mm: fix low-high watermark distance on small systems

It is observed that watermark_scale_factor when used to reduce thundering herds in direct reclaim, reduces the direct reclaims, but results in unnecessary reclaim due to kswapd running for long after being woken up. The tests are done with 4 GB of RAM and the tests done are multibuild and another which opens a set of apps sequentially on Android and repeating the sequence N times. The tests are done on 4.9 kernel.

The issue is caused by watermark_scale_factor creating larger than required gap between low and high watermarks. The following results are with watermark_scale_factor of 120.

                       wsf-120-default  wsf-120-reduced-low-high-gap
workingset_activate    15120206         8319182
pgpgin                 269795482        147928581
allocstall             1406             1498
pgsteal_kswapd         68676960         38105142
slabs_scanned          94181738         49085755

Even though kswapd does 80% more steal in the default case, it doesn't make any significant improvement to the direct reclaims. The excessive reclaims cause more pgpgin and increases app launch latencies.

The min-low gap is untouched by the patch. The low-high gap is made a percentage of min-low gap. The fraction was derived considering these,

1) The existing watermark_scale_factor logic was designed to fix issues on high RAM machines and I assume that the current low-high gap works well on those.
2) The gap should be reduced on low RAM targets where thrashing is observed.

The multiplier 4 was chosen empirically which was seen to fix the thrashing on <8GB devices.

With watermark_scale_factor as default 10, the low-high gap for different memory sizes.

           default-low_high_gap-pages  this-patch-low_high_gap-pages
16M        4                           4
512M       131                         131
1024M      262                         256
2048M      524                         362
4096M      1048                        512
8192M      2097                        724
16384M     4194                        1717
32768M     8388                        4858
65536M     16777                       13743
102400M    26214                       26214
143360M    36700                       36700

Signed-off-by Vinayak Menon <[email protected]>

---
 mm/page_alloc.c | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

mm/page_alloc.c +14 -1

index c0f0b1b..ac75985 100644
@@ -7223,7 +7223,20 @@ static void __setup_per_zone_wmarks(void)
 				      watermark_scale_factor, 10000));
 
 		zone->watermark[WMARK_LOW]  = min_wmark_pages(zone) + tmp;
-		zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
+
+		/*
+		 * Set the kswapd low-high distance as a percentage of
+		 * min-low distance in such a way that the distance
+		 * increases non-linearly with available memory. This
+		 * ensures enough free memory without causing thrashing
+		 * on small machines due to excessive reclaim by kswapd.
+		 * Ensure a minimum distance on very small machines.
+		 */
+		tmp = clamp_t(u64, mult_frac(tmp,
+				int_sqrt(4 * zone->managed_pages), 10000),
+					min_wmark_pages(zone) >> 2, tmp);
+
+		zone->watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
 
 		spin_unlock_irqrestore(&zone->lock, flags);
 	}

Signature · 2 lines

QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a
member of the Code Aurora Forum, hosted by The Linux Foundation

Hi Vinayak,

Quoted 12 lines

It is observed that watermark_scale_factor when used to reduce thundering herds in direct reclaim, reduces the direct reclaims, but results in unnecessary reclaim due to kswapd running for long after being woken up. The tests are done with 4 GB of RAM and the tests done are multibuild and another which opens a set of apps sequentially on Android and repeating the sequence N times. The tests are done on 4.9 kernel.

The issue is caused by watermark_scale_factor creating larger than required gap between low and high watermarks. The following results are with watermark_scale_factor of 120.

Quoted 6 lines

                       wsf-120-default  wsf-120-reduced-low-high-gap
workingset_activate    15120206         8319182
pgpgin                 269795482        147928581
allocstall             1406             1498
pgsteal_kswapd         68676960         38105142
slabs_scanned          94181738         49085755

"required gap" you mentiond is very dependent for your workload. You had an experiment with wsf-120. It means user wanted to be more aggressive for kswapd while your load is not enough to make meomry consumption spike. Couldn't you decrease wfs?

Don't get me wrong. I don't want you test all of wfs with varios workload to prove your logic is better. What I want to say here is it's heuristic so it couldn't be perfect for every workload so if you change to non-linear, you could be better but others might be not.

