Messages in this thread | | | Date | Wed, 27 May 2009 09:08:13 +0900 | From | KAMEZAWA Hiroyuki <> | Subject | Re: [RFC][PATCH 5/5] (experimental) chase and free cache only swap |
| |
On Tue, 26 May 2009 20:14:00 +0200 Johannes Weiner <hannes@cmpxchg.org> wrote:
> On Tue, May 26, 2009 at 12:18:34PM +0900, KAMEZAWA Hiroyuki wrote: > > > > From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> > > > > Just a trial/example patch. > > I'd like to consider more. Better implementation idea is welcome. > > > > When the system does swap-in/swap-out repeatedly, there are > > cache-only swaps in general. > > Typically, > > - swapped out in past but on memory now while vm_swap_full() returns true > > pages are cache-only swaps. (swap_map has no references.) > > > > This cache-only swaps can be an obstacles for smooth page reclaiming. > > Current implemantation is very naive, just scan & free. > > I think we can just remove that vm_swap_full() check in do_swap_page() > and try to remove the page from swap cache unconditionally. > I'm not sure why reclaim swap entry only at write fault.
> If it's still mapped someplace else, we let it cached. If not, there > is not much use for keeping it around and we free it. > yes.
> When I removed it and did benchmarks, I couldn't spot any difference > in the timings, though. Did you measure the benefits of your patch > somehow? My patche has no "performance benefit". (My patch description may be bad.) I just checked that cache-only-swap can be big.(by sysrq-m)
Even when we remove vm_swap_full() in do_swap_page(), swapin-readahead + trylock-at-zap + vmscan makes "unused" swap caches easily. It reaches 1M in 2hours test of heavy swap program.
But yes, I admit I don't like scan & free. I'm now thinking some mark-and-sweep approach...but it tends to consume memory and to be racy ;)
> > According to the git history tree, vm_swap_full() was initially only > used to aggressively drop cache entries even when they are mapped. > > Rik put it into vmscan to reclaim swap cache _at all_ for activated > pages. But I think unconditionally dropping the cache entry makes > sense if the page gets shuffled around on the LRU list. Better to > re-allocate a swap slot close to the new LRU buddies on the next scan. > > And having this all covered, the need for the scanning your patch does > should be gone, unless I missed something. > Considering memcg, global lru scanning is no help ;( And I'm writing this patch for memcg.
It seems I should make this 5/5 patch as an independent one and test 1-4/5 first.
Thank you for review.
Regards, -Kame
| |