Re: [PATCH 1/3] vm: kswapd incmin

[Nick Piggin <nickpiggin@yahoo.com.au>]

Marcelo Tosatti wrote:
> Hi Nick,
> 
> Looks nice, much easier to read than before.
> 

Hi Marcelo,

Thanks! That was one of the main aims.

> One comment: you change the pagecache/slab scanning ratio by moving
> shrink_slab() outside of the zone loop. 
> 
> This means that for each kswapd iteration will scan "lru_pages" 
> SLAB entries, instead of "lru_pages*NR_ZONES" entries.
> 
> Can you comment on that?
> 

I believe I have tried to get it right, let me explain. lru_pages
is just used as the divisor for the ratio between lru scanning
and slab scanning. So long as it is kept constant across calls to
shrink_slab, there should be no change in behaviour.

The the nr_scanned variable is the other half of the equation that
controls slab shrinking. I have changed from:

   lru_pages = total_node_lru_pages;
   for each zone in node {
      shrink_zone();
      shrink_slab(zone_scanned, lru_pages);
   }

To:

   lru_pages = 0;
   for each zone in node {
      shrink_zone();
      lru_pages += zone_lru_pages;
   }
   shrink_slab(total_zone_scanned, lru_pages);

So the ratio remains basically the same
[eg. 10/100 + 20/100 + 30/100 = (10+20+30)/100]

2 reasons for doing this. The first is just efficiency and better
rounding of the divisions.

The second is that within the for_each_zone loop, we are able to
set all_unreclaimable without worrying about slab, because the
final shrink_slab at the end will clear all_unreclaimable if any
zones have had slab pages freed up.

I believe it generally should result in more consistent reclaim
across zones, and also matches direct reclaim better.

Hope this made sense,
Nick

-- 
SUSE Labs, Novell Inc.

Send instant messages to your online friends http://au.messenger.yahoo.com