Re: [PATCH 10/10] mm: balance LRU lists based on relative thrashing

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Mon, Jun 20, 2016 at 04:42:08PM +0900, Minchan Kim wrote:
> On Fri, Jun 17, 2016 at 01:01:29PM -0400, Johannes Weiner wrote:
> > On Fri, Jun 17, 2016 at 04:49:45PM +0900, Minchan Kim wrote:
> > > On Thu, Jun 16, 2016 at 11:12:07AM -0400, Johannes Weiner wrote:
> > > > On Wed, Jun 15, 2016 at 11:23:41AM +0900, Minchan Kim wrote:
> > > > > Do we want to retain [1]?
> > > > > 
> > > > > This patch motivates from swap IO could be much faster than file IO
> > > > > so that it would be natural if we rely on refaulting feedback rather
> > > > > than forcing evicting file cache?
> > > > > 
> > > > > [1] e9868505987a, mm,vmscan: only evict file pages when we have plenty?
> > > > 
> > > > Yes! We don't want to go after the workingset, whether it be cache or
> > > > anonymous, while there is single-use page cache lying around that we
> > > > can reclaim for free, with no IO and little risk of future IO. Anon
> > > > memory doesn't have this equivalent. Only cache is lazy-reclaimed.
> > > > 
> > > > Once the cache refaults, we activate it to reflect the fact that it's
> > > > workingset. Only when we run out of single-use cache do we want to
> > > > reclaim multi-use pages, and *then* we balance workingsets based on
> > > > cost of refetching each side from secondary storage.
> > > 
> > > If pages in inactive file LRU are really single-use page cache, I agree.
> > > 
> > > However, how does the logic can work like that?
> > > If reclaimed file pages were part of workingset(i.e., refault happens),
> > > we give the pressure to anonymous LRU but get_scan_count still force to
> > > reclaim file lru until inactive file LRU list size is enough low.
> > > 
> > > With that, too many file workingset could be evicted although anon swap
> > > is cheaper on fast swap storage?
> > > 
> > > IOW, refault mechanisme works once inactive file LRU list size is enough
> > > small but small inactive file LRU doesn't guarantee it has only multiple
> > > -use pages. Hm, Isn't it a problem?
> > 
> > It's a trade-off between the cost of detecting a new workingset from a
> > stream of use-once pages, and the cost of use-once pages impose on the
> > established workingset.
> > 
> > That's a pretty easy choice, if you ask me. I'd rather ask cache pages
> > to prove they are multi-use than have use-once pages put pressure on
> > the workingset.
> 
> Make sense.
> 
> > 
> > Sure, a spike like you describe is certainly possible, where a good
> > portion of the inactive file pages will be re-used in the near future,
> > yet we evict all of them in a burst of memory pressure when we should
> > have swapped. That's a worst case scenario for the use-once policy in
> > a workingset transition.
> 
> So, the point is how such case it happens frequently. A scenario I can
> think of is that if we use one-cgroup-per-app, many file pages would be
> inactive LRU while active LRU is almost empty until reclaim kicks in.
> Because normally, parallel reclaim work during launching new app makes
> app's startup time really slow. That's why mobile platform uses notifiers
> to get free memory in advance via kiling/reclaiming. Anyway, once we get
> amount of free memory and lauching new app in a new cgroup, pages would
> live his born LRU list(ie, anon: active file: inactive) without aging.
> 
> Then, activity manager can set memory.high of less important app-cgroup
> to reclaim it with high value swappiness because swap device is much
> faster on that system and much bigger anonymous pages compared to file-
> backed pages. Surely, activity manager will expect lots of anonymous
> pages be able to swap out but unlike expectation, he will see such spike
> easily with reclaiming file-backed pages a lot and refault until inactive
> file LRU is enough small.
> 
> I think it's enough possible scenario in small system one-cgroup-per-
> app.

That's the workingset transition I was talking about. The algorithm is
designed to settle towards stable memory patterns. We can't possibly
remove one of the key components of this - the use-once policy - to
speed up a few seconds of workingset transition when it comes at the
risk of potentially thrashing the workingset for *hours*.

The fact that swap IO can be faster than filesystem IO doesn't change
this at all. The point is that the reclaim and refetch IO cost of
use-once cache is ZERO. Causing swap IO to make room for more and more
unused cache pages doesn't make any sense, no matter the swap speed.

I really don't see the relevance of this discussion to this patch set.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>



[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux]     [Linux OMAP]     [Linux MIPS]     [ECOS]     [Asterisk Internet PBX]     [Linux API]