On Tue, Apr 14, 2020 at 09:31:24AM +0800, Huang, Ying wrote:
> Andrea Righi <andrea.righi@xxxxxxxxxxxxx> writes:
>
> > On Mon, Apr 13, 2020 at 09:00:34PM +0800, Huang, Ying wrote:
> >> Andrea Righi <andrea.righi@xxxxxxxxxxxxx> writes:
> >>
> >> [snip]
> >>
> >> > diff --git a/mm/swap_state.c b/mm/swap_state.c
> >> > index ebed37bbf7a3..c71abc8df304 100644
> >> > --- a/mm/swap_state.c
> >> > +++ b/mm/swap_state.c
> >> > @@ -20,6 +20,7 @@
> >> > #include <linux/migrate.h>
> >> > #include <linux/vmalloc.h>
> >> > #include <linux/swap_slots.h>
> >> > +#include <linux/oom.h>
> >> > #include <linux/huge_mm.h>
> >> >
> >> > #include <asm/pgtable.h>
> >> > @@ -507,6 +508,14 @@ static unsigned long swapin_nr_pages(unsigned long offset)
> >> > max_pages = 1 << READ_ONCE(page_cluster);
> >> > if (max_pages <= 1)
> >> > return 1;
> >> > + /*
> >> > + * If current task is using too much memory or swapoff is running
> >> > + * simply use the max readahead size. Since we likely want to load a
> >> > + * lot of pages back into memory, using a fixed-size max readhaead can
> >> > + * give better performance in this case.
> >> > + */
> >> > + if (oom_task_origin(current))
> >> > + return max_pages;
> >> >
> >> > hits = atomic_xchg(&swapin_readahead_hits, 0);
> >> > pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,
> >>
> >> Thinks this again. If my understanding were correct, the accessing
> >> pattern during swapoff is sequential, why swap readahead doesn't work?
> >> If so, can you root cause that firstly?
> >
> > Theoretically if the pattern is sequential the current heuristic should
> > already select a big readahead size, but apparently it's not doing that.
> >
> > I'll repeat my tests tracing the readahead size during swapoff to see
> > exactly what's going on here.
>
> I haven't verify it. It may be helpful to call lookup_swap_cache()
> before swapin_readahead() in unuse_pte_range(). The theory behind it is
> to update the swap readahead statistics via lookup_swap_cache().
I did more tests trying to collect some useful information.
In particular I've been focusing at tracing the distribution of the
values returned by swapin_nr_pages() in different scenarios.
To do so I made swapin_nr_pages() trace-able and I used the following
bcc command to measure the distrubution of the returned values:
# argdist-bpfcc -c -C 'r::swapin_nr_pages(unsigned long offset):unsigned long:$retval'
I've collected this metric in the following scenarios:
- 5.6 vanilla
- 5.6 + lookup_swap_cache() before swapin_readahead() in
unuse_pte_range()
- 5.6 + atomic_inc(&swapin_readahead_hits) before swapin_readahead()
in unuse_pte_range()
- 5.6 + swapin_readahead_hits=last_readahead_pages (in the atomic way)
before swapin_readahead() in unuse_pte_range()
Each kernel has been tested both with swappiness=0 and swappiness=60.
Results are pretty much identical changing the swappiness, so I'm just
reporting the default case here (w/ swappiness=60).
Result
======
= swapoff performance (elapsed time) =
vanilla 22.09s
lookup_swap_cache() 23.87s
hits++ 16.10s
hits=last_ra_pages 8.81s
= swapin_nr_pages() $retval distribution =
5.6 vanilla:
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval
COUNT EVENT
36948 $retval = 8
44151 $retval = 4
49290 $retval = 1
527771 $retval = 2
5.6 lookup_swap_cache() before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval
COUNT EVENT
13093 $retval = 1
56703 $retval = 8
123067 $retval = 2
366118 $retval = 4
5.6 atomic_inc(&swapin_readahead_hits) before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval
COUNT EVENT
2589 $retval = 1
8016 $retval = 2
40021 $retval = 8
566038 $retval = 4
5.6 swapin_readahead_hits=last_readahead_pages before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval
COUNT EVENT
785 $retval = 2
1072 $retval = 1
21844 $retval = 4
644168 $retval = 8
In the vanilla case, the readahead heuristic seems to choose 2 pages
most of the time. This is because we are not properly considering the
hits (hits are always 0 in the swapoff code path) and, as you correctly
pointed out, we can fix this by calling lookup_swap_cache() in
unuse_pte_range() before calling swapin_readahead().
With this change the distribution of the readahead size moves more
toward 4 pages, but we still have some 2s. That looks good, however it
doesn't seem to speed up swapoff very much... maybe because calling
lookup_swap_cache() introduces a small overhead? (still need to
investigate about this theory).
In the next test I've tried to always increment hits by 1 before calling
swapin_readahead() in unuse_pte_range(). This is basically cheating,
because I'm faking the hit ratio, forcing the heuristic to use a larger
readahead size; in fact, the readahead size moves even more toward 4
pages and swapoff performance are a little better now.
Pushing even more the "cheating" I can pretend that the previous
readahead was all hits (swapin_readahead_hits=last_readahead_pages), so
I'm forcing the heuristic to move toward the max size and keep using it.
The result here is pretty much identical to my fixed-size patch, because
swapin_nr_pages() returns the max readahead size pretty much all the
time during swapoff (8 pages or, more in general, vm.page-cluster).
Personally I don't like very much forcing the heuristic in this way,
it'd be nice if it would just work by accounting the proper hit ratio
(so just by adding lookup_swap_cache() as you correctly suggested), but
this solution doesn't seem to improve performance in reality. For this
reason I still think we should consider the swapoff scenario like a
special one and somehow bypass the readahead heuristic and always return
the max readahead size.
Looking at the hits of the previous step in the swapoff case just
doesn't work, because we may have some misses, but they will become hits
very soon, since we are reading all the swapped out pages back into
memory. This is why using the max readahead size gives better
swapoff performance.
What do you think?
Thanks,
-Andrea
