* [PATCH] mm: Only force scan in reclaim when none of the LRUs are big enough.
@ 2014-03-16 3:36 Hugh Dickins
2014-03-27 20:41 ` Rik van Riel
` (2 more replies)
0 siblings, 3 replies; 4+ messages in thread
From: Hugh Dickins @ 2014-03-16 3:36 UTC (permalink / raw)
To: Johannes Weiner
Cc: Suleiman Souhlal, Andrew Morton, Mel Gorman, Rik van Riel,
Rafael Aquini, Michal Hocko, Yuanhan Liu, Seth Jennings, Bob Liu,
Minchan Kim, Luigi Semenzato, linux-mm, linux-kernel
From: Suleiman Souhlal <suleiman@google.com>
Prior to this change, we would decide whether to force scan a LRU
during reclaim if that LRU itself was too small for the current
priority. However, this can lead to the file LRU getting force
scanned even if there are a lot of anonymous pages we can reclaim,
leading to hot file pages getting needlessly reclaimed.
To address this, we instead only force scan when none of the
reclaimable LRUs are big enough.
Gives huge improvements with zswap. For example, when doing -j20
kernel build in a 500MB container with zswap enabled, runtime (in
seconds) is greatly reduced:
x without this change
+ with this change
N Min Max Median Avg Stddev
x 5 700.997 790.076 763.928 754.05 39.59493
+ 5 141.634 197.899 155.706 161.9 21.270224
Difference at 95.0% confidence
-592.15 +/- 46.3521
-78.5293% +/- 6.14709%
(Student's t, pooled s = 31.7819)
Should also give some improvements in regular (non-zswap) swap cases.
Yes, hughd found significant speedup using regular swap, with several
memcgs under pressure; and it should also be effective in the non-memcg
case, whenever one or another zone LRU is forced too small.
Signed-off-by: Suleiman Souhlal <suleiman@google.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
---
I apologize to everyone for holding on to this so long: I think it's
a very helpful patch (which we've been using in Google for months now).
Been sitting on my TODO list, now prompted to send by related patches
https://lkml.org/lkml/2014/3/13/217
https://lkml.org/lkml/2014/3/14/277
Certainly worth considering all three together, but my understanding
is that they're actually three independent attacks on different ways
in which we currently squeeze an LRU too small; and this patch from
Suleiman seems to be the most valuable of the three, at least for
the workloads I've tried it on. But I'm not much of a page reclaim
performance tester: please try it out to see if it's good for you.
Thanks!
mm/vmscan.c | 72 +++++++++++++++++++++++++++++---------------------
1 file changed, 42 insertions(+), 30 deletions(-)
We did experiment with different ways of writing the patch, I'm afraid
the way it came out best indents deeper, making it look more than it is.
--- 3.14-rc6/mm/vmscan.c 2014-02-02 18:49:07.949302116 -0800
+++ linux/mm/vmscan.c 2014-03-15 19:31:44.948977032 -0700
@@ -1852,6 +1852,8 @@ static void get_scan_count(struct lruvec
bool force_scan = false;
unsigned long ap, fp;
enum lru_list lru;
+ bool some_scanned;
+ int pass;
/*
* If the zone or memcg is small, nr[l] can be 0. This
@@ -1971,39 +1973,49 @@ static void get_scan_count(struct lruvec
fraction[1] = fp;
denominator = ap + fp + 1;
out:
- for_each_evictable_lru(lru) {
- int file = is_file_lru(lru);
- unsigned long size;
- unsigned long scan;
-
- size = get_lru_size(lruvec, lru);
- scan = size >> sc->priority;
-
- if (!scan && force_scan)
- scan = min(size, SWAP_CLUSTER_MAX);
-
- switch (scan_balance) {
- case SCAN_EQUAL:
- /* Scan lists relative to size */
- break;
- case SCAN_FRACT:
+ some_scanned = false;
+ /* Only use force_scan on second pass. */
+ for (pass = 0; !some_scanned && pass < 2; pass++) {
+ for_each_evictable_lru(lru) {
+ int file = is_file_lru(lru);
+ unsigned long size;
+ unsigned long scan;
+
+ size = get_lru_size(lruvec, lru);
+ scan = size >> sc->priority;
+
+ if (!scan && pass && force_scan)
+ scan = min(size, SWAP_CLUSTER_MAX);
+
+ switch (scan_balance) {
+ case SCAN_EQUAL:
+ /* Scan lists relative to size */
+ break;
+ case SCAN_FRACT:
+ /*
+ * Scan types proportional to swappiness and
+ * their relative recent reclaim efficiency.
