Re: [PATCH v5 00/31] kmemcg shrinkers

On 05/09/2013 06:03 PM, Mel Gorman wrote:

On Thu, May 09, 2013 at 11:18:23PM +1000, Dave Chinner wrote:

quoted

quoted

quoted

Mel, I have identified the overly aggressive behavior you noticed to be a bug
in the at-least-one-pass patch, that would ask the shrinkers to scan the full
batch even when total_scan < batch. They would do their best for it, and
eventually succeed. I also went further, and made that the behavior of direct
reclaim only - The only case that really matter for memcg, and one in which
we could argue that we are more or less desperate for small squeezes in memory.
Thank you very much for spotting this.

I haven't seen the relevant code yet but in general I do not think it is
a good idea for direct reclaim to potentially reclaim all of slabs like
this. Direct reclaim does not necessarily mean the system is desperate
for small amounts of memory. Lets take a few examples where it would be
a poor decision to reclaim all the slab pages within direct reclaim.

1. Direct reclaim triggers because kswapd is stalled writing pages for
   memcg (see code near comment "memcg doesn't have any dirty pages
   throttling"). A memcg dirtying its limit of pages may cause a lot of
   direct reclaim and dumping all the slab pages

2. Direct reclaim triggers because kswapd is writing pages out to swap.
   Similar to memcg above, kswapd failing to make forward progress triggers
   direct reclaim which then potentially reclaims all slab

3. Direct reclaim triggers because kswapd waits on congestion as there
   are too many pages under writeback. In this case, a large amounts of
   writes to slow storage like USB could result in all slab being reclaimed

4. The system has been up a long time, memory is fragmented and the page
   allocator enters direct reclaim/compaction to allocate THPs. It would
   be very unfortunate if allocating a THP reclaimed all the slabs

All that is potentially bad and likely to make Dave put in his cranky
pants. I would much prefer if direct reclaim and kswapd treated slab
similarly and not ask the shrinkers to do a full scan unless the alternative
is OOM kill.

Just keep in mind that I really don't care about micro-behaviours of
the shrinker algorithm. What I look at is the overall cache balance
under steady state workloads, the response to step changes in
workload and what sort of overhead is seen to maintain system
balance under memory pressure. So unless a micro-behaviour has an
impact at the macro level, I just don't care one way or the other.

Ok, that's fine by me because I think what you are worried about can
happen too easily right now.  A system in a steady state of streaming
IO can decide to reclaim excessively in direct reclaim becomes active --
a macro level change for a steady state workload.

However, Glauber has already said he will either make a priority check in
direct reclaim or make it memcg specific. I'm happy with either as either
should avoid a large impact at a macro level in response to a small change
in the workload pattern.

quoted

But I can put on cranky panks if you want, Mel. :)

Unjustified cranky pants just isn't the same :)

quoted

quoted

quoted

Running postmark on the final result (at least on my 2-node box) show something
a lot saner. We are still stealing more inodes than before, but by a factor of
around 15 %. Since the correct balance is somewhat heuristic anyway - I
personally think this is acceptable. But I am waiting to hear from you on this
matter. Meanwhile, I am investigating further to try to pinpoint where exactly
this comes from. It might either be because of the new node-aware behavior, or
because of the increased calculation precision in the first patch.

I'm going to defer to Dave as to whether that increased level of slab
reclaim is acceptable or not.

Depends on how it changes the balance of the system. I won't know
that until I run some new tests.

Thanks

Ok guys

The "problem" (change of behavior, actually), lies somewhere between
those two consecutive patches:

    dcache: convert to use new lru list infrastructure
    list_lru: per-node list infrastructure

I cannot pinpoint it for sure because the results I've got for the first
one were quite weird, and we have actually stolen *a lot* *less* inodes
with this patch. I decided to re-run the test just to be sure, but I am
already back home, so I will grab the results tomorrow.

The fact that the stealing of inodes increases after the list_lru patch
seems to indicate that this is because we are now able to shrink in
parallel due to the per node lists. It is only reasonable that we will
be able to do more work, and it is consistent with expectations.

However, to confirm that, I think it would be beneficial to disable one
of the nodes in my system and then run it again (which I will have to do
tomorrow). Meanwhile, of course, other tests and validations from Dave
are welcome.