In such context, current linear linear scale factor is simple enough to understand. IMO, if we really want to enhance watermark, the low/high wmark shold be adaptable according to memory spike. One of rough idea is to change low/high wmark based on kswapd_[high|low]_wmark_hit_quickly.

Quoted 67 lines

Even though kswapd does 80% more steal in the default case, it doesn't make any significant improvement to the direct reclaims. The excessive reclaims cause more pgpgin and increases app launch latencies.

The min-low gap is untouched by the patch. The low-high gap is made a percentage of min-low gap. The fraction was derived considering these,

1) The existing watermark_scale_factor logic was designed to fix issues on high RAM machines and I assume that the current low-high gap works well on those.
2) The gap should be reduced on low RAM targets where thrashing is observed.

The multiplier 4 was chosen empirically which was seen to fix the thrashing on <8GB devices.

With watermark_scale_factor as default 10, the low-high gap for different memory sizes.

           default-low_high_gap-pages  this-patch-low_high_gap-pages
16M        4                           4
512M       131                         131
1024M      262                         256
2048M      524                         362
4096M      1048                        512
8192M      2097                        724
16384M     4194                        1717
32768M     8388                        4858
65536M     16777                       13743
102400M    26214                       26214
143360M    36700                       36700

Signed-off-by Vinayak Menon <[email protected]>

---
 mm/page_alloc.c | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

mm/page_alloc.c +14 -1

index c0f0b1b..ac75985 100644
@@ -7223,7 +7223,20 @@ static void __setup_per_zone_wmarks(void)
 				      watermark_scale_factor, 10000));
 
 		zone->watermark[WMARK_LOW]  = min_wmark_pages(zone) + tmp;
-		zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
+
+		/*
+		 * Set the kswapd low-high distance as a percentage of
+		 * min-low distance in such a way that the distance
+		 * increases non-linearly with available memory. This
+		 * ensures enough free memory without causing thrashing
+		 * on small machines due to excessive reclaim by kswapd.
+		 * Ensure a minimum distance on very small machines.
+		 */
+		tmp = clamp_t(u64, mult_frac(tmp,
+				int_sqrt(4 * zone->managed_pages), 10000),
+					min_wmark_pages(zone) >> 2, tmp);
+
+		zone->watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
 
 		spin_unlock_irqrestore(&zone->lock, flags);
 	}

Signature · 2 lines

QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a
member of the Code Aurora Forum, hosted by The Linux Foundation

Quoted 1 line · Minchan Kim

Hi Vinayak,

Thanks for your comments Minchan.

Quoted 20 lines · Vinayak Menon

Quoted 12 lines

It is observed that watermark_scale_factor when used to reduce thundering herds in direct reclaim, reduces the direct reclaims, but results in unnecessary reclaim due to kswapd running for long after being woken up. The tests are done with 4 GB of RAM and the tests done are multibuild and another which opens a set of apps sequentially on Android and repeating the sequence N times. The tests are done on 4.9 kernel.

The issue is caused by watermark_scale_factor creating larger than required gap between low and high watermarks. The following results are with watermark_scale_factor of 120.

Quoted 6 lines

                       wsf-120-default  wsf-120-reduced-low-high-gap
workingset_activate    15120206         8319182
pgpgin                 269795482        147928581
allocstall             1406             1498
pgsteal_kswapd         68676960         38105142
slabs_scanned          94181738         49085755

Quoted 4 lines

"required gap" you mentiond is very dependent for your workload. You had an experiment with wsf-120. It means user wanted to be more aggressive for kswapd while your load is not enough to make meomry consumption spike. Couldn't you decrease wfs?

I did try reducing the wsf for both multibuild and Android workloads. But that results in kswapd waking up late and thus latency issues due to higher direct reclaims. As I understand the problem, the wsf in its current form helps in tuning the kswapd wakeups (and note that I have not touched the wsf logic to calculate min-low gap), but the issue arises due to the depth to which kswapd scans the LRUs in a single run, causing thrashing, due to the higher low-high gap. From experiments, it looks like kswapd waking

up few more times and doing shorter steals is better than kswapd stealing more in a single run. The latter does not better direct reclaims and causes thrashing too.