+ */
+ scan = div64_u64(scan * fraction[file],
+ denominator);
+ break;
+ case SCAN_FILE:
+ case SCAN_ANON:
+ /* Scan one type exclusively */
+ if ((scan_balance == SCAN_FILE) != file)
+ scan = 0;
+ break;
+ default:
+ /* Look ma, no brain */
+ BUG();
+ }
+ nr[lru] = scan;
/*
- * Scan types proportional to swappiness and
- * their relative recent reclaim efficiency.
+ * Skip the second pass and don't force_scan,
+ * if we found something to scan.
*/
- scan = div64_u64(scan * fraction[file], denominator);
- break;
- case SCAN_FILE:
- case SCAN_ANON:
- /* Scan one type exclusively */
- if ((scan_balance == SCAN_FILE) != file)
- scan = 0;
- break;
- default:
- /* Look ma, no brain */
- BUG();
+ some_scanned |= !!scan;
}
- nr[lru] = scan;
}
}
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [PATCH] mm: Only force scan in reclaim when none of the LRUs are big enough.
2014-03-16 3:36 [PATCH] mm: Only force scan in reclaim when none of the LRUs are big enough Hugh Dickins
@ 2014-03-27 20:41 ` Rik van Riel
2014-03-28 18:10 ` Rafael Aquini
2014-04-01 19:49 ` Andrew Morton
2 siblings, 0 replies; 4+ messages in thread
From: Rik van Riel @ 2014-03-27 20:41 UTC (permalink / raw)
To: Hugh Dickins, Johannes Weiner
Cc: Suleiman Souhlal, Andrew Morton, Mel Gorman, Rafael Aquini,
Michal Hocko, Yuanhan Liu, Seth Jennings, Bob Liu, Minchan Kim,
Luigi Semenzato, linux-mm, linux-kernel
On 03/15/2014 11:36 PM, Hugh Dickins wrote:
> From: Suleiman Souhlal <suleiman@google.com>
>
> Prior to this change, we would decide whether to force scan a LRU
> during reclaim if that LRU itself was too small for the current
> priority. However, this can lead to the file LRU getting force
> scanned even if there are a lot of anonymous pages we can reclaim,
> leading to hot file pages getting needlessly reclaimed.
>
> To address this, we instead only force scan when none of the
> reclaimable LRUs are big enough.
>
> Gives huge improvements with zswap. For example, when doing -j20
> kernel build in a 500MB container with zswap enabled, runtime (in
> seconds) is greatly reduced:
>
> x without this change
> + with this change
> N Min Max Median Avg Stddev
> x 5 700.997 790.076 763.928 754.05 39.59493
> + 5 141.634 197.899 155.706 161.9 21.270224
> Difference at 95.0% confidence
> -592.15 +/- 46.3521
> -78.5293% +/- 6.14709%
> (Student's t, pooled s = 31.7819)
>
> Should also give some improvements in regular (non-zswap) swap cases.
>
> Yes, hughd found significant speedup using regular swap, with several
> memcgs under pressure; and it should also be effective in the non-memcg
> case, whenever one or another zone LRU is forced too small.
>
> Signed-off-by: Suleiman Souhlal <suleiman@google.com>
> Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [PATCH] mm: Only force scan in reclaim when none of the LRUs are big enough.
