[PATCH v2 02/28] vmscan: take at least one pass with shrinkers

Glauber Costa glommer at parallels.com
Wed Apr 10 07:30:32 UTC 2013


On 04/10/2013 06:51 AM, Joonsoo Kim wrote:
> As you can see, before this patch, do_shrinker_shrink() for
> "huge_zero_page_shrinker" is not called until we call shrink_slab() more
> than 13 times. *Frequency* we call do_shrinker_shrink() actually is
> largely different with before. With this patch, we actually call
> do_shrinker_shrink() for "huge_zero_page_shrinker" 12 times more
> than before. Can we be convinced that there will be no problem?
> 
> This is why I worry about this change.
> Am I worried too much? :)

Yes, you are. The amount of times shrink_slab is called is completely
unpredictable. Changing the size of cached data structures is a lot more
likely to change this than this shrinker change, for instance.

Not to mention, the amount of times shrink_slab() is called is not
changed directly here. But rather, the amount of times an individual
shrinker actually does work.




More information about the Containers mailing list