On Tue, Nov 29, 2011 at 04:20:14PM -0800, Andrew Morton wrote:
> On Wed, 23 Nov 2011 14:34:14 +0100
> Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
>
> > From: Johannes Weiner <jweiner@xxxxxxxxxx>
> >
> > The amount of dirtyable pages should not include the full number of
> > free pages: there is a number of reserved pages that the page
> > allocator and kswapd always try to keep free.
> >
> > The closer (reclaimable pages - dirty pages) is to the number of
> > reserved pages, the more likely it becomes for reclaim to run into
> > dirty pages:
> >
> > +----------+ ---
> > | anon | |
> > +----------+ |
> > | | |
> > | | -- dirty limit new -- flusher new
> > | file | | |
> > | | | |
> > | | -- dirty limit old -- flusher old
> > | | |
> > +----------+ --- reclaim
> > | reserved |
> > +----------+
> > | kernel |
> > +----------+
> >
> > This patch introduces a per-zone dirty reserve that takes both the
> > lowmem reserve as well as the high watermark of the zone into account,
> > and a global sum of those per-zone values that is subtracted from the
> > global amount of dirtyable pages. The lowmem reserve is unavailable
> > to page cache allocations and kswapd tries to keep the high watermark
> > free. We don't want to end up in a situation where reclaim has to
> > clean pages in order to balance zones.
> >
> > Not treating reserved pages as dirtyable on a global level is only a
> > conceptual fix. In reality, dirty pages are not distributed equally
> > across zones and reclaim runs into dirty pages on a regular basis.
> >
> > But it is important to get this right before tackling the problem on a
> > per-zone level, where the distance between reclaim and the dirty pages
> > is mostly much smaller in absolute numbers.
> >
> > ...
> >
> > --- a/mm/page-writeback.c
> > +++ b/mm/page-writeback.c
> > @@ -327,7 +327,8 @@ static unsigned long highmem_dirtyable_memory(unsigned long total)
> > &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
> >
> > x += zone_page_state(z, NR_FREE_PAGES) +
> > - zone_reclaimable_pages(z);
> > + zone_reclaimable_pages(z) -
> > + zone->dirty_balance_reserve;
>
> Doesn't compile. s/zone/z/.
>
> Which makes me suspect it wasn't tested on a highmem box. This is
> rather worrisome, as highmem machines tend to have acute and unique
> zone balancing issues.
You are right, so I ran fs_mark on an x86 machine with 8GB and a
32-bit kernel.
fs_mark -S 0 -d work-01 -d work-02 -d work-03 -d work-04 -D 128 -N 128 -L 16 -n 512 -s 655360
This translates to 4 threads doing 16 iterations over a new set of 512
files each time, where each file is 640k in size, which adds up to 20G
of written data per run. The results are gathered over 5 runs. Data
are written to an ext4 on a standard consumer rotational disk.
The overall runtime for the loads were the same:
seconds
mean(stddev)
vanilla: 242.061(0.953)
patched: 242.726(1.714)
Allocation counts confirm that allocation placement does not change:
pgalloc_dma pgalloc_normal pgalloc_high
min|median|max
vanilla: 0.000|0.000|0.000 3733291.000|3742709.000|4034662.000 5189412.000|5202220.000|5208743.000
patched: 0.000|0.000|0.000 3716148.000|3733269.000|4032205.000 5212301.000|5216834.000|5227756.000
Kswapd in both kernels did the same amount of work in each zone over
the course of the workload; direct reclaim was never invoked:
pgscan_kswapd_dma pgscan_kswapd_normal pgscan_kswapd_high
min|median|max
vanilla: 0.000|0.000|0.000 109919.000|115773.000|117952.000 3235879.000|3246707.000|3255205.000
patched: 0.000|0.000|0.000 104169.000|114845.000|117657.000 3241327.000|3246835.000|3257843.000
pgsteal_dma pgsteal_normal pgsteal_high
min|median|max
vanilla: 0.000|0.000|0.000 109912.000|115766.000|117945.000 3235318.000|3246632.000|3255098.000
patched: 0.000|0.000|0.000 104163.000|114839.000|117651.000 3240765.000|3246760.000|3257768.000
and the distribution of scans over time was equivalent, with no new
hickups or scan spikes:
pgscan_kswapd_dma/s pgscan_kswapd_normal/s pgscan_kswapd_high/s
min|median|max
vanilla: 0.000|0.000|0.000 0.000|144.000|2100.000 0.000|15582.500|44916.000
patched: 0.000|0.000|0.000 0.000|152.000|2058.000 0.000|15361.000|44453.000
pgsteal_dma/s pgsteal_normal/s pgsteal_high/s
min|median|max
vanilla: 0.000|0.000|0.000 0.000|144.000|2094.000 0.000|15582.500|44916.000
patched: 0.000|0.000|0.000 0.000|152.000|2058.000 0.000|15361.000|44453.000
fs_mark 1G
The same fs_mark load was run on the system limited to 1G memory
(booted with mem=1G), to have a highmem zone that is much smaller
compared to the rest of the system.
