Re: [PATCH 3/3] mm, vmscan: Prevent kswapd sleeping prematurely due to mismatched classzone_idx

From: Vlastimil Babka
Date: Mon Feb 20 2017 - 11:43:02 EST


On 02/15/2017 10:22 AM, Mel Gorman wrote:
> kswapd is woken to reclaim a node based on a failed allocation request
> from any eligible zone. Once reclaiming in balance_pgdat(), it will
> continue reclaiming until there is an eligible zone available for the
> zone it was woken for. kswapd tracks what zone it was recently woken for
> in pgdat->kswapd_classzone_idx. If it has not been woken recently, this
> zone will be 0.
>
> However, the decision on whether to sleep is made on kswapd_classzone_idx
> which is 0 without a recent wakeup request and that classzone does not
> account for lowmem reserves. This allows kswapd to sleep when a low
> small zone such as ZONE_DMA is balanced for a GFP_DMA request even if
> a stream of allocations cannot use that zone. While kswapd may be woken
> again shortly in the near future there are two consequences -- the pgdat
> bits that control congestion are cleared prematurely and direct reclaim
> is more likely as kswapd slept prematurely.
>
> This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an invalid
> index) when there has been no recent wakeups. If there are no wakeups,
> it'll decide whether to sleep based on the highest possible zone available
> (MAX_NR_ZONES - 1). It then becomes critical that the "pgdat balanced"
> decisions during reclaim and when deciding to sleep are the same. If there is
> a mismatch, kswapd can stay awake continually trying to balance tiny zones.
>
> simoop was used to evaluate it again. Two of the preparation patches regressed
> the workload so they are included as the second set of results. Otherwise
> this patch looks artifically excellent
>
> 4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
> mmots-20170209 clear-v1r25 keepawake-v1r25
> Amean p50-Read 22325202.49 ( 0.00%) 19491134.58 ( 12.69%) 22092755.48 ( 1.04%)
> Amean p95-Read 26102988.80 ( 0.00%) 24294195.20 ( 6.93%) 26101849.04 ( 0.00%)
> Amean p99-Read 30935176.53 ( 0.00%) 30397053.16 ( 1.74%) 29746220.52 ( 3.84%)
> Amean p50-Write 976.44 ( 0.00%) 1077.22 (-10.32%) 952.73 ( 2.43%)
> Amean p95-Write 15471.29 ( 0.00%) 36419.56 (-135.40%) 3140.27 ( 79.70%)
> Amean p99-Write 35108.62 ( 0.00%) 102000.36 (-190.53%) 8843.73 ( 74.81%)
> Amean p50-Allocation 76382.61 ( 0.00%) 87485.22 (-14.54%) 76349.22 ( 0.04%)
> Amean p95-Allocation 127777.39 ( 0.00%) 204588.52 (-60.11%) 108630.26 ( 14.98%)
> Amean p99-Allocation 187937.39 ( 0.00%) 631657.74 (-236.10%) 139094.26 ( 25.99%)
>
> With this patch on top, all the latencies relative to the baseline are
> improved, particularly write latencies. The read latencies are still high
> for the number of threads but it's worth noting that this is mostly due
> to the IO scheduler and not directly related to reclaim. The vmstats are
> a bit of a mix but the relevant ones are as follows;
>
> 4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
> mmots-20170209 clear-v1r25keepawake-v1r25
> Swap Ins 0 0 0
> Swap Outs 0 608 0
> Direct pages scanned 6910672 3132699 6357298
> Kswapd pages scanned 57036946 82488665 56986286
> Kswapd pages reclaimed 55993488 63474329 55939113
> Direct pages reclaimed 6905990 2964843 6352115

These stats are confusing me. The earlier description suggests that this patch
should cause less direct reclaim and more kswapd reclaim, but compared to
"clear-v1r25" it does the opposite? Was clear-v1r25 overreclaiming then? (when
considering direct + kswapd combined)

> Kswapd efficiency 98% 76% 98%
> Kswapd velocity 12494.375 17597.507 12488.065
> Direct efficiency 99% 94% 99%
> Direct velocity 1513.835 668.306 1393.148
> Page writes by reclaim 0.000 4410243.000 0.000
> Page writes file 0 4409635 0
> Page writes anon 0 608 0
> Page reclaim immediate 1036792 14175203 1042571
>
> Swap-outs are equivalent to baseline
> Direct reclaim is reduced but not eliminated. It's worth noting
> that there are two periods of direct reclaim for this workload. The
> first is when it switches from preparing the files for the actual
> test itself. It's a lot of file IO followed by a lot of allocs
> that reclaims heavily for a brief window. After that, direct
> reclaim is intermittent when the workload spawns a number of
> threads periodically to do work. kswapd simply cannot wake and
> reclaim fast enough between the low and min watermarks. It could
> be mitigated using vm.watermark_scale_factor but not through
> special tricks in kswapd.
> Page writes from reclaim context are at 0 which is the ideal
> Pages immediately reclaimed after IO completes is back at the baseline
>
> On UMA, there is almost no change so this is not expected to be a universal
> win.
>
> Signed-off-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>

[...]

> @@ -3328,6 +3330,22 @@ static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
> return sc.order;
> }
>
> +/*
> + * pgdat->kswapd_classzone_idx is the highest zone index that a recent
> + * allocation request woke kswapd for. When kswapd has not woken recently,
> + * the value is MAX_NR_ZONES which is not a valid index. This compares a
> + * given classzone and returns it or the highest classzone index kswapd
> + * was recently woke for.
> + */
> +static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
> + enum zone_type classzone_idx)
> +{
> + if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
> + return classzone_idx;
> +
> + return max(pgdat->kswapd_classzone_idx, classzone_idx);

A bit paranoid comment: this should probably read pgdat->kswapd_classzone_idx to
a local variable with READ_ONCE(), otherwise something can set it to
MAX_NR_ZONES between the check and max(), and compiler can decide to reread.
Probably not an issue with current callers, but I'd rather future-proof it.

Thanks,
Vlastimil