On Thu, May 03, 2018 at 08:15:59PM +1000, Nicholas Piggin wrote: > On Thu, 03 May 2018 20:03:55 +1000 > Stewart Smith <stew...@linux.vnet.ibm.com> wrote: > > > Nicholas Piggin <npig...@gmail.com> writes: > > > On Thu, 3 May 2018 14:36:47 +0530 > > > Akshay Adiga <akshay.ad...@linux.vnet.ibm.com> wrote: > > > > > >> On Tue, May 01, 2018 at 01:47:23PM +1000, Nicholas Piggin wrote: > > >> > On Mon, 30 Apr 2018 14:42:08 +0530 > > >> > Akshay Adiga <akshay.ad...@linux.vnet.ibm.com> wrote: > > >> > > > >> > > Powersaving for stop0_lite and stop1_lite is observed to be quite > > >> > > similar > > >> > > and both states resume without state loss. Using context_switch test > > >> > > [1] > > >> > > we observe that stop0_lite has slightly lower latency, hence removing > > >> > > stop1_lite. > > >> > > > > >> > > [1] linux/tools/testing/selftests/powerpc/benchmarks/context_switch.c > > >> > > > > >> > > Signed-off-by: Akshay Adiga <akshay.ad...@linux.vnet.ibm.com> > > >> > > > >> > I'm okay for removing stop1_lite and stop2_lite -- SMT switching > > >> > is very latency critical. If we decide to actually start saving > > >> > real power then SMT should already have been switched. > > >> > > > >> > So I would put stop1_lite and stop2_lite removal in the same patch. > > >> > > >> I can do this. > > >> > > >> > > > >> > Then what do we have? stop0_lite, stop0, stop1 for our fast idle > > >> > states. > > >> > > >> Currently we were looking at stop0_lite , stop1 as the fast idle states > > >> because stop0 and stop1 have similar latency and powersaving. > > >> Having so many low latency states does not make sense. > > >> > > >> > > > >> > I would be against removing stop0 if that is our fastest way to > > >> > release SMT resources, even if there is only a small advantage. Why > > >> > not remove stop1 instead? > > >> > > > >> SMT-folding comes into picture only when we have at least one thread > > >> running in the core. stop0 and stop1 has exactly same power-saving and > > >> both will release SMT resources if at least one thread in the core is > > >> running. > > > > > > Right, but you don't know that other threads are running or will remain > > > running when you enter stop. If not, then latency is higher for stop1, > > > no? So we need to be using stop0. > > > > > >> > > >> As soon as all threads are idle core enters stop0/stop1, where stop1 > > >> does a bit more powersaving than stop0. > > >> > > >> > We also need to better evaluate stop0_lite. How much advantage does > > >> > that have over snooze? > > >> > > >> I evaluated snooze and stop0_lite, there is an additional ipi latency of > > >> a few microseconds in case of stop0_lite. So snooze cannot still be > > >> replaced by stop0_lite. > > > > > > I meant the other way around. Replace stop0_lite with snooze. > > > > > > So we would have snooze, stop0, stop2, and stop4 and/or 5. > > > > Slightly stupid question: should we be disabling these here or should > > Linux be better and deciding what states to use? > > Yeah not a bad question, I don't have a good answer. I don't know how > smart Linux is at deciding what to use and when. > > I am pretty sure the way we set our _lite states wrong -- we don't > want to go into stop2_lite as a deeper sleep state than stop0 for > example, because that then prevents SMT folding.
I think we should keep both stop0 and stop1, i was not able to get a good enough reason to remove stop0. I a diffrent patch we need to tweak residencies so that we can bias to more useful stop states.
