Message-ID: <20090615105651.GD23198@csn.ul.ie>
Date:	Mon, 15 Jun 2009 11:56:51 +0100
From:	Mel Gorman <mel@....ul.ie>
To:	KOSAKI Motohiro <kosaki.motohiro@...fujitsu.com>
Cc:	Andrew Morton <akpm@...ux-foundation.org>, riel@...hat.com,
	cl@...ux-foundation.org, fengguang.wu@...el.com,
	linuxram@...ibm.com, linux-mm@...ck.org,
	linux-kernel@...r.kernel.org, Johannes Weiner <hannes@...xchg.org>
Subject: Re: [PATCH 0/3] Fix malloc() stall in zone_reclaim() and bring
	behaviour more in line with expectations V3

On Mon, Jun 15, 2009 at 06:42:37PM +0900, KOSAKI Motohiro wrote:
> Hi
> 
> > On Thu, Jun 11, 2009 at 04:30:06PM -0700, Andrew Morton wrote:
> > > On Thu, 11 Jun 2009 11:47:50 +0100
> > > Mel Gorman <mel@....ul.ie> wrote:
> > > 
> > > > The big change with this release is that the patch reintroducing
> > > > zone_reclaim_interval has been dropped as Ram reports the malloc() stalls
> > > > have been resolved. If this bug occurs again, the counter will be there to
> > > > help us identify the situation.
> > > 
> > > What is the exact relationship between this work and the somewhat
> > > mangled "[PATCH for mmotm 0/5] introduce swap-backed-file-mapped count
> > > and fix
> > > vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch"
> > > series?
> > > 
> > 
> > The patch series "Fix malloc() stall in zone_reclaim() and bring
> > behaviour more in line with expectations V3" replaces
> > vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch.
> > 
> > Portions of the patch series "Introduce swap-backed-file-mapped count" are
> > potentially follow-on work if a failure case can be identified. The series
> > brings the kernel behaviour more in line with documentation, but it's easier
> > to fix the documentation.
> 
> Agreed.
> 
> 
> > > That five-patch series had me thinking that it was time to drop 
> > > 
> > > vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch
> > 
> > This patch gets replaced. All the lessons in the new patch are included.
> > They could be merged together.
> 
> Sure.
> 
> 
> > > vmscan-drop-pf_swapwrite-from-zone_reclaim.patch
> > 
> > This patch is wrong, but only sortof. It should be dropped or replaced with
> > another version. Kosaki, could you resubmit this patch except that you check
> > if RECLAIM_SWAP is set in zone_reclaim_mode when deciding whether to set
> > PF_SWAPWRITE or not please?
> 
> OK. I'll test it again with your patch.
> 
> > Your patch is correct if zone_reclaim_mode 1, but incorrect if it's 7 for
> > example.
> 
> May I ask your worry?
> 

Simply that I believe the intention of PF_SWAPWRITE here was to allow
zone_reclaim() to aggressively reclaim memory if the reclaim_mode allowed
it as it was a statement that off-node accesses are really not desired.

> Parhaps, my patch description was wrong. I should wrote patch effective
> to separate small and large server.
> 
> First, our dirty page limitation is sane. Thus we don't need to care
> all pages of system are dirty.
> 

For dirty pages, that seems fair, but what about pages that have been recently
unmapped and are being written to swap? If PF_SWAPWRITE is not set, then is it
possible we'll fail to pageout due to IO congestion even though zone_reclaim()
queues only a small number of pages?

> Thus, on large server, turning off PF_SWAPWRITE don't cause off-node allocation.
> There are always clean and droppable page in system.
> 

And it would be expected that if the scan-avoidance heuristics are working,
then we know there are pages that can be reclaimed. However, if the reclaim
mode is allowing the writing, unmapping and swapping of pages, there could
still be a lot of IO to generate.

> In the other hand, on small server, we need to concern write-back race because
> system memory are relatively small.
> Thus, turning off PF_SWAPWRITE might cause off-node allocation.
> 
>   - typically, small servers are latency aware than larger one.
>   - zone reclaim is not the feature of gurantee no off-node allocation.
>   - on small server, off-node allocation penalty is not much rather larger
>     one in many case.
> 
> I sitll think this patch is valueable.
> 

Ok. I am not fully convinced but I'll not block it either if believe it's
necessary. My current understanding is that this patch only makes a difference
if the server is IO congested in which case the system is struggling anyway
and an off-node access is going to be relatively small penalty overall.
Conceivably, having PF_SWAPWRITE set makes things worse in that situation
and the patch makes some sense.

