On 19 May 2014 19:55, Daniel Lezcano daniel.lezcano@linaro.org wrote:
Concerning the requirement. Did you notice an improvement ? Was it measurable ?
You meant power saving with it? No, I didn't as that wasn't something I was concerned about. The first objective was to get core isolated for data plane activities and I saw with my initial patch that I could get rid of it. Quite surely it will save some power as well.
Is it possible that adds an extra latency ?
Didn't get this exactly. You meant stopping and starting back the clockevent device will add some latency? Yes, surely some latency would be added but would be platform specific..
In case of NO_HZ_FULL we are more concerned about isolated as we are talking about isolation in seconds (for LNG we are trying for infinite isolation). And that latency isn't going to be a bottleneck for us.
NOTE: This wouldn't happen without NO_HZ_FULL as we always have a hrtimer queued for few ms.
Why not split the second patch per arch + clocksource dir ?
Yes, I am not going to send that patch to mainline as is. It will be divided, but don't know how. I thought of doing it in platform specific way as well, which would mean some drivers for ARM would be in arch/arm/ and others in drivers/clocksource/ .. And so wasn't sure yet.
Do you suggest doing it this way: 1. arch/arm/ 2. arch/mips ... last: drivers/clocksource ??
Don't know, but ARM drivers would be spread across patches then.