2014-03-16 3:36 [PATCH] mm: Only force scan in reclaim when none of the LRUs are big enough Hugh Dickins
2014-03-27 20:41 ` Rik van Riel
@ 2014-03-28 18:10 ` Rafael Aquini
2014-04-01 19:49 ` Andrew Morton
2 siblings, 0 replies; 4+ messages in thread
From: Rafael Aquini @ 2014-03-28 18:10 UTC (permalink / raw)
To: Hugh Dickins
Cc: Johannes Weiner, Suleiman Souhlal, Andrew Morton, Mel Gorman,
Rik van Riel, Michal Hocko, Yuanhan Liu, Seth Jennings, Bob Liu,
Minchan Kim, Luigi Semenzato, linux-mm, linux-kernel
On Sat, Mar 15, 2014 at 08:36:02PM -0700, Hugh Dickins wrote:
> From: Suleiman Souhlal <suleiman@google.com>
>
> Prior to this change, we would decide whether to force scan a LRU
> during reclaim if that LRU itself was too small for the current
> priority. However, this can lead to the file LRU getting force
> scanned even if there are a lot of anonymous pages we can reclaim,
> leading to hot file pages getting needlessly reclaimed.
>
> To address this, we instead only force scan when none of the
> reclaimable LRUs are big enough.
>
> Gives huge improvements with zswap. For example, when doing -j20
> kernel build in a 500MB container with zswap enabled, runtime (in
> seconds) is greatly reduced:
>
> x without this change
> + with this change
> N Min Max Median Avg Stddev
> x 5 700.997 790.076 763.928 754.05 39.59493
> + 5 141.634 197.899 155.706 161.9 21.270224
> Difference at 95.0% confidence
> -592.15 +/- 46.3521
> -78.5293% +/- 6.14709%
> (Student's t, pooled s = 31.7819)
>
> Should also give some improvements in regular (non-zswap) swap cases.
>
> Yes, hughd found significant speedup using regular swap, with several
> memcgs under pressure; and it should also be effective in the non-memcg
> case, whenever one or another zone LRU is forced too small.
>
> Signed-off-by: Suleiman Souhlal <suleiman@google.com>
> Signed-off-by: Hugh Dickins <hughd@google.com>
> ---
>
Acked-by: Rafael Aquini <aquini@redhat.com>
> I apologize to everyone for holding on to this so long: I think it's
> a very helpful patch (which we've been using in Google for months now).
> Been sitting on my TODO list, now prompted to send by related patches
>
> https://lkml.org/lkml/2014/3/13/217
> https://lkml.org/lkml/2014/3/14/277
>
> Certainly worth considering all three together, but my understanding
> is that they're actually three independent attacks on different ways
> in which we currently squeeze an LRU too small; and this patch from
> Suleiman seems to be the most valuable of the three, at least for
> the workloads I've tried it on. But I'm not much of a page reclaim
> performance tester: please try it out to see if it's good for you.
> Thanks!
>
> mm/vmscan.c | 72 +++++++++++++++++++++++++++++---------------------
> 1 file changed, 42 insertions(+), 30 deletions(-)
>
> We did experiment with different ways of writing the patch, I'm afraid
> the way it came out best indents deeper, making it look more than it is.
