diff mbox series

[v29,03/13] mm/damon: Adaptively adjust regions

Message ID 20210520075629.4332-4-sj38.park@gmail.com (mailing list archive)
State New, archived
Headers show
Series Introduce Data Access MONitor (DAMON) | expand

Commit Message

SeongJae Park May 20, 2021, 7:56 a.m. UTC 
From: SeongJae Park <sjpark@amazon.de>

Even somehow the initial monitoring target regions are well constructed
to fulfill the assumption (pages in same region have similar access
frequencies), the data access pattern can be dynamically changed.  This
will result in low monitoring quality.  To keep the assumption as much
as possible, DAMON adaptively merges and splits each region based on
their access frequency.

For each ``aggregation interval``, it compares the access frequencies of
adjacent regions and merges those if the frequency difference is small.
Then, after it reports and clears the aggregated access frequency of
each region, it splits each region into two or three regions if the
total number of regions will not exceed the user-specified maximum
number of regions after the split.

In this way, DAMON provides its best-effort quality and minimal overhead
while keeping the upper-bound overhead that users set.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
Reviewed-by: Leonard Foerster <foersleo@amazon.de>
---
 include/linux/damon.h |  23 +++--
 mm/damon/core.c       | 214 +++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 227 insertions(+), 10 deletions(-)

Comments

Fernand Sieber May 25, 2021, 3:17 p.m. UTC | #1 
Hi SeongJae,

The code looks good. Some questions for this patch:

The region merge threshold is computed on the access diff. Should the 
diff threshold be exponential as diffs in low number of access are 
likely to be more important? I.e if the threshold is 5, a region A with 
0 accesses will be merged with a region B with 4 accesses (diff=4), but 
a region C with 50 access won't be merged with a region D with 60 
accesses (diff=10), however it seems to me that keeping a good 
granularity between A and B is more important than between C and D for 
FPR. What do you think?

When the number of regions is less than half max region, region split 
kicks in and doubles the number of region. This means that the number of 
region will grow close to max region, then slowly decay as region 
merges, until it reaches half max regions, then double again. This seems 
to create a non-uniform region number distribution over time, with large 
cycles. Also we do a lot of work when we double and no work otherwise. 
Not sure what's the impact on measurement quality but intuitively seems 
like keeping the number of regions constant over time would yield more 
consistent metrics? How about we rather always split regions at each 
iteration, and for each region we give a split probability?

Kind regards,

--Fernand
SeongJae Park May 25, 2021, 3:39 p.m. UTC | #2 
From: SeongJae Park <sjpark@amazon.de>

Hello Fernand,


Thank you for the questions!

On Tue, 25 May 2021 17:17:05 +0200 <sieberf@amazon.com> wrote:

> Hi SeongJae,
> 
> The code looks good. Some questions for this patch:
> 
> The region merge threshold is computed on the access diff. Should the 
> diff threshold be exponential as diffs in low number of access are 
> likely to be more important? I.e if the threshold is 5, a region A with 
> 0 accesses will be merged with a region B with 4 accesses (diff=4), but 
> a region C with 50 access won't be merged with a region D with 60 
> accesses (diff=10), however it seems to me that keeping a good 
> granularity between A and B is more important than between C and D for 
> FPR. What do you think?

That totally makes sense if we have interest in only cold pages.  However,
DAMON is for more general use cases.  In some cases, people would have interest
in hot pages.  Using exponential diff might make the regions merging more
aggressive, and result in smaller overhead.  But, I think the amount of the
problem and benefit is unclear for now.  I was unable to find the overhead
becomes problematically high in my tests with production systems.  I think we
could add another option for this later, after we find it becomes a real
problem.

> 
> When the number of regions is less than half max region, region split 
> kicks in and doubles the number of region. This means that the number of 
> region will grow close to max region, then slowly decay as region 
> merges, until it reaches half max regions, then double again. This seems 
> to create a non-uniform region number distribution over time, with large 
> cycles. Also we do a lot of work when we double and no work otherwise. 
> Not sure what's the impact on measurement quality but intuitively seems 
> like keeping the number of regions constant over time would yield more 
> consistent metrics? How about we rather always split regions at each 
> iteration, and for each region we give a split probability?

