Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Tue, 31 Mar 2020 17:08:55 +0100
From:   Jonathan Cameron <Jonathan.Cameron@Huawei.com>
To:     SeongJae Park <sjpark@amazon.com>
CC:     <akpm@linux-foundation.org>, SeongJae Park <sjpark@amazon.de>,
        <aarcange@redhat.com>, <acme@kernel.org>,
        <alexander.shishkin@linux.intel.com>, <amit@kernel.org>,
        <brendan.d.gregg@gmail.com>, <brendanhiggins@google.com>,
        <cai@lca.pw>, <colin.king@canonical.com>, <corbet@lwn.net>,
        <dwmw@amazon.com>, <jolsa@redhat.com>, <kirill@shutemov.name>,
        <mark.rutland@arm.com>, <mgorman@suse.de>, <minchan@kernel.org>,
        <mingo@redhat.com>, <namhyung@kernel.org>, <peterz@infradead.org>,
        <rdunlap@infradead.org>, <riel@surriel.com>, <rientjes@google.com>,
        <rostedt@goodmis.org>, <shuah@kernel.org>, <sj38.park@gmail.com>,
        <vbabka@suse.cz>, <vdavydov.dev@gmail.com>,
        <yang.shi@linux.alibaba.com>, <ying.huang@intel.com>,
        <linux-mm@kvack.org>, <linux-doc@vger.kernel.org>,
        <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH v7 05/15] mm/damon: Adaptively adjust regions
Message-ID: <20200331170855.0000024f@Huawei.com>
In-Reply-To: <20200318112722.30143-6-sjpark@amazon.com>
References: <20200318112722.30143-1-sjpark@amazon.com>
        <20200318112722.30143-6-sjpark@amazon.com>
Organization: Huawei Technologies Research and Development (UK) Ltd.
MIME-Version: 1.0
Content-Type: text/plain; charset="US-ASCII"
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Wed, 18 Mar 2020 12:27:12 +0100
SeongJae Park <sjpark@amazon.com> wrote:

> From: SeongJae Park <sjpark@amazon.de>
> 
> At the beginning of the monitoring, DAMON constructs the initial regions
> by evenly splitting the memory mapped address space of the process into
> the user-specified minimal number of regions.  In this initial state,
> the assumption of the regions (pages in same region have similar access
> frequencies) is normally not kept and thus the monitoring quality could
> be low.  To keep the assumption as much as possible, DAMON adaptively
> merges and splits each region.
> 
> 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 regions if the total number
> of regions is smaller than the half of the user-specified maximum number
> of regions.
> 
> In this way, DAMON provides its best-effort quality and minimal overhead
> while keeping the bounds users set for their trade-off.
> 
> Signed-off-by: SeongJae Park <sjpark@amazon.de>

A few more edge cases in here, and a suggestion that might be more costly
but lead to simpler code.

Jonathan

> ---
>  include/linux/damon.h |   6 +-
>  mm/damon.c            | 148 ++++++++++++++++++++++++++++++++++++++++--
>  2 files changed, 145 insertions(+), 9 deletions(-)
> 
> diff --git a/include/linux/damon.h b/include/linux/damon.h
> index f1945df6e6b4..7562b85b1ec0 100644
> --- a/include/linux/damon.h
> +++ b/include/linux/damon.h
> @@ -42,6 +42,7 @@ struct damon_ctx {
>  	unsigned long sample_interval;
>  	unsigned long aggr_interval;
>  	unsigned long min_nr_regions;
> +	unsigned long max_nr_regions;
>  
>  	struct timespec64 last_aggregation;
>  
> @@ -54,8 +55,9 @@ struct damon_ctx {
>  };
>  
>  int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids);
> -int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
> -		unsigned long aggr_int, unsigned long min_nr_reg);
> +int damon_set_attrs(struct damon_ctx *ctx,
> +		unsigned long sample_int, unsigned long aggr_int,
> +		unsigned long min_nr_reg, unsigned long max_nr_reg);
>  int damon_start(struct damon_ctx *ctx);
>  int damon_stop(struct damon_ctx *ctx);
>  
> diff --git a/mm/damon.c b/mm/damon.c
> index 018016793555..23c0de3b502e 100644
> --- a/mm/damon.c
> +++ b/mm/damon.c
> @@ -342,9 +342,12 @@ static int damon_three_regions_of(struct damon_task *t,
>   * regions is wasteful.  That said, because we can deal with small noises,
>   * tracking every mapping is not strictly required but could even incur a high
>   * overhead if the mapping frequently changes or the number of mappings is
> - * high.  Nonetheless, this may seems very weird.  DAMON's dynamic regions
> - * adjustment mechanism, which will be implemented with following commit will
> - * make this more sense.
> + * high.  The adaptive regions adjustment mechanism will further help to deal
> + * with the noises by simply identifying the unmapped areas as a region that
> + * has no access.  Moreover, applying the real mappings that would have many
> + * unmapped areas inside will make the adaptive mechanism quite complex.  That
> + * said, too huge unmapped areas inside the monitoring target should be removed
> + * to not take the time for the adaptive mechanism.
>   *
>   * For the reason, we convert the complex mappings to three distinct regions
>   * that cover every mapped areas of the address space.  Also the two gaps
> @@ -533,6 +536,121 @@ static void kdamond_reset_aggregated(struct damon_ctx *c)
>  	}
>  }
>  
> +#define sz_damon_region(r) (r->vm_end - r->vm_start)
> +
> +/*
> + * Merge two adjacent regions into one region
> + */
> +static void damon_merge_two_regions(struct damon_region *l,
> +				struct damon_region *r)
> +{
> +	l->nr_accesses = (l->nr_accesses * sz_damon_region(l) +
> +			r->nr_accesses * sz_damon_region(r)) /
> +			(sz_damon_region(l) + sz_damon_region(r));
> +	l->vm_end = r->vm_end;
> +	damon_destroy_region(r);
> +}
> +
> +#define diff_of(a, b) (a > b ? a - b : b - a)
> +
> +/*
> + * Merge adjacent regions having similar access frequencies
> + *
> + * t		task that merge operation will make change
> + * thres	merge regions having '->nr_accesses' diff smaller than this
> + */
> +static void damon_merge_regions_of(struct damon_task *t, unsigned int thres)
> +{
> +	struct damon_region *r, *prev = NULL, *next;
> +
> +	damon_for_each_region_safe(r, next, t) {
> +		if (!prev || prev->vm_end != r->vm_start ||
> +		    diff_of(prev->nr_accesses, r->nr_accesses) > thres) {
> +			prev = r;
> +			continue;
> +		}
> +		damon_merge_two_regions(prev, r);
> +	}
> +}
> +
> +/*
> + * Merge adjacent regions having similar access frequencies
> + *
> + * threshold	merge regions havind nr_accesses diff larger than this
> + *
> + * 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)
> +{
> +	struct damon_task *t;
> +
> +	damon_for_each_task(c, t)
> +		damon_merge_regions_of(t, threshold);
> +}
> +
> +/*
> + * Split a region into two small regions
> + *
> + * 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(ctx, r->vm_start + sz_r, r->vm_end);
> +	r->vm_end = new->vm_start;