>
> --- 3.14-rc6/mm/vmscan.c 2014-02-02 18:49:07.949302116 -0800
> +++ linux/mm/vmscan.c 2014-03-15 19:31:44.948977032 -0700
> @@ -1852,6 +1852,8 @@ static void get_scan_count(struct lruvec
> bool force_scan = false;
> unsigned long ap, fp;
> enum lru_list lru;
> + bool some_scanned;
> + int pass;
>
> /*
> * If the zone or memcg is small, nr[l] can be 0. This
> @@ -1971,39 +1973,49 @@ static void get_scan_count(struct lruvec
> fraction[1] = fp;
> denominator = ap + fp + 1;
> out:
> - for_each_evictable_lru(lru) {
> - int file = is_file_lru(lru);
> - unsigned long size;
> - unsigned long scan;
> -
> - size = get_lru_size(lruvec, lru);
> - scan = size >> sc->priority;
> -
> - if (!scan && force_scan)
> - scan = min(size, SWAP_CLUSTER_MAX);
> -
> - switch (scan_balance) {
> - case SCAN_EQUAL:
> - /* Scan lists relative to size */
> - break;
> - case SCAN_FRACT:
> + some_scanned = false;
> + /* Only use force_scan on second pass. */
> + for (pass = 0; !some_scanned && pass < 2; pass++) {
> + for_each_evictable_lru(lru) {
> + int file = is_file_lru(lru);
> + unsigned long size;
> + unsigned long scan;
> +
> + size = get_lru_size(lruvec, lru);
> + scan = size >> sc->priority;
> +
> + if (!scan && pass && force_scan)
> + scan = min(size, SWAP_CLUSTER_MAX);
> +
> + switch (scan_balance) {
> + case SCAN_EQUAL:
> + /* Scan lists relative to size */
> + break;
> + case SCAN_FRACT:
> + /*
> + * Scan types proportional to swappiness and
> + * their relative recent reclaim efficiency.
> + */
> + scan = div64_u64(scan * fraction[file],
> + denominator);
> + break;
> + case SCAN_FILE:
> + case SCAN_ANON:
> + /* Scan one type exclusively */
> + if ((scan_balance == SCAN_FILE) != file)
> + scan = 0;
> + break;
> + default:
> + /* Look ma, no brain */
> + BUG();
> + }
> + nr[lru] = scan;
> /*
> - * Scan types proportional to swappiness and
> - * their relative recent reclaim efficiency.
> + * Skip the second pass and don't force_scan,
> + * if we found something to scan.
> */
> - scan = div64_u64(scan * fraction[file], denominator);
> - break;
> - case SCAN_FILE:
> - case SCAN_ANON:
> - /* Scan one type exclusively */
> - if ((scan_balance == SCAN_FILE) != file)
> - scan = 0;
> - break;
> - default:
> - /* Look ma, no brain */
> - BUG();
> + some_scanned |= !!scan;
> }
> - nr[lru] = scan;
> }
> }
>
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [PATCH] mm: Only force scan in reclaim when none of the LRUs are big enough.
2014-03-16 3:36 [PATCH] mm: Only force scan in reclaim when none of the LRUs are big enough Hugh Dickins
2014-03-27 20:41 ` Rik van Riel
2014-03-28 18:10 ` Rafael Aquini
@ 2014-04-01 19:49 ` Andrew Morton
2 siblings, 0 replies; 4+ messages in thread
From: Andrew Morton @ 2014-04-01 19:49 UTC (permalink / raw)
To: Hugh Dickins
Cc: Johannes Weiner, Suleiman Souhlal, Mel Gorman, Rik van Riel,
Rafael Aquini, Michal Hocko, Yuanhan Liu, Seth Jennings, Bob Liu,
Minchan Kim, Luigi Semenzato, linux-mm, linux-kernel
On Sat, 15 Mar 2014 20:36:02 -0700 (PDT) Hugh Dickins <hughd@google.com> wrote:
> From: Suleiman Souhlal <suleiman@google.com>
>
> Prior to this change, we would decide whether to force scan a LRU
> during reclaim if that LRU itself was too small for the current
> priority. However, this can lead to the file LRU getting force
> scanned even if there are a lot of anonymous pages we can reclaim,
> leading to hot file pages getting needlessly reclaimed.
Struggling a bit here. You're referring to this code?
size = get_lru_size(lruvec, lru);
scan = size >> sc->priority;
if (!scan && force_scan)
scan = min(size, SWAP_CLUSTER_MAX);
So we're talking about the case where the LRU is so small that it
contains fewer than (1<<sc->priority) pages?
If so, then I'd expect that in normal operation this situation rarely
occurs? Surely the LRUs normally contain many more pages than this.
> To address this, we instead only force scan when none of the
> reclaimable LRUs are big enough.
