mm: Proactive compaction

For some applications, we need to allocate almost all memory as hugepages. However, on a running system, higher-order allocations can fail if the memory is fragmented. Linux kernel currently does on-demand compaction as we request more hugepages, but this style of compaction incurs very high latency. Experiments with one-time full memory compaction (followed by hugepage allocations) show that kernel is able to restore a highly fragmented memory state to a fairly compacted memory state within <1 sec for a 32G system. Such data suggests that a more proactive compaction can help us allocate a large fraction of memory as hugepages keeping allocation latencies low.

For a more proactive compaction, the approach taken here is to define a new tunable called 'proactiveness' which dictates bounds for external fragmentation wrt HUGETLB_PAGE_ORDER order which kcompactd tries to maintain.

The tunable is exposed through sysfs:
/sys/kernel/mm/compaction/proactiveness

<snip>

For some applications, we need to allocate almost all memory as hugepages. However, on a running system, higher-order allocations can fail if the memory is fragmented. Linux kernel currently does on-demand compaction as we request more hugepages, but this style of compaction incurs very high latency. Experiments with one-time full memory compaction (followed by hugepage allocations) show that kernel is able to restore a highly fragmented memory state to a fairly compacted memory state within <1 sec for a 32G system. Such data suggests that a more proactive compaction can help us allocate a large fraction of memory as hugepages keeping allocation latencies low.

For a more proactive compaction, the approach taken here is to define a new tunable called 'proactiveness' which dictates bounds for external fragmentation wrt HUGETLB_PAGE_ORDER order which kcompactd tries to maintain.

The tunable is exposed through sysfs:
/sys/kernel/mm/compaction/proactiveness

It takes value in range [0, 100], with a default of 20.

Note that a previous version of this patch [1] was found to introduce too many tunables (per-order extfrag{low, high}), but this one reduces them to just one (proactiveness). Also, the new tunable is an opaque value instead of asking for specific bounds of "external fragmentation", which would have been difficult to estimate. The internal interpretation of this opaque value allows for future fine-tuning.

Currently, we use a simple translation from this tunable to [low, high] "fragmentation score" thresholds (low=100-proactiveness, high=low+10%). The score for a node is defined as weighted mean of per-zone external fragmentation wrt HUGETLB_PAGE_ORDER order. A zone's present_pages determines its weight.

To periodically check per-node score, we reuse per-node kcompactd threads, which are woken up every 500 milliseconds to check the same. If a node's score exceeds its high threshold (as derived from user-provided proactiveness value), proactive compaction is started until its score reaches its low threshold value. By default, proactiveness is set to 20, which implies threshold values of low=80 and high=90.

This patch is largely based on ideas from Michal Hocko posted here:
https://lore.kernel.org/linux-mm/[email protected]/

Performance data
================

System: x64_64, 1T RAM, 80 CPU threads.
Kernel: 5.6.0-rc3 + this patch

echo madvise | sudo tee /sys/kernel/mm/transparent_hugepage/enabled echo madvise | sudo tee /sys/kernel/mm/transparent_hugepage/defrag

Before starting the driver, the system was fragmented from a userspace program that allocates all memory and then for each 2M aligned section, frees 3/4 of base pages using munmap. The workload is mainly anonymous userspace pages, which are easy to move around. I intentionally avoided unmovable pages in this test to see how much latency we incur when hugepage allocations hit direct compaction.

1. Kernel hugepage allocation latencies

With the system in such a fragmented state, a kernel driver then allocates as many hugepages as possible and measures allocation latency:

(all latency values are in microseconds)

- With vanilla 5.6.0-rc3

echo 0 | sudo tee /sys/kernel/mm/compaction/node-*/proactiveness

percentile latency
–––––––––– –––––––

	   5    7894
	  10    9496
	  25   12561
	  30   15295
	  40   18244
	  50   21229
	  60   27556
	  75   30147
	  80   31047
	  90   32859
	  95   33799

Total 2M hugepages allocated = 383859 (749G worth of hugepages out of 762G total free => 98% of free memory could be allocated as hugepages)

- With 5.6.0-rc3 + this patch, with proactiveness=20

echo 20 | sudo tee /sys/kernel/mm/compaction/node-*/proactiveness

percentile latency
–––––––––– –––––––

	   5       2
	  10       2
	  25       3
	  30       3
	  40       3
	  50       4
	  60       4
	  75       4
	  80       4
	  90       5
	  95     429

Total 2M hugepages allocated = 384105 (750G worth of hugepages out of 762G total free => 98% of free memory could be allocated as hugepages)

2. JAVA heap allocation

In this test, we first fragment memory using the same method as for (1).