We may well have a sampling address that is in the wrong region.
It should have little effect on the stats as will fix on next sample
but in my view still worth cleaning up.

> +
> +	damon_insert_region(new, r, damon_next_region(r));
> +}
> +
> +static void damon_split_regions_of(struct damon_ctx *ctx, struct damon_task *t)
> +{
> +	struct damon_region *r, *next;
> +	unsigned long sz_left_region;
> +
> +	damon_for_each_region_safe(r, next, t) {
> +		/*
> +		 * Randomly select size of left sub-region to be at least
> +		 * 10 percent and at most 90% of original region
> +		 */
> +		sz_left_region = (prandom_u32_state(&ctx->rndseed) % 9 + 1) *
> +			(r->vm_end - r->vm_start) / 10;
> +		/* Do not allow blank region */
> +		if (sz_left_region == 0)
> +			continue;
> +		damon_split_region_at(ctx, r, sz_left_region);
> +	}
> +}
> +
> +/*
> + * splits every target regions into two randomly-sized regions
> + *
> + * This function splits every target regions into two random-sized regions if
> + * current total number of the regions is smaller than the 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_task *t;
> +	unsigned int nr_regions = 0;
> +
> +	damon_for_each_task(ctx, t)
> +		nr_regions += nr_damon_regions(t);
> +	if (nr_regions > ctx->max_nr_regions / 2)
> +		return;
> +
> +	damon_for_each_task(ctx, t)
> +		damon_split_regions_of(ctx, t);
> +}
> +
>  /*
>   * Check whether current monitoring should be stopped
>   *
> @@ -571,21 +689,29 @@ static int kdamond_fn(void *data)
>  	struct damon_task *t;
>  	struct damon_region *r, *next;
>  	struct mm_struct *mm;
> +	unsigned int max_nr_accesses;
>  
>  	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
>  	kdamond_init_regions(ctx);
>  	while (!kdamond_need_stop(ctx)) {
> +		max_nr_accesses = 0;
>  		damon_for_each_task(ctx, t) {
>  			mm = damon_get_mm(t);
>  			if (!mm)
>  				continue;
> -			damon_for_each_region(r, t)
> +			damon_for_each_region(r, t) {
>  				kdamond_check_access(ctx, mm, r);
> +				max_nr_accesses = max(r->nr_accesses,
> +						max_nr_accesses);
> +			}
>  			mmput(mm);
>  		}
>  
> -		if (kdamond_aggregate_interval_passed(ctx))
> +		if (kdamond_aggregate_interval_passed(ctx)) {
> +			kdamond_merge_regions(ctx, max_nr_accesses / 10);
>  			kdamond_reset_aggregated(ctx);
> +			kdamond_split_regions(ctx);
> +		}

I wonder if it would be simpler to split the sampling address setup and
mkold from the access check.  We would have to walk regions twice,
but not have to bother separately dealing with updating some regions
if they are modified in the above block.

Also, the above has some overhead, so will bias that first sample each
time the block above runs.  If we do the mkold afterwards it will make
much less difference.

>  
>  		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
>  	}
> @@ -702,24 +828,32 @@ int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids)
>   * @sample_int:		time interval between samplings
>   * @aggr_int:		time interval between aggregations
>   * @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.
>   *
>   * 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 min_nr_reg)
> +int damon_set_attrs(struct damon_ctx *ctx,
> +			unsigned long sample_int, unsigned long aggr_int,
> +			unsigned long min_nr_reg, unsigned long max_nr_reg)
>  {
>  	if (min_nr_reg < 3) {
>  		pr_err("min_nr_regions (%lu) should be bigger than 2\n",
>  				min_nr_reg);
>  		return -EINVAL;
>  	}
> +	if (min_nr_reg >= ctx->max_nr_regions) {
> +		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->min_nr_regions = min_nr_reg;
> +	ctx->max_nr_regions = max_nr_reg;
>  
>  	return 0;
>  }