>
> Gives huge improvements with zswap. For example, when doing -j20
> kernel build in a 500MB container with zswap enabled, runtime (in
> seconds) is greatly reduced:
>
> x without this change
> + with this change
> N Min Max Median Avg Stddev
> x 5 700.997 790.076 763.928 754.05 39.59493
> + 5 141.634 197.899 155.706 161.9 21.270224
> Difference at 95.0% confidence
> -592.15 +/- 46.3521
> -78.5293% +/- 6.14709%
> (Student's t, pooled s = 31.7819)
And yet the patch makes a large difference. What am I missing here?
> --- 3.14-rc6/mm/vmscan.c 2014-02-02 18:49:07.949302116 -0800
> +++ linux/mm/vmscan.c 2014-03-15 19:31:44.948977032 -0700
> @@ -1971,39 +1973,49 @@ static void get_scan_count(struct lruvec
> fraction[1] = fp;
> denominator = ap + fp + 1;
> out:
> - for_each_evictable_lru(lru) {
> - int file = is_file_lru(lru);
> - unsigned long size;
> - unsigned long scan;
> -
> - size = get_lru_size(lruvec, lru);
> - scan = size >> sc->priority;
> -
> - if (!scan && force_scan)
> - scan = min(size, SWAP_CLUSTER_MAX);
> -
> - switch (scan_balance) {
> - case SCAN_EQUAL:
> - /* Scan lists relative to size */
> - break;
> - case SCAN_FRACT:
> + some_scanned = false;
> + /* Only use force_scan on second pass. */
That's a poor comment.
> + for (pass = 0; !some_scanned && pass < 2; pass++) {
> + for_each_evictable_lru(lru) {
> + int file = is_file_lru(lru);
> + unsigned long size;
> + unsigned long scan;
> +
> + size = get_lru_size(lruvec, lru);
> + scan = size >> sc->priority;
> +
> + if (!scan && pass && force_scan)
> + scan = min(size, SWAP_CLUSTER_MAX);
> +
> + switch (scan_balance) {
> + case SCAN_EQUAL:
> + /* Scan lists relative to size */
> + break;
> + case SCAN_FRACT:
> + /*
> + * Scan types proportional to swappiness and
> + * their relative recent reclaim efficiency.
> + */
> + scan = div64_u64(scan * fraction[file],
> + denominator);
> + break;
> + case SCAN_FILE:
> + case SCAN_ANON:
> + /* Scan one type exclusively */
> + if ((scan_balance == SCAN_FILE) != file)
> + scan = 0;
> + break;
> + default:
> + /* Look ma, no brain */
> + BUG();
> + }
> + nr[lru] = scan;
> /*
> - * Scan types proportional to swappiness and
> - * their relative recent reclaim efficiency.
> + * Skip the second pass and don't force_scan,
> + * if we found something to scan.
And so is that. Both comments explain *what* the code is doing (which
was fairly obvious from the code!) but they fail to explain *why* the
code is doing what it does.
> */
> - scan = div64_u64(scan * fraction[file], denominator);
> - break;
> - case SCAN_FILE:
> - case SCAN_ANON:
> - /* Scan one type exclusively */
> - if ((scan_balance == SCAN_FILE) != file)
> - scan = 0;
> - break;
> - default:
> - /* Look ma, no brain */
> - BUG();
> + some_scanned |= !!scan;
Also the "and don't force_scan" part appears to be flatly untrue. Either
the comment is wrong or the code should be along the lines of
if (scan) {
some_scanned = true;
force_scan = false;
}
Can we fix these things please? And retest if necessary.
> }
> - nr[lru] = scan;
> }
> }
>
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2014-04-01 19:49 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2014-03-16 3:36 [PATCH] mm: Only force scan in reclaim when none of the LRUs are big enough Hugh Dickins
2014-03-27 20:41 ` Rik van Riel
2014-03-28 18:10 ` Rafael Aquini
2014-04-01 19:49 ` Andrew Morton
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).