Quoted 6 lines

Don't get me wrong. I don't want you test all of wfs with varios workload to prove your logic is better. What I want to say here is it's heuristic so it couldn't be perfect for every workload so if you change to non-linear, you could be better but others might be not.

Yes I understand your point. But since mmtests and Android tests showed similar results, I thought the heuristic may just work across workloads. I assume from Johannes's tests on 140GB machine (from the commit msg of the patch which introduced wsf) that the current low-high gap works well without thrashing on bigger machines. This made me assume that the behavior is non-linear. So the non-linear behavior will not make any difference to higher RAM machines as the low-high remains almost same as shown in the table below. But I understand your point, for a different workload on smaller machines, I am not sure the benefit I see would be observed, though that's the same problem with current wsf too.

Quoted 6 lines

In such context, current linear linear scale factor is simple enough to understand. IMO, if we really want to enhance watermark, the low/high wmark shold be adaptable according to memory spike. One of rough idea is to change low/high wmark based on kswapd_[high|low]_wmark_hit_quickly.

That seems like a nice idea to me. But considering the current case with and without this patch, the kswapd_low_wmark_hit quickly is actually less without the patch. But that comes at the cost of thrashing. Which then would mean we need to detect thrashing to adjust the watermark. We may get the thrashing data from workingset refaults, but I am not sure if we can take it as an input to adjust watermark since thrashing can be due to other reasons too. Maybe there are ways to make this adaptive, like using Johannes's memdelay feature to detect more time spent in direct reclaims and then raise the low watermark, and then use time spent in refault and stabilized direct reclaim time to bring down or stop raising the low watermark.
But do we need that adaptive logic now if this (or other similar) patch just works across workloads for small machines ? What is your suggestion ?

Quoted 26 lines · Vinayak Menon

Quoted 1 line · Minchan Kim

Hi Vinayak,

Thanks for your comments Minchan.

Quoted 19 lines · Vinayak Menon

Quoted 12 lines

It is observed that watermark_scale_factor when used to reduce thundering herds in direct reclaim, reduces the direct reclaims, but results in unnecessary reclaim due to kswapd running for long after being woken up. The tests are done with 4 GB of RAM and the tests done are multibuild and another which opens a set of apps sequentially on Android and repeating the sequence N times. The tests are done on 4.9 kernel.

The issue is caused by watermark_scale_factor creating larger than required gap between low and high watermarks. The following results are with watermark_scale_factor of 120.

Quoted 6 lines

                       wsf-120-default  wsf-120-reduced-low-high-gap
workingset_activate    15120206         8319182
pgpgin                 269795482        147928581
allocstall             1406             1498
pgsteal_kswapd         68676960         38105142
slabs_scanned          94181738         49085755

Quoted 4 lines

"required gap" you mentiond is very dependent for your workload. You had an experiment with wsf-120. It means user wanted to be more aggressive for kswapd while your load is not enough to make meomry consumption spike. Couldn't you decrease wfs?

Quoted 5 lines

I did try reducing the wsf for both multibuild and Android workloads. But that results in kswapd waking up late and thus latency issues due to higher direct reclaims. As I understand the problem, the wsf in its current form helps in tuning the kswapd wakeups (and note that I have not touched the wsf logic to calculate min-low gap), but the issue arises due to the depth to which kswapd scans the LRUs in a single run, causing thrashing, due to the higher low-high gap. From experiments, it looks like kswapd waking

wsf conducts kswapd sleep time as well as wakeup time.

"This factor controls the aggressiveness of kswapd. It defines the amount of memory left in a node/system before kswapd is woken up and how much memory needs to be free before kswapd goes back to sleep."

Quoted 3 lines

up few more times and doing shorter steals is better than kswapd stealing more in a single run. The latter does not better direct reclaims and causes thrashing too.

That's the tradeoff of kswapd aggressiveness to avoid high rate direct reclaim.

Quoted 11 lines

Quoted 4 lines

Don't get me wrong. I don't want you test all of wfs with varios workload to prove your logic is better. What I want to say here is it's heuristic so it couldn't be perfect for every workload so if you change to non-linear, you could be better but others might be not.