Agreed, I think this makes sense.  I also planning to make the probability
adaptively changes based on current monitoring result, in future.
Nevertheless, I want to keep the logic as simple as possible for now, unless we
see clear problem and benefit there.


Thanks,
SeongJae Park

> 
> Kind regards,
> 
> --Fernand
diff mbox series

Patch

diff --git a/include/linux/damon.h b/include/linux/damon.h
index 67db309ad61b..0bd5d6913a6c 100644
--- a/include/linux/damon.h
+++ b/include/linux/damon.h
@@ -12,6 +12,9 @@ 
 #include <linux/time64.h>
 #include <linux/types.h>
 
+/* Minimal region size.  Every damon_region is aligned by this. */
+#define DAMON_MIN_REGION	PAGE_SIZE
+
 /**
  * struct damon_addr_range - Represents an address region of [@start, @end).
  * @start:	Start address of the region (inclusive).
@@ -85,6 +88,8 @@  struct damon_ctx;
  * prepared for the next access check.
  * @check_accesses should check the accesses to each region that made after the
  * last preparation and update the number of observed accesses of each region.
+ * It should also return max number of observed accesses that made as a result
+ * of its update.  The value will be used for regions adjustment threshold.
  * @reset_aggregated should reset the access monitoring results that aggregated
  * by @check_accesses.
  * @target_valid should check whether the target is still valid for the
@@ -95,7 +100,7 @@  struct damon_primitive {
 	void (*init)(struct damon_ctx *context);
 	void (*update)(struct damon_ctx *context);
 	void (*prepare_access_checks)(struct damon_ctx *context);
-	void (*check_accesses)(struct damon_ctx *context);
+	unsigned int (*check_accesses)(struct damon_ctx *context);
 	void (*reset_aggregated)(struct damon_ctx *context);
 	bool (*target_valid)(void *target);
 	void (*cleanup)(struct damon_ctx *context);
@@ -172,7 +177,9 @@  struct damon_callback {
  * @primitive:	Set of monitoring primitives for given use cases.
  * @callback:	Set of callbacks for monitoring events notifications.
  *
- * @region_targets:	Head of monitoring targets (&damon_target) list.
+ * @min_nr_regions:	The minimum number of adaptive monitoring regions.
+ * @max_nr_regions:	The maximum number of adaptive monitoring regions.
+ * @adaptive_targets:	Head of monitoring targets (&damon_target) list.
  */
 struct damon_ctx {
 	unsigned long sample_interval;
@@ -191,7 +198,9 @@  struct damon_ctx {
 	struct damon_primitive primitive;
 	struct damon_callback callback;
 
-	struct list_head region_targets;
+	unsigned long min_nr_regions;
+	unsigned long max_nr_regions;
+	struct list_head adaptive_targets;
 };
 
 #define damon_next_region(r) \
@@ -207,10 +216,10 @@  struct damon_ctx {
 	list_for_each_entry_safe(r, next, &t->regions_list, list)
 
 #define damon_for_each_target(t, ctx) \
-	list_for_each_entry(t, &(ctx)->region_targets, list)
+	list_for_each_entry(t, &(ctx)->adaptive_targets, list)
 
 #define damon_for_each_target_safe(t, next, ctx)	\
-	list_for_each_entry_safe(t, next, &(ctx)->region_targets, list)
+	list_for_each_entry_safe(t, next, &(ctx)->adaptive_targets, list)
 