> 
> > > vmscan-zone_reclaim-use-may_swap.patch
> > > 
> > 
> > This is a tricky one. Kosaki, I think this patch is a little dangerous. With
> > this applied, pages get unmapped whether RECLAIM_SWAP is set or not. This
> > means that zone_reclaim() now has more work to do when it's enabled and it
> > incurs a number of minor faults for no reason as a result of trying to avoid
> > going off-node. I don't believe that is desirable because it would manifest
> > as high minor fault counts on NUMA and would be difficult to pin down why
> > that was happening.
> 
> (cc to hanns)
> 
> First, if this patch should be dropped, commit bd2f6199 
> (vmscan: respect higher order in zone_reclaim()) should be too. I think.
> the combination of lumply reclaim and !may_unmap is really ineffective.

Whether it's ineffective or not, it's what the user has asked for. They
want a high-order page found if possible within the limits of
zone_reclaim_mode. If it fails, they will enter full direct reclaim
later in the path and try again.

How effective lumpy reclaim is in this case really depends on what the
system has been used for in the past. It's impossible to know in advance
how effective lumpy reclaim will be in every case.

> it might cause isolate neighbor pages and give up unmapping and pages put
> back tail of lru.
> it mean to shuffle lru list.
> 
> I don't think it is desirable.
> 

With Kamezawa Hiroyuki's patch that avoids unnecessary shuffles of the LRU
list due to lumpy reclaim, the situation might be better?

> Second, we did learned that "mapped or not mapped" is not appropriate
> reclaim boosting between split-lru discussion.
> So, I think to make consistent is better. if no considerness of may_unmap
> makes serious performance issue, we need to fix try_to_free_pages() path too.
> 

I don't understand this paragraph.

If zone_reclaim_mode is set to 1, I don't believe the expected behaviour is
for pages to be unmapped from page tables. I think it will lead to mysterious
bug reports of higher numbers of minor faults when running applications on
NUMA machines in some situations.

> 
> Third, if we consider MPI program on NUMA, each process only access
> a part of array data frequently and never touch rest part of array.
> So, AFAIK "rarely, but access" is rarely, no freqent access is not major performance source.
> 
> I have one question. your "difficultness of pinning down" is major issue?
> 

Yes. If an administrator notices that minor fault rates are higher than
expected, it's going to be very difficult for them to understand why
it is happening and why setting reclaim_mode to 0 apparently fixes the
problem. oprofile for example might just show that a lot of time is being
spent in the fault paths but not explain why.

> 
> > 
> > I think the code makes more sense than the documentation and it's the
> > documentation that should be fixed. Our current behaviour is to discard
> > clean, swap-backed, unmapped pages that require no further IO. This is
> > reasonable behaviour for zone_reclaim_mode == 1 so maybe the patch
> > should change the documentation to
> > 
> >         1       = Zone reclaim discards clean unmapped disk-backed pages
> >         2       = Zone reclaim writes dirty pages out
> >         4       = Zone reclaim unmaps and swaps pages
> > 
> > If you really wanted to strict about the meaning of RECLAIM_SWAP, then
> > something like the following would be reasonable;
> > 
> > 	.may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
> > 	.may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
> > 
> > because a system administrator is not going to distinguish between
> > unmapping and swap. I would assume at least that RECLAIM_SWAP implies
> > unmapping pages for swapping but an updated documentation wouldn't hurt
> > with
> > 
> > 	4       = Zone reclaim unmaps and swaps pages
> > 
> > > (they can be removed cleanly, but I haven't tried compiling the result)
> > > 
> > > but your series is based on those.
> > > 
> > 
> > The patchset only depends on
> > vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch
> > and then only because of merge conflicts. All the lessons in
> > vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch are
> > incorporated.
> > 
> > > We have 142 MM patches queued, and we need to merge next week.
> > > 
> > 
> > I'm sorry my timing for coming out with the zone_reclaim() patches sucks
> > and that I failed to spot these patches earlier. Despite the abundance
> > of evidence, I'm not trying to be deliberatly awkward :/
> > 

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists