Re: [PATCH 3/5] mm/vmscan: Throttle reclaim when no progress is being made

[Mel Gorman]

On Wed, Sep 22, 2021 at 07:46:58AM +1000, NeilBrown wrote:
> On Tue, 21 Sep 2021, Mel Gorman wrote:
> > On Tue, Sep 21, 2021 at 09:31:30AM +1000, NeilBrown wrote:
> > > On Mon, 20 Sep 2021, Mel Gorman wrote:
> > > > +
> > > > +		reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS, HZ/10);
> > > 
> > > We always seem to pass "HZ/10" to reclaim_throttle().  Should we just
> > > hard-code that in the one place inside reclaim_throttle() itself?
> > > 
> > 
> > do_writepages passes in HZ/50. I'm not sure if these values even have
> > any special meaning, I think it's more likely they were pulled out of
> > the air based on the speed of some disk in the past and then copied.
> > It's another reason why I want the wakeups to be based on events within
> > the mm as much as possible.
> 
> Yes, I saw the HZ/50 shortly after writing that email :-)
> I agree with your guess for the source of these numbers.  I still think
> we should pull them all from the same piece of air.
> Hopefully, once these changes are properly understood and the events
> reliably come as expected, we can make it quite large (HZ?) with minimal
> cost.
> 

I'd prefer to do it as a separate patch. At some point congestion_wait
worked and the original timeouts may have been selected based on testing
(I severely doubt it but I'm trying to be optimistic). However, we can
at least centralise the decision based on "reason" with this

---8<---