: 0.55 0.50 0.23 1/226 2076

T: 0 ( 2037) P:99 I:125 C: 72724 Min: 4 Act: 6 Avg: 5 Max: 42

T: 1 ( 2038) P:99 I:625 C: 14541 Min: 4 Act: 6 Avg: 6 Max: 31

T: 2 ( 2039) P:99 I:1125 C: 8069 Min: 4 Act: 6 Avg: 6 Max: 38

T: 3 ( 2040) P:99 I:1625 C: 5582 Min: 4 Act: 7 Avg: 6 Max: 38

Those max numbers are too high, and they seem to be outliers given that the avg always stays at an acceptable level

+1 Pekka Varis over 1 year ago in reply to Lisandro Bravo

TI__Mastermind 27050 points

I'm not sure what kind of performance you were expecting, those look like very competitive results for any A53 or any <2W embedded SoC running a reasonable set of applications. In general for Linux you should see something like few per 1M or 100k outliers. So average in the single digit microseconds, with something like couple per million way above that. See https://www.osadl.org/Latency-plot-of-system-in-rack-e-slot.qa-latencyplot-res6.0.html?shadow=1 and other platforms for examples.

https://e2e.ti.com/support/processors-group/processors/f/processors-forum/1183526/faq-linux-how-do-i-test-the-real-time-performance-of-an-am3x-am4x-am6x-soc is an example of what you can do to tune further.

Pekka

Processors

Processors forum

PROCESSOR-SDK-AM62X: power management framework