Then, we start a Java process with a heap size set to 700G and request the heap to be allocated with THP hugepages. We also set THP to madvise to allow hugepage backing of this heap.

/usr/bin/time
java -Xms700G -Xmx700G -XX:+UseTransparentHugePages -XX:+AlwaysPreTouch

The above command allocates 700G of Java heap using hugepages.

- With vanilla 5.6.0-rc3

17.39user 1666.48system 27:37.89elapsed

- With 5.6.0-rc3 + this patch, with proactiveness=20

8.35user 194.58system 3:19.62elapsed

Elapsed time remains around 3:15, as proactiveness is further increased.

Note that proactive compaction happens throughout the runtime of these workloads. The situation of one-time compaction, sufficient to supply hugepages for following allocation stream, can probably happen for more extreme proactiveness values, like 80 or 90.

In the above Java workload, proactiveness is set to 20. The test starts with a node's score of 80 or higher, depending on the delay between the fragmentation step and starting the benchmark, which gives more-or-less time for the initial round of compaction. As the benchmark consumes hugepages, node's score quickly rises above the high threshold (90) and proactive compaction starts again, which brings down the score to the low threshold level (80). Repeat.

bpftrace also confirms proactive compaction running 20+ times during the runtime of this Java benchmark. kcompactd threads consume 100% of one of the CPUs while it tries to bring a node's score within thresholds.

Backoff behavior
================

Above workloads produce a memory state which is easy to compact. However, if memory is filled with unmovable pages, proactive compaction should essentially back off. To test this aspect:

- Created a kernel driver that allocates almost all memory as hugepages followed by freeing first 3/4 of each hugepage.
- Set proactiveness=40
- Note that proactive_compact_node() is deferred maximum number of times with HPAGE_FRAG_CHECK_INTERVAL_MSEC of wait between each check (=> ~30 seconds between retries).

[1] https://patchwork.kernel.org/patch/11098289/

Signed-off-by Nitin Gupta <[email protected]>

To: Mel Gorman <[email protected]>

mm/compaction.c +7 -0

+
+/*
+ * A zone's fragmentation score is the external fragmentation wrt to the
+ * HUGETLB_PAGE_ORDER scaled by the zone's size. It returns a value in the
+ * range [0, 100].
+
+ * The scaling factor ensures that proactive compaction focuses on larger

mm/compaction.c +12 -0

 		.whole_zone = (prio == MIN_COMPACT_PRIORITY),
 		.ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
 		.ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)

@@ -2412,6 +2515,42 @@ enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
 	return rc;
 }
 
 

@@ -2500,6 +2640,63 @@ void compaction_unregister_node(struct node *node)
 }
 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
 
+#ifdef CONFIG_SYSFS
+
+#define COMPACTION_ATTR_RO(_name) \
+	static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
+
+#define COMPACTION_ATTR(_name) \
+	static struct kobj_attribute _name##_attr = \
+		__ATTR(_name, 0644, _name##_show, _name##_store)
+
+static struct kobject *compaction_kobj;
+
+static ssize_t proactiveness_store(struct kobject *kobj,

Quoted 11 lines · Nitin Gupta

For some applications, we need to allocate almost all memory as hugepages. However, on a running system, higher-order allocations can fail if the memory is fragmented. Linux kernel currently does on-demand compaction as we request more hugepages, but this style of compaction incurs very high latency. Experiments with one-time full memory compaction (followed by hugepage allocations) show that kernel is able to restore a highly fragmented memory state to a fairly compacted memory state within <1 sec for a 32G system. Such data suggests that a more proactive compaction can help us allocate a large fraction of memory as hugepages keeping allocation latencies low.

Quoted 7 lines

For a more proactive compaction, the approach taken here is to define a new tunable called 'proactiveness' which dictates bounds for external fragmentation wrt HUGETLB_PAGE_ORDER order which kcompactd tries to maintain.

The tunable is exposed through sysfs:
/sys/kernel/mm/compaction/proactiveness

I would prefer sysctl. Why?

During the mm evolution we seem to have end up with stuff scattered over several
places:

/proc/sys aka sysctl:
/proc/sys/vm/compact_unevictable_allowed /proc/sys/vm/compact_memory - write-only one-time action trigger!

/sys/kernel/mm:
e.g. /sys/kernel/mm/transparent_hugepage/

This is unfortunate enough, and (influenced by my recent dive into sysctl perhaps :), I would have preferred sysctl only. In this case it's consistent
that we have sysctls for compaction already, while this introduces a whole new
compaction directory in the /sys/kernel/mm/ space.

