Re: [RFC PATCH 00/20] Cleanup and optimise the page allocator

[Andi Kleen <andi@firstfloor.org>]

Mel Gorman <mel@csn.ul.ie> writes:

> The complexity of the page allocator has been increasing for some time
> and it has now reached the point where the SLUB allocator is doing strange
> tricks to avoid the page allocator. This is obviously bad as it may encourage
> other subsystems to try avoiding the page allocator as well.

Congratulations! That was long overdue. Haven't read the patches yet though.

> Patch 15 reduces the number of times interrupts are disabled by reworking
> what free_page_mlock() does. However, I notice that the cost of calling
> TestClearPageMlocked() is still quite high and I'm guessing it's because
> it's a locked bit operation. It's be nice if it could be established if
> it's safe to use an unlocked version here. Rik, can you comment?

What machine was that again?

> Patch 16 avoids using the zonelist cache on non-NUMA machines

My suspicion is that it can be even dropped on most small (all?) NUMA systems.

> Patch 20 gets rid of hot/cold freeing of pages because it incurs cost for
> what I believe to be very dubious gain. I'm not sure we currently gain
> anything by it but it's further discussed in the patch itself.

Yes the hot/cold thing was always quite dubious.

> Counters are surprising expensive, we spent a good chuck of our time in
> functions like __dec_zone_page_state and __dec_zone_state. In a profiled
> run of kernbench, the time spent in __dec_zone_state was roughly equal to
> the combined cost of the rest of the page free path. A quick check showed
> that almost half of the time in that function is spent on line 233 alone
> which for me is;
>
> 	(*p)--;
>
> That's worth a separate investigation but it might be a case that
> manipulating int8_t on the machine I was using for profiling is unusually
> expensive. 

What machine was that?

In general I wouldn't expect even on a system with slow char
operations to be that expensive. It sounds more like a cache miss or a
cache line bounce. You could possibly confirm by using appropiate
performance counters.

> Converting this to an int might be faster but the increased
> memory consumption and cache footprint might be a problem. Opinions?

One possibility would be to move the zone statistics to allocated
per cpu data. Or perhaps just stop counting per zone at all and
only count per cpu.

> The downside is that the patches do increase text size because of the
> splitting of the fast path into one inlined blob and the slow path into a
> number of other functions. On my test machine, text increased by 1.2K so
> I might revisit that again and see how much of a difference it really made.
>
> That all said, I'm seeing good results on actual benchmarks with these
> patches.
>
> o On many machines, I'm seeing a 0-2% improvement on kernbench. The dominant

Neat.

> So, by and large it's an improvement of some sort.

That seems like an understatement.

-Andi
-- 
ak@linux.intel.com -- Speaking for myself only.