Date:	Fri, 25 Sep 2015 15:09:07 -0400
From:	Johannes Weiner <hannes@...xchg.org>
To:	Mel Gorman <mgorman@...hsingularity.net>
Cc:	Andrew Morton <akpm@...ux-foundation.org>,
	Rik van Riel <riel@...hat.com>,
	Vlastimil Babka <vbabka@...e.cz>,
	David Rientjes <rientjes@...gle.com>,
	Joonsoo Kim <iamjoonsoo.kim@....com>,
	Michal Hocko <mhocko@...nel.org>,
	Linux-MM <linux-mm@...ck.org>,
	LKML <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 07/10] mm, page_alloc: Delete the zonelist_cache

On Mon, Sep 21, 2015 at 11:52:39AM +0100, Mel Gorman wrote:
> The zonelist cache (zlc) was introduced to skip over zones that were
> recently known to be full. This avoided expensive operations such as the
> cpuset checks, watermark calculations and zone_reclaim. The situation
> today is different and the complexity of zlc is harder to justify.
> 
> 1) The cpuset checks are no-ops unless a cpuset is active and in general
>    are a lot cheaper.
> 
> 2) zone_reclaim is now disabled by default and I suspect that was a large
>    source of the cost that zlc wanted to avoid. When it is enabled, it's
>    known to be a major source of stalling when nodes fill up and it's
>    unwise to hit every other user with the overhead.
> 
> 3) Watermark checks are expensive to calculate for high-order
>    allocation requests. Later patches in this series will reduce the cost
>    of the watermark checking.
> 
> 4) The most important issue is that in the current implementation it
>    is possible for a failed THP allocation to mark a zone full for order-0
>    allocations and cause a fallback to remote nodes.
> 
> The last issue could be addressed with additional complexity but as the
> benefit of zlc is questionable, it is better to remove it.  If stalls
> due to zone_reclaim are ever reported then an alternative would be to
> introduce deferring logic based on a timeout inside zone_reclaim itself
> and leave the page allocator fast paths alone.
> 
> The impact on page-allocator microbenchmarks is negligible as they don't
> hit the paths where the zlc comes into play. Most page-reclaim related
> workloads showed no noticeable difference as a result of the removal.
> 
> The impact was noticeable in a workload called "stutter". One part uses a
> lot of anonymous memory, a second measures mmap latency and a third copies
> a large file. In an ideal world the latency application would not notice
> the mmap latency.  On a 2-node machine the results of this patch are
> 
> stutter
>                              4.3.0-rc1             4.3.0-rc1
>                               baseline              nozlc-v4
> Min         mmap     20.9243 (  0.00%)     20.7716 (  0.73%)
> 1st-qrtle   mmap     22.0612 (  0.00%)     22.0680 ( -0.03%)
> 2nd-qrtle   mmap     22.3291 (  0.00%)     22.3809 ( -0.23%)
> 3rd-qrtle   mmap     25.2244 (  0.00%)     25.2396 ( -0.06%)
> Max-90%     mmap     48.0995 (  0.00%)     28.3713 ( 41.02%)
> Max-93%     mmap     52.5557 (  0.00%)     36.0170 ( 31.47%)
> Max-95%     mmap     55.8173 (  0.00%)     47.3163 ( 15.23%)
> Max-99%     mmap     67.3781 (  0.00%)     70.1140 ( -4.06%)
> Max         mmap  24447.6375 (  0.00%)  12915.1356 ( 47.17%)
> Mean        mmap     33.7883 (  0.00%)     27.7944 ( 17.74%)
> Best99%Mean mmap     27.7825 (  0.00%)     25.2767 (  9.02%)
> Best95%Mean mmap     26.3912 (  0.00%)     23.7994 (  9.82%)
> Best90%Mean mmap     24.9886 (  0.00%)     23.2251 (  7.06%)
> Best50%Mean mmap     22.0157 (  0.00%)     22.0261 ( -0.05%)
> Best10%Mean mmap     21.6705 (  0.00%)     21.6083 (  0.29%)
> Best5%Mean  mmap     21.5581 (  0.00%)     21.4611 (  0.45%)
> Best1%Mean  mmap     21.3079 (  0.00%)     21.1631 (  0.68%)
> 
> Note that the maximum stall latency went from 24 seconds to 12 which is still
> bad but an improvement.  The milage varies considerably 2-node machine on an
> earlier test went from 494 seconds to 47 seconds and  a 4-node machine that
> tested an earlier version of this patch went from a worst case stall time of
> 6 seconds to 67ms. The nature of the benchmark is inherently unpredictable
> as it is hammering the system and the milage will vary between machines.
> 
> There is a secondary impact with potentially more direct reclaim because
> zones are now being considered instead of being skipped by zlc. In this
> particular test run it did not occur so will not be described. However,
> in at least one test the following was observed
> 
> 1. Direct reclaim rates were higher. This was likely due to direct reclaim
>   being entered instead of the zlc disabling a zone and busy looping.
>   Busy looping may have the effect of allowing kswapd to make more
>   progress and in some cases may be better overall. If this is found then
>   the correct action is to put direct reclaimers to sleep on a waitqueue
>   and allow kswapd make forward progress. Busy looping on the zlc is even
>   worse than when the allocator used to blindly call congestion_wait().
> 
> 2. There was higher swap activity as direct reclaim was active.
> 
> 3. Direct reclaim efficiency was lower. This is related to 1 as more
>   scanning activity also encountered more pages that could not be
>   immediately reclaimed
> 
> In that case, the direct page scan and reclaim rates are noticeable but
> it is not considered a problem for a few reasons
> 
> 1. The test is primarily concerned with latency. The mmap attempts are also
>    faulted which means there are THP allocation requests. The ZLC could
>    cause zones to be disabled causing the process to busy loop instead
>    of reclaiming.  This looks like elevated direct reclaim activity but
>    it's the correct action to take based on what processes requested.
> 
> 2. The test hammers reclaim and compaction heavily. The number of successful
>    THP faults is highly variable but affects the reclaim stats. It's not a
>    realistic or reasonable measure of page reclaim activity.
> 
> 3. No other page-reclaim intensive workload that was tested showed a problem.
> 
> 4. If a workload is identified that benefitted from the busy looping then it
>    should be fixed by having direct reclaimers sleep on a wait queue until
>    woken by kswapd instead of busy looping. We had this class of problem before
>    when congestion_waits() with a fixed timeout was a brain damaged decision
>    but happened to benefit some workloads.
> 
> If a workload is identified that relied on the zlc to busy loop then it
> should be fixed correctly and have a direct reclaimer sleep on a waitqueue
> until woken by kswapd.
> 
> Signed-off-by: Mel Gorman <mgorman@...hsingularity.net>
> Acked-by: David Rientjes <rientjes@...gle.com>
> Acked-by: Christoph Lameter <cl@...ux.com>
> Acked-by: Vlastimil Babka <vbabka@...e.cz>
> Acked-by: Michal Hocko <mhocko@...e.com>
> ---
>  include/linux/mmzone.h |  74 -----------------
>  mm/page_alloc.c        | 212 -------------------------------------------------
>  2 files changed, 286 deletions(-)

This patch and its results look great!

And I agree, should this affect the balance between kswapd and direct
reclaim, it should be fixed explicitely and not rely on something as
unrelated as the zonelist cache.

Acked-by: Johannes Weiner <hannes@...xchg.org>
--
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