On Mon 22-06-20 17:25:01, Mel Gorman wrote:
> On Mon, Jun 22, 2020 at 04:23:04PM +0200, Michal Hocko wrote:
> > On Mon 22-06-20 11:04:39, Mel Gorman wrote:
> > > On Mon, Jun 22, 2020 at 06:40:20PM +0900, ????????? wrote:
> > > > >But more importantly, I have hard time to follow why we need both
> > > > >zone_watermark_fast and zone_watermark_ok now. They should be
> > > > >essentially the same for anything but order == 0. For order 0 the
> > > > >only difference between the two is that zone_watermark_ok checks for
> > > > >ALLOC_HIGH resp ALLOC_HARDER, ALLOC_OOM. So what is exactly fast about
> > > > >the former and why do we need it these days?
> > > > >
> > > >
> > > > I think the author, Mel, may ansewr. But I think the wmark_fast may
> > > > fast by 1) not checking more condition about wmark and 2) using inline
> > > > rather than function. According to description on commit 48ee5f3696f6,
> > > > it seems to bring about 4% improvement.
> > > >
> > >
> > > The original intent was that watermark checks were expensive as some of the
> > > calculations are only necessary when a zone is relatively low on memory
> > > and the check does not always have to be 100% accurate. This is probably
> > > still true given that __zone_watermark_ok() makes a number of calculations
> > > depending on alloc flags even if a zone is almost completely free.
> >
> > OK, so we are talking about
> > if (alloc_flags & ALLOC_HIGH)
> > min -= min / 2;
> >
> > if (unlikely((alloc_flags & (ALLOC_HARDER|ALLOC_OOM))) {
> > /*
> > * OOM victims can try even harder than normal ALLOC_HARDER
> > * users on the grounds that it's definitely going to be in
> > * the exit path shortly and free memory. Any allocation it
> > * makes during the free path will be small and short-lived.
> > */
> > if (alloc_flags & ALLOC_OOM)
> > min -= min / 2;
> > else
> > min -= min / 4;
> > }
> >
> > Is this something even measurable and something that would justify a
> > complex code? If we really want to keep it even after these changes
> > which are making the two closer in the cost then can we have it
> > documented at least?
>
> It was originally documented as being roughly 4% for a page allocator
> micro-benchmark but that was 4 years ago and I do not even remember what
> type of machine that was on. Chances are the relative cost is different
> now but I haven't measured it as the microbenchmark in question doesn't
> even compile with recent kernels.
Thanks for the clarification.
> For many allocations, the bulk of the
> allocation cost is zeroing the page so I have no particular objection
> to zone_watermark_fast being removed if it makes the code easier to
> read. While I have not looked recently, the cost of allocation in general
> and the increasing scope of the zone->lock with larger NUMA nodes for
> high-order allocations like THP are more of a concern than two branches
> and potentially two minor calculations.
OK, then I would rather go with the code simplification for the future
maintainability. If somebody can test this and provide good numbers then
we can reintroduce a fast check.
Thanks!
--
Michal Hocko
SUSE Labs