seconds
mean(stddev)
vanilla: 238.428(3.810)
patched: 241.392(0.221)
In this case, allocation placement did shift slightly towards lower
zones, to protect the tiny highmem zone from being unreclaimable due
to dirty pages:
pgalloc_dma pgalloc_normal pgalloc_high
min|median|max
vanilla: 20658.000|21863.000|23231.000 4017580.000|4023331.000|4038774.000 1057246.000|1076280.000|1083824.000
patched: 25403.000|27679.000|28556.000 4163538.000|4172116.000|4179151.000 917054.000| 922206.000| 933609.000
However, while there were in total more allocations in the DMA and
Normal zone, the utilization peaks of the zones individually were
actually reduced due to smoother distribution:
DMA min nr_free_pages Normal min nr_free_pages HighMem min nr_free_pages
vanilla: 1244.000 14819.000 432.000
patched: 1337.000 14850.000 439.000
Keep in mind that the lower zones are only used more often for
allocation because they are providing dirtyable memory in this
scenario, i.e. they have space to spare.
With increasing lowmem usage for stuff that is truly lowmem, like
dcache and page tables, the amount of memory we consider dirtyable
(free pages + file pages) shrinks, so when highmem is not allowed to
take anymore dirty pages, we will not thrash on the lower zones:
either they have space left or the dirtiers are already being
throttled in balance_dirty_pages().
Reclaim numbers suggests that kswapd can easily keep up with the the
allocation frequency increase in the Normal zone. But for DMA, it
looks like the unpatched kernel flooded the zone with dirty pages
every once in a while, making it ineligible for allocations until
those pages were cleaned. Through better distribution, the patch
improves reclaim efficiency (reclaimed/scanned) from 32% to 100% for
DMA:
pgscan_kswapd_dma pgscan_kswapd_normal pgscan_kswapd_high
min|median|max
vanilla: 39734.000|41248.000|41965.000 3692050.000|3696209.000|3716653.000 970411.000|987483.000|991469.000
patched: 21204.000|23901.000|25141.000 3874782.000|3879125.000|3888302.000 793141.000|795631.000|803482.000
pgsteal_dma pgsteal_normal pgsteal_high
min|median|max
vanilla: 12932.000|14044.000|16957.000 3692025.000|3696183.000|3716626.000 966050.000|987386.000|991405.000
patched: 21204.000|23901.000|25141.000 3874771.000|3879095.000|3888284.000 792079.000|795572.000|803370.000
And the increased reclaim efficiency in the DMA zone indeed correlates
with the reduced likelyhood of reclaim running into dirty pages:
DMA Normal Highmem
nr_vmscan_write nr_vmscan_immediate_reclaim
vanilla:
26.0 19614.0 0.0 0.0 1174.0 0.0
0.0 21737.0 0.0 1.0 0.0 0.0
0.0 22101.0 0.0 0.0 0.0 0.0
0.0 21906.0 0.0 0.0 0.0 0.0
0.0 21880.0 0.0 0.0 0.0 0.0
patched:
0.0 0.0 0.0 1.0 502.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 1.0 0.0 0.0
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
