Re: [PATCH]: 1/4 batch mark_page_accessed()

[Nikita Danilov]

Andrew Morton writes:
 > Marcelo Tosatti <marcelo.tosatti@xxxxxxxxxxxx> wrote:
 > >
 > > Because the ordering of LRU pages should be enhanced in respect to locality, 
 > >  with the mark_page_accessed batching you group together tasks accessed pages 
 > >  and move them at once to the active list. 
 > > 
 > >  You maintain better locality ordering, while decreasing the precision of aging/
 > >  temporal locality.
 > > 
 > >  Which should enhance disk writeout performance.
 > 
 > I'll buy that explanation.  Although I'm a bit sceptical that it is
 > measurable.

cluster-pageout.patch that was sent together with
mark_page_accessed-batching.patch has roughly similar effect: page-out
is done in file order, ignoring local LRU order at end of the inactive
list. It did improve performance in page-out intensive micro-benchmark:

Averaged number of microseconds it takes to dirty 1GB of
16-times-larger-than-RAM ext3 file mmaped in 1GB chunks:

without-patch:   average:    74188417.156250
               deviation:    10538258.613280

   with-patch:   average:    69449001.583333
               deviation:    12621756.615280

 > 
 > Was that particular workload actually performing significant amounts of
 > writeout in vmscan.c?  (We should have direct+kswapd counters for that, but
 > we don't.  /proc/vmstat:pgrotated will give us an idea).
 > 
 > 
 > >  On the other hand, without batching you mix the locality up in LRU - the LRU becomes 
 > >  more precise in terms of "LRU aging", but less ordered in terms of sequential 
 > >  access pattern.
 > > 
 > >  The disk IO intensive reaim has very significant gain from the batching, its
 > >  probably due to the enhanced LRU ordering (what Nikita says).
 > > 
 > >  The slowdown is probably due to the additional atomic_inc by page_cache_get(). 
 > > 
 > >  Is there no way to avoid such page_cache_get there (and in lru_cache_add also)?
 > 
 > Not really.  The page is only in the pagevec at that time - if someone does
 > a put_page() on it the page will be freed for real, and will then be
 > spilled onto the LRU.  Messy.

I don't think that atomic_inc will be particularly
costly. generic_file_{write,read}() call find_get_page() just before
calling mark_page_accessed(), so cache-line with page reference counter
is most likely still exclusive owned by this CPU.

Nikita.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/