OLS 2007 Day 3 (excess.org)

Posted Jul 10, 2007 3:41 UTC (Tue) by mmarq (guest, #2332) [Link]

. In Memory partition instead of creating only a ZONE_MOVABLE to reserve space for movable pages why not create another ZONE_SWAPPABLE for grouping movable pages that are good candidates for swapping out

- a) size space for this new 'ZONE_SWAPPABLE' should not be a big issue, because by itself this reserved space can be a relatively small space, because is not directly dependent of 'big pages' needs, and because is not allocable or reclaimable by any normal processes.

- b) it actually can make the MASTER TABLE more efficient to, and i *guess* is possible, by addressing the whole ZONE_SWAPPABLE as 1 or 2 single pages ?...

- c) ZONE_MOVABLE can be more efficient to, because it will end up having to deal with less target pages of the overall space.

- d) It can be a 'swap cache', that is a 'warm cache', capable of minimizing disk writes even with a high swappiness, and as 'the' place for disk swap pre-fetching. Actually it should/could (?) be viewed by the OS as an extension of swap space.

- e) it will allow an effective and very fast mechanism for marking pages swappables to run ahead of reclamation, constructing a relatively small table, and delegating 'page moving' to the page migration or page compaction code, i. e., no immediate need to move pages out to disk swap.

- f) actually it can provide the means to make 'kswapd' to only work, i. e. to be triggered, based on the memory pressure of the ZONE_SWAPPABLE, which because of its serial FIFO like nature, i *guess* can be less frequent then if the all RAM space is considered. Also the speed of 'kswapd' operations can be faster because there is no need of accessing scattered zones of memory to do its job.