Quoted 41 lines

It takes value in range [0, 100], with a default of 20.

Note that a previous version of this patch [1] was found to introduce too many tunables (per-order extfrag{low, high}), but this one reduces them to just one (proactiveness). Also, the new tunable is an opaque value instead of asking for specific bounds of "external fragmentation", which would have been difficult to estimate. The internal interpretation of this opaque value allows for future fine-tuning.

Currently, we use a simple translation from this tunable to [low, high] "fragmentation score" thresholds (low=100-proactiveness, high=low+10%). The score for a node is defined as weighted mean of per-zone external fragmentation wrt HUGETLB_PAGE_ORDER order. A zone's present_pages determines its weight.

To periodically check per-node score, we reuse per-node kcompactd threads, which are woken up every 500 milliseconds to check the same. If a node's score exceeds its high threshold (as derived from user-provided proactiveness value), proactive compaction is started until its score reaches its low threshold value. By default, proactiveness is set to 20, which implies threshold values of low=80 and high=90.

This patch is largely based on ideas from Michal Hocko posted here:
https://lore.kernel.org/linux-mm/[email protected]/

Performance data
================

System: x64_64, 1T RAM, 80 CPU threads.
Kernel: 5.6.0-rc3 + this patch

echo madvise | sudo tee /sys/kernel/mm/transparent_hugepage/enabled echo madvise | sudo tee /sys/kernel/mm/transparent_hugepage/defrag

Before starting the driver, the system was fragmented from a userspace program that allocates all memory and then for each 2M aligned section, frees 3/4 of base pages using munmap. The workload is mainly anonymous userspace pages, which are easy to move around. I intentionally avoided unmovable pages in this test to see how much latency we incur when hugepage allocations hit direct compaction.

Quoted 3 lines

1. Kernel hugepage allocation latencies

With the system in such a fragmented state, a kernel driver then

allocates

Quoted 63 lines

as many hugepages as possible and measures allocation latency:

(all latency values are in microseconds)

- With vanilla 5.6.0-rc3

echo 0 | sudo tee /sys/kernel/mm/compaction/node-*/proactiveness

percentile latency
–––––––––– –––––––

         5    7894
        10    9496
        25   12561
        30   15295
        40   18244
        50   21229
        60   27556
        75   30147
        80   31047
        90   32859
        95   33799

Total 2M hugepages allocated = 383859 (749G worth of hugepages out of 762G total free => 98% of free memory could be allocated as hugepages)

- With 5.6.0-rc3 + this patch, with proactiveness=20

echo 20 | sudo tee /sys/kernel/mm/compaction/node-*/proactiveness

percentile latency
–––––––––– –––––––

         5       2
        10       2
        25       3
        30       3
        40       3
        50       4
        60       4
        75       4
        80       4
        90       5
        95     429

Total 2M hugepages allocated = 384105 (750G worth of hugepages out of 762G total free => 98% of free memory could be allocated as hugepages)

2. JAVA heap allocation

In this test, we first fragment memory using the same method as for (1).

Then, we start a Java process with a heap size set to 700G and request the heap to be allocated with THP hugepages. We also set THP to madvise to allow hugepage backing of this heap.

/usr/bin/time
java -Xms700G -Xmx700G -XX:+UseTransparentHugePages -XX:+AlwaysPreTouch

The above command allocates 700G of Java heap using hugepages.

- With vanilla 5.6.0-rc3

17.39user 1666.48system 27:37.89elapsed

Quoted 3 lines

- With 5.6.0-rc3 + this patch, with proactiveness=20

8.35user 194.58system 3:19.62elapsed

I still wonder how the single additional CPU during compaction resulted in such
an improvement. Isn't this against the Amdahl's law? :)

Quoted 33 lines

Elapsed time remains around 3:15, as proactiveness is further increased.

Note that proactive compaction happens throughout the runtime of these workloads. The situation of one-time compaction, sufficient to supply hugepages for following allocation stream, can probably happen for more extreme proactiveness values, like 80 or 90.

In the above Java workload, proactiveness is set to 20. The test starts with a node's score of 80 or higher, depending on the delay between the fragmentation step and starting the benchmark, which gives more-or-less time for the initial round of compaction. As the benchmark consumes hugepages, node's score quickly rises above the high threshold (90) and proactive compaction starts again, which brings down the score to the low threshold level (80). Repeat.

bpftrace also confirms proactive compaction running 20+ times during the runtime of this Java benchmark. kcompactd threads consume 100% of one of the CPUs while it tries to bring a node's score within thresholds.