 #ifdef CONFIG_DAMON
 
@@ -224,11 +233,13 @@  struct damon_target *damon_new_target(unsigned long id);
 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t);
 void damon_free_target(struct damon_target *t);
 void damon_destroy_target(struct damon_target *t);
+unsigned int damon_nr_regions(struct damon_target *t);
 
 struct damon_ctx *damon_new_ctx(void);
 void damon_destroy_ctx(struct damon_ctx *ctx);
 int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
-		unsigned long aggr_int, unsigned long primitive_upd_int);
+		unsigned long aggr_int, unsigned long primitive_upd_int,
+		unsigned long min_nr_reg, unsigned long max_nr_reg);
 
 int damon_start(struct damon_ctx **ctxs, int nr_ctxs);
 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs);
diff --git a/mm/damon/core.c b/mm/damon/core.c
index 94db494dcf70..b36b6bdd94e2 100644
--- a/mm/damon/core.c
+++ b/mm/damon/core.c
@@ -10,8 +10,12 @@ 
 #include <linux/damon.h>
 #include <linux/delay.h>
 #include <linux/kthread.h>
+#include <linux/random.h>
 #include <linux/slab.h>
 
+/* Get a random number in [l, r) */
+#define damon_rand(l, r) (l + prandom_u32_max(r - l))
+
 static DEFINE_MUTEX(damon_lock);
 static int nr_running_ctxs;
 
@@ -87,7 +91,7 @@  struct damon_target *damon_new_target(unsigned long id)
 
 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t)
 {
-	list_add_tail(&t->list, &ctx->region_targets);
+	list_add_tail(&t->list, &ctx->adaptive_targets);
 }
 
 static void damon_del_target(struct damon_target *t)
@@ -110,6 +114,17 @@  void damon_destroy_target(struct damon_target *t)
 	damon_free_target(t);
 }
 
+unsigned int damon_nr_regions(struct damon_target *t)
+{
+	struct damon_region *r;
+	unsigned int nr_regions = 0;
+
+	damon_for_each_region(r, t)
+		nr_regions++;
+
+	return nr_regions;
+}
+
 struct damon_ctx *damon_new_ctx(void)
 {
 	struct damon_ctx *ctx;
@@ -127,7 +142,10 @@  struct damon_ctx *damon_new_ctx(void)
 
 	mutex_init(&ctx->kdamond_lock);
 
-	INIT_LIST_HEAD(&ctx->region_targets);
+	ctx->min_nr_regions = 10;
+	ctx->max_nr_regions = 1000;
+
+	INIT_LIST_HEAD(&ctx->adaptive_targets);
 
 	return ctx;
 }
@@ -157,6 +175,8 @@  void damon_destroy_ctx(struct damon_ctx *ctx)
  * @sample_int:		time interval between samplings
  * @aggr_int:		time interval between aggregations
  * @primitive_upd_int:	time interval between monitoring primitive updates
+ * @min_nr_reg:		minimal number of regions
+ * @max_nr_reg:		maximum number of regions
  *
  * This function should not be called while the kdamond is running.
  * Every time interval is in micro-seconds.
@@ -164,15 +184,49 @@  void damon_destroy_ctx(struct damon_ctx *ctx)
  * Return: 0 on success, negative error code otherwise.
  */
 int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
-		    unsigned long aggr_int, unsigned long primitive_upd_int)
+		    unsigned long aggr_int, unsigned long primitive_upd_int,
+		    unsigned long min_nr_reg, unsigned long max_nr_reg)
 {
+	if (min_nr_reg < 3) {
+		pr_err("min_nr_regions (%lu) must be at least 3\n",
+				min_nr_reg);
+		return -EINVAL;
+	}
+	if (min_nr_reg > max_nr_reg) {
+		pr_err("invalid nr_regions.  min (%lu) > max (%lu)\n",
+				min_nr_reg, max_nr_reg);
+		return -EINVAL;
+	}
+
 	ctx->sample_interval = sample_int;
 	ctx->aggr_interval = aggr_int;
 	ctx->primitive_update_interval = primitive_upd_int;
+	ctx->min_nr_regions = min_nr_reg;
+	ctx->max_nr_regions = max_nr_reg;
 
