On Fri, Jun 08, 2012 at 10:03:24AM +0300, Pekka Enberg wrote:
On Fri, Jun 8, 2012 at 9:58 AM, Anton Vorontsov anton.vorontsov@linaro.org wrote:
If you're saying that we should set up a timer in the userland and constantly read /proc/vmstat, then we will cause CPU wake up every 100ms, which is not acceptable. Well, we can try to introduce deferrable timers for the userspace. But then it would still add a lot more overhead for our task, as this solution adds other two context switches to read and parse /proc/vmstat. I guess this is not a show-stopper though, so we can discuss this.
Leonid, Pekka, what do you think about the idea?
That's exactly the kind of half-assed ABI that lead to people inventing out-of-tree lowmem notifiers in the first place.
:-)
Well, at least powersaving-wise, the solution w/ userland deferred timers would be much better then just looping over /proc/vmstat each 100ms, and it is comparable to vmevent. Not pretty, but still would work.
I'd be more interested to know what people think of Minchan's that gets rid of vmstat sampling.
I answered to Minchan's post. The thing is that Minchan's idea is not a substitution for vmevent. To me it seems like a shrinker w/ some pre-filter.
Thanks,