Yes I understand your point. But since mmtests and Android tests showed similar results, I thought the heuristic may just work across workloads. I assume from Johannes's tests on 140GB machine (from the commit msg of the patch which introduced wsf) that the current low-high gap works well without thrashing on bigger machines. This made me assume that the behavior is non-linear. So the non-linear behavior will not make any difference to higher RAM machines as the low-high remains almost same as shown in the table below. But I understand your point, for a different workload on smaller machines, I am not sure the benefit I see would be observed, though that's the same problem with current wsf too.

True. That's why I don't want to make it complicate. Later, if someone complains "linear is better for his several testing", are you happy to rollback to it?

You might argue it's same problem now but at least as-is code is simple to understand.

Quoted 6 lines

Quoted 3 lines

up few more times and doing shorter steals is better than kswapd stealing more in a single run. The latter does not better direct reclaims and causes thrashing too.

That's the tradeoff of kswapd aggressiveness to avoid high rate direct reclaim.

We can call it trade off only if increasing the aggressiveness of kswapd reduces the direct reclaims ? But as shown by the data I had shared, the aggressiveness does not improve direct reclaims. It is just causing unnecessary reclaim. i.e. a much lower low-high gap gives the same benefit on direct reclaims with far less reclaim.

Quoted 12 lines

Quoted 12 lines

Quoted 5 lines

Don't get me wrong. I don't want you test all of wfs with varios workload to prove your logic is better. What I want to say here is it's heuristic so it couldn't be perfect for every workload so if you change to non-linear, you could be better but others might be not.

Yes I understand your point. But since mmtests and Android tests showed similar results, I thought the heuristic may just work across workloads. I assume from Johannes's tests on 140GB machine (from the commit msg of the patch which introduced wsf) that the current low-high gap works well without thrashing on bigger machines. This made me assume that the behavior is non-linear. So the non-linear behavior will not make any difference to higher RAM machines as the low-high remains almost same as shown in the table below. But I understand your point, for a different workload on smaller machines, I am not sure the benefit I see would be observed, though that's the same problem with current wsf too.

Quoted 5 lines

True. That's why I don't want to make it complicate. Later, if someone complains "linear is better for his several testing", are you happy to rollback to it?

You might argue it's same problem now but at least as-is code is simple to understand.

Yes I agree that there can be workloads on low RAM that may see a side effect. But since popular use case like those on Android

and also the mmtests shows the problem, and fixed by the patch, can we try to pick it and see if someone complains ? I see that there were other reports of this https://lkml.org/lkml/2017/11/24/167 . Do you suggest a tunable approach taken by the patch in that link ? So that varying use cases can be accommodated. I wanted to avoid a new tunable if some heuristic like the patch does just works.

Quoted 8 lines · Vinayak Menon

Quoted 4 lines

Quoted 2 lines

up few more times and doing shorter steals is better than kswapd stealing more in a single run. The latter does not better direct reclaims and causes thrashing too.

That's the tradeoff of kswapd aggressiveness to avoid high rate direct reclaim.

We can call it trade off only if increasing the aggressiveness of kswapd reduces the direct reclaims ? But as shown by the data I had shared, the aggressiveness does not improve direct reclaims. It is just causing unnecessary reclaim. i.e. a much lower low-high gap gives the same benefit on direct reclaims with far less reclaim.

Said again, it depends on workload. I can make simple test to break it easily.

Quoted 1 line

I don't think repeated app launching on android doesn't reflect real user scenario. Anyone don't do that in real life except some guys want to show benchmark result in youtube.

About mmtests, what kinds of tests did you perform? So what's the result? If you reduced thrashing, how much the test result is improved? Every tests are improved? Need not vmstat but result from the benchmark. Such wide testing would make more conviction.

Quoted 4 lines

and also the mmtests shows the problem, and fixed by the patch, can we try to pick it and see if someone complains ? I see that there were other reports of this https://lkml.org/lkml/2017/11/24/167 . Do you suggest a tunable approach taken by the patch in that link ? So that varying use cases can be accommodated. I wanted to avoid a new tunable if some heuristic like the patch does just works.

Actually, I don't want to touch it unless we have more nice feedback algorithm.

Anyway, it's just my opinion. I did best effort to explain. I will defer to maintainer.

Thanks.