On Thu, Jun 02, 2016 at 07:50:23AM +0200, Mike Galbraith wrote: > > > Nope, those two have different meanings. We pass SD_BALANCE_WAKE to > > > identify a ttwu() wakeup, just as we pass SD_BALANCE_FORK to say we're > > > waking a child. SD_WAKE_AFFINE means exactly what it says, but is only > > > applicable to ttwu() wakeups. > > > > I don't disagree, but want to add that, SD_WAKE_AFFINE has no meaning that > > is so > > special and so important for anyone to use the flag to tune anything. If > > you want > > to do any SD_BALANCE_*, waker CPU is a valid candidate if allowed, that is > > it. > > That flag lets the user specifically tell us that he doesn't want us to > bounce his tasks around the box, cache misses be damned. The user may > _know_ that say cross node migrations hurt his load more than help, and > not want us to do that, thus expresses himself by turning the flag off > at whatever level. People do that. You can force them to take other > measures, but why do that? Agreed, and with this patch, just disable SD_BALANCE_WAKE.
> > IIUC your XXX mark and your comment "Prefer wake_affine over balance > > flags", you > > said the same thing: SD_WAKE_AFFINE should be part of SD_BALANCE_WAKE, and > > should > > be part of all SD_BALANCE_* flags, > > Peter wrote that, but I don't read it the way you do. I read as if the > user wants the benefits of affine wakeups, he surely doesn't want us to > send the wakee off to god know where on every wakeup _instead_ of > waking affine, he wants to balance iff he can't have an affine wakeup. That is another matter within SD_BALANCE_WAKE we may further define: how much effort to scan or how frequent bouncing etc the user wants. This is now defined by SD_WAKE_AFFINE flag, which I certainly don't think is good. > > > If wake_wide() says we do not want an affine wakeup, but you apply > > > SD_WAKE_AFFINE meaning to SD_BALANCE_WAKE and turn it on in ->flags, > > > we'll give the user a free sample of full balance cost, no? > > > > Yes, and otherwise we don't select anything? That is just bad engough > > whether worse > > or not. So the whole fuss I made is really that this is a right thing to > > start with. :) > > Nope, leaving tasks where they were is not a bad thing. Lots of stuff > likes the scheduler best when it leaves them the hell alone :) That > works out well all around, balance cycles are spent in userspace > instead, scheduler produces wins by doing nothing, perfect. > Again, agreed, and with this patch, just disable SD_BALANCE_WAKE. :)
