We are running a survey to help us improve the experience for all of our members. If you see the survey appear, please take the time to tell us about your experience if you can.
Do calculation in bigger memory arrays in a quad core A53 SoC, using all 4 cores and observe a huge jitter in execution times. That is even the case if data are always same and at same location - so really repeatable situation. In an example the range for operation is between 90...110ms.
I am aware that my OS (Linux, test different kernels up to 6.8) isn't an RT OS. (RT patches make things even worse here). But without other load of usual "housekeeping" I would not expect differences of several ms or even more than 20ms in doing calculation . Also I don't see that behaviour in a Dual Cortex A9 or running that task on A53 disabling 3 cores - which is no real option from performance point of view.
So I suspect that very little and short interrupts or scheduler events (that can't be avoided) can have massive impact on cache or memory access. In kernelshark I can see that cores really go to idle for some ms instead of continue operations. To make it clear - there are no dependencies in that operation - so each core can run its parts. I even tried to isolate calculation cores from "normal" system. That improves situation a bit. But as soon as there is additional (even light) load on that system core, the jitter of the processing in the 3 other cores increase again.
Is it a known issue on A53 and is there something I can do (without completely changing software, which isn't possible) to get less jitter here?