 	return 0;
 }
 
+/* Returns the size upper limit for each monitoring region */
+static unsigned long damon_region_sz_limit(struct damon_ctx *ctx)
+{
+	struct damon_target *t;
+	struct damon_region *r;
+	unsigned long sz = 0;
+
+	damon_for_each_target(t, ctx) {
+		damon_for_each_region(r, t)
+			sz += r->ar.end - r->ar.start;
+	}
+
+	if (ctx->min_nr_regions)
+		sz /= ctx->min_nr_regions;
+	if (sz < DAMON_MIN_REGION)
+		sz = DAMON_MIN_REGION;
+
+	return sz;
+}
+
 static bool damon_kdamond_running(struct damon_ctx *ctx)
 {
 	bool running;
@@ -339,6 +393,149 @@  static void kdamond_reset_aggregated(struct damon_ctx *c)
 	}
 }
 
+#define sz_damon_region(r) (r->ar.end - r->ar.start)
+
+/*
+ * Merge two adjacent regions into one region
+ */
+static void damon_merge_two_regions(struct damon_region *l,
+				struct damon_region *r)
+{
+	unsigned long sz_l = sz_damon_region(l), sz_r = sz_damon_region(r);
+
+	l->nr_accesses = (l->nr_accesses * sz_l + r->nr_accesses * sz_r) /
+			(sz_l + sz_r);
+	l->ar.end = r->ar.end;
+	damon_destroy_region(r);
+}
+
+#define diff_of(a, b) (a > b ? a - b : b - a)
+
+/*
+ * Merge adjacent regions having similar access frequencies
+ *
+ * t		target affected by this merge operation
+ * thres	'->nr_accesses' diff threshold for the merge
+ * sz_limit	size upper limit of each region
+ */
+static void damon_merge_regions_of(struct damon_target *t, unsigned int thres,
+				   unsigned long sz_limit)
+{
+	struct damon_region *r, *prev = NULL, *next;
+
+	damon_for_each_region_safe(r, next, t) {
+		if (prev && prev->ar.end == r->ar.start &&
+		    diff_of(prev->nr_accesses, r->nr_accesses) <= thres &&
+		    sz_damon_region(prev) + sz_damon_region(r) <= sz_limit)
+			damon_merge_two_regions(prev, r);
+		else
+			prev = r;
+	}
+}
+
+/*
+ * Merge adjacent regions having similar access frequencies
+ *
+ * threshold	'->nr_accesses' diff threshold for the merge
+ * sz_limit	size upper limit of each region
+ *
+ * This function merges monitoring target regions which are adjacent and their
+ * access frequencies are similar.  This is for minimizing the monitoring
+ * overhead under the dynamically changeable access pattern.  If a merge was
+ * unnecessarily made, later 'kdamond_split_regions()' will revert it.
+ */
+static void kdamond_merge_regions(struct damon_ctx *c, unsigned int threshold,
+				  unsigned long sz_limit)
+{
+	struct damon_target *t;
+
+	damon_for_each_target(t, c)
+		damon_merge_regions_of(t, threshold, sz_limit);
+}
+
+/*
+ * Split a region in two
+ *
+ * r		the region to be split
+ * sz_r		size of the first sub-region that will be made
+ */
+static void damon_split_region_at(struct damon_ctx *ctx,
+				  struct damon_region *r, unsigned long sz_r)
+{
+	struct damon_region *new;
+
+	new = damon_new_region(r->ar.start + sz_r, r->ar.end);
+	if (!new)
+		return;
+
+	r->ar.end = new->ar.start;
+
+	damon_insert_region(new, r, damon_next_region(r));
+}
+
+/* Split every region in the given target into 'nr_subs' regions */
+static void damon_split_regions_of(struct damon_ctx *ctx,
+				     struct damon_target *t, int nr_subs)
+{
+	struct damon_region *r, *next;
+	unsigned long sz_region, sz_sub = 0;
+	int i;
+
+	damon_for_each_region_safe(r, next, t) {
+		sz_region = r->ar.end - r->ar.start;
+
+		for (i = 0; i < nr_subs - 1 &&
+				sz_region > 2 * DAMON_MIN_REGION; i++) {
+			/*
+			 * Randomly select size of left sub-region to be at
+			 * least 10 percent and at most 90% of original region
+			 */
+			sz_sub = ALIGN_DOWN(damon_rand(1, 10) *
+					sz_region / 10, DAMON_MIN_REGION);
+			/* Do not allow blank region */
+			if (sz_sub == 0 || sz_sub >= sz_region)
+				continue;
+
+			damon_split_region_at(ctx, r, sz_sub);
+			sz_region = sz_sub;
+		}
+	}
+}
+
+/*
+ * Split every target region into randomly-sized small regions
+ *
+ * This function splits every target region into random-sized small regions if
+ * current total number of the regions is equal or smaller than half of the
+ * user-specified maximum number of regions.  This is for maximizing the
+ * monitoring accuracy under the dynamically changeable access patterns.  If a
+ * split was unnecessarily made, later 'kdamond_merge_regions()' will revert
+ * it.
+ */
+static void kdamond_split_regions(struct damon_ctx *ctx)
+{
+	struct damon_target *t;
+	unsigned int nr_regions = 0;
+	static unsigned int last_nr_regions;
+	int nr_subregions = 2;
+
+	damon_for_each_target(t, ctx)
+		nr_regions += damon_nr_regions(t);
+
+	if (nr_regions > ctx->max_nr_regions / 2)
+		return;
+
+	/* Maybe the middle of the region has different access frequency */
+	if (last_nr_regions == nr_regions &&
+			nr_regions < ctx->max_nr_regions / 3)
+		nr_subregions = 3;
+
+	damon_for_each_target(t, ctx)
+		damon_split_regions_of(ctx, t, nr_subregions);
+
+	last_nr_regions = nr_regions;
+}
+
 /*
  * Check whether it is time to check and apply the target monitoring regions
  *
@@ -395,6 +592,8 @@  static int kdamond_fn(void *data)
 	struct damon_ctx *ctx = (struct damon_ctx *)data;
 	struct damon_target *t;
 	struct damon_region *r, *next;
+	unsigned int max_nr_accesses = 0;
+	unsigned long sz_limit = 0;
 