Backoff behavior
================

Above workloads produce a memory state which is easy to compact. However, if memory is filled with unmovable pages, proactive compaction should essentially back off. To test this aspect:

- Created a kernel driver that allocates almost all memory as hugepages followed by freeing first 3/4 of each hugepage.
- Set proactiveness=40
- Note that proactive_compact_node() is deferred maximum number of times with HPAGE_FRAG_CHECK_INTERVAL_MSEC of wait between each check (=> ~30 seconds between retries).

Quoted 4 lines

[1] https://patchwork.kernel.org/patch/11098289/

Signed-off-by Nitin Gupta <[email protected]>

To: Mel Gorman <[email protected]>

I hope Mel can also comment on this, but in general I agree.

...

mm/compaction.c static inline bool is_via_compact_memory(int order)

Quoted 4 lines

mm/compaction.c +4 -0

+
+/*
+ * A zone's fragmentation score is the external fragmentation wrt to the
+ * HUGETLB_PAGE_ORDER scaled by the zone's size. It returns a value in

the

Quoted 3 lines

+ * range [0, 100].
+
+ * The scaling factor ensures that proactive compaction focuses on

larger

mm/compaction.c static inline bool is_via_compact_memory(int order) · line 1893

Quoted 10 lines

+ * zones like ZONE_NORMAL, rather than smaller, specialized zones like
+ * ZONE_DMA32. For smaller zones, the score value remains close to zero,
+ * and thus never exceeds the high threshold for proactive compaction.
+ */

mm/compaction.c +6 -0

+static int fragmentation_score_zone(struct zone *zone)
+{
+     unsigned long score;
+
+     score = zone->present_pages *
+                     extfrag_for_order(zone, HUGETLB_PAGE_ORDER);

HPAGE_PMD_ORDER would be a better match than HUGETLB_PAGE_ORDER, even if it might be the same number. hugetlb pages are pre-reserved, unlike THP.

mm/compaction.c static inline bool is_via_compact_memory(int order) · line 1895

Quoted 2 lines

mm/compaction.c +2 -0

+     score = div64_ul(score,
+                     node_present_pages(zone->zone_pgdat->node_id) + 1);

zone->zone_pgdat->node_present_pages is more direct

+ return score;

mm/compaction.c static enum compact_result

Quoted 5 lines

mm/compaction.c +3 -0

+}
+
+/*

@@ -2309,6 +2411,7 @@ static enum compact_result

compact_zone_order(struct zone *zone, int order,

mm/compaction.c static enum compact_result compact_zone_order(struct zone *zone, int order, · line 2414

Quoted 4 lines

mm/compaction.c +1 -0

              .alloc_flags = alloc_flags,
              .classzone_idx = classzone_idx,
              .direct_compaction = true,
+             .proactive_compaction = false,

false, 0, NULL etc are implicitly initialized with this kind of initialization
(also in other places of the patch)

mm/compaction.c enum compact_result try_to_compact_pages(gfp_t

Quoted 4 lines

mm/compaction.c +0 -0

              .whole_zone = (prio == MIN_COMPACT_PRIORITY),
              .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
              .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)

@@ -2412,6 +2515,42 @@ enum compact_result try_to_compact_pages(gfp_t

gfp_mask, unsigned int order,

Quoted 25 lines

mm/compaction.c +17 -0

      return rc;
 }
 
 

@@ -2500,6 +2640,63 @@ void compaction_unregister_node(struct node *node)
 }
 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
 
+#ifdef CONFIG_SYSFS
+
+#define COMPACTION_ATTR_RO(_name) \
+     static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
+
+#define COMPACTION_ATTR(_name) \
+     static struct kobj_attribute _name##_attr = \
+             __ATTR(_name, 0644, _name##_show, _name##_store)
+
+static struct kobject *compaction_kobj;
+
+static ssize_t proactiveness_store(struct kobject *kobj,
+             struct kobj_attribute *attr, const char *buf, size_t count)
+{
+     int err;
+     unsigned long input;
+

mm/compaction.c void compaction_unregister_node(struct node *node) · line 2664

Quoted 5 lines

mm/compaction.c +5 -0

+     err = kstrtoul(buf, 10, &input);
+     if (err)
+             return err;
+     if (input > 100)
+             return -EINVAL;

The sysctl way also allows to specify min/max in the descriptor and use the generic handler