 	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
 
@@ -403,6 +602,8 @@  static int kdamond_fn(void *data)
 	if (ctx->callback.before_start && ctx->callback.before_start(ctx))
 		set_kdamond_stop(ctx);
 
+	sz_limit = damon_region_sz_limit(ctx);
+
 	while (!kdamond_need_stop(ctx)) {
 		if (ctx->primitive.prepare_access_checks)
 			ctx->primitive.prepare_access_checks(ctx);
@@ -413,13 +614,17 @@  static int kdamond_fn(void *data)
 		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
 
 		if (ctx->primitive.check_accesses)
-			ctx->primitive.check_accesses(ctx);
+			max_nr_accesses = ctx->primitive.check_accesses(ctx);
 
 		if (kdamond_aggregate_interval_passed(ctx)) {
+			kdamond_merge_regions(ctx,
+					max_nr_accesses / 10,
+					sz_limit);
 			if (ctx->callback.after_aggregation &&
 					ctx->callback.after_aggregation(ctx))
 				set_kdamond_stop(ctx);
 			kdamond_reset_aggregated(ctx);
+			kdamond_split_regions(ctx);
 			if (ctx->primitive.reset_aggregated)
 				ctx->primitive.reset_aggregated(ctx);
 		}
@@ -427,6 +632,7 @@  static int kdamond_fn(void *data)
 		if (kdamond_need_update_primitive(ctx)) {
 			if (ctx->primitive.update)
 				ctx->primitive.update(ctx);
+			sz_limit = damon_region_sz_limit(ctx);
 		}
 	}
 	damon_for_each_target(t, ctx) {