Re: Re: [PATCH v7 04/15] mm/damon: Implement region based sampling

From: SeongJae Park
Date: Wed Apr 01 2020 - 04:23:09 EST


On Tue, 31 Mar 2020 17:02:33 +0100 Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx> wrote:

> On Wed, 18 Mar 2020 12:27:11 +0100
> SeongJae Park <sjpark@xxxxxxxxxx> wrote:
>
> > From: SeongJae Park <sjpark@xxxxxxxxx>
> >
> > This commit implements DAMON's basic access check and region based
> > sampling mechanisms. This change would seems make no sense, mainly
> > because it is only a part of the DAMON's logics. Following two commits
> > will make more sense.
> >
> > Basic Access Check
> > ------------------
> >
> > DAMON basically reports what pages are how frequently accessed. Note
> > that the frequency is not an absolute number of accesses, but a relative
> > frequency among the pages of the target workloads.
> >
> > Users can control the resolution of the reports by setting two time
> > intervals, ``sampling interval`` and ``aggregation interval``. In
> > detail, DAMON checks access to each page per ``sampling interval``,
> > aggregates the results (counts the number of the accesses to each page),
> > and reports the aggregated results per ``aggregation interval``. For
> > the access check of each page, DAMON uses the Accessed bits of PTEs.
> >
> > This is thus similar to common periodic access checks based access
> > tracking mechanisms, which overhead is increasing as the size of the
> > target process grows.
> >
> > Region Based Sampling
> > ---------------------
> >
> > To avoid the unbounded increase of the overhead, DAMON groups a number
> > of adjacent pages that assumed to have same access frequencies into a
> > region. As long as the assumption (pages in a region have same access
> > frequencies) is kept, only one page in the region is required to be
> > checked. Thus, for each ``sampling interval``, DAMON randomly picks one
> > page in each region and clears its Accessed bit. After one more
> > ``sampling interval``, DAMON reads the Accessed bit of the page and
> > increases the access frequency of the region if the bit has set
> > meanwhile. Therefore, the monitoring overhead is controllable by
> > setting the number of regions.
> >
> > Nonetheless, this scheme cannot preserve the quality of the output if
> > the assumption is not kept. Following commit will introduce how we can
> > make the guarantee with best effort.
> >
> > Signed-off-by: SeongJae Park <sjpark@xxxxxxxxx>
>
> Hi.
>
> A few comments inline.
>
> I've still not replicated your benchmarks so may well have some more
> feedback once I've managed that on one of our servers.

Appreciate your comments. If you need any help for the replication, please let
me know. I basically use my parsec3 wrapper scripts[1] to run parsec3 and
splash2x workloads and `damo` tool, which resides in the kernel tree at
`/tools/damon/`.

For example, below commands will reproduce ethp applied splash2x/fft run.

$ echo "2M null 5 null null null hugepage
2M null null 5 1s null nohugepage" > ethp
$ parsec3_on_ubuntu/run.sh splash2x.fft
$ linux/tools/damon/damo schemes -c ethp `pidof fft`

[1] https://github.com/sjp38/parsec3_on_ubuntu

>
> Thanks,
>
> Jonathan
>
> > ---
> > include/linux/damon.h | 24 ++
> > mm/damon.c | 553 ++++++++++++++++++++++++++++++++++++++++++
> > 2 files changed, 577 insertions(+)
> >
[...]
> > diff --git a/mm/damon.c b/mm/damon.c
> > index d7e6226ab7f1..018016793555 100644
> > --- a/mm/damon.c
> > +++ b/mm/damon.c
> > @@ -10,8 +10,14 @@
> > #define pr_fmt(fmt) "damon: " fmt
> >
> > #include <linux/damon.h>
> > +#include <linux/delay.h>
> > +#include <linux/kthread.h>
> > #include <linux/mm.h>
> > #include <linux/module.h>
> > +#include <linux/page_idle.h>
> > +#include <linux/random.h>
> > +#include <linux/sched/mm.h>
> > +#include <linux/sched/task.h>
> > #include <linux/slab.h>
> >
[...]
> > +/*
> > + * Size-evenly split a region into 'nr_pieces' small regions
> > + *
> > + * Returns 0 on success, or negative error code otherwise.
> > + */
> > +static int damon_split_region_evenly(struct damon_ctx *ctx,
> > + struct damon_region *r, unsigned int nr_pieces)
> > +{
> > + unsigned long sz_orig, sz_piece, orig_end;
> > + struct damon_region *piece = NULL, *next;
> > + unsigned long start;
> > +
> > + if (!r || !nr_pieces)
> > + return -EINVAL;
> > +
> > + orig_end = r->vm_end;
> > + sz_orig = r->vm_end - r->vm_start;
> > + sz_piece = sz_orig / nr_pieces;
> > +
> > + if (!sz_piece)
> > + return -EINVAL;
> > +
> > + r->vm_end = r->vm_start + sz_piece;
> > + next = damon_next_region(r);
> > + for (start = r->vm_end; start + sz_piece <= orig_end;
> > + start += sz_piece) {
> > + piece = damon_new_region(ctx, start, start + sz_piece);
> piece may be n

Yes, that name is short and more intuitive. I will rename so.

> > + damon_insert_region(piece, r, next);
> > + r = piece;
> > + }
> > + /* complement last region for possible rounding error */
> > + if (piece)
> > + piece->vm_end = orig_end;
>
> Update the sampling address to ensure it's in the region?

I think `piece->vm_end` should be equal or smaller than `orig_end` and
therefore the sampling address of `piece` will be still in the region.

>
> > +
> > + return 0;
> > +}
> > +
[...]
> > +static void damon_pte_pmd_mkold(pte_t *pte, pmd_t *pmd)
> > +{
> > + if (pte) {
> > + if (pte_young(*pte)) {
> > + clear_page_idle(pte_page(*pte));
> > + set_page_young(pte_page(*pte));
> > + }
> > + *pte = pte_mkold(*pte);
> > + return;
> > + }
> > +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> > + if (pmd) {
> > + if (pmd_young(*pmd)) {
> > + clear_page_idle(pmd_page(*pmd));
> > + set_page_young(pmd_page(*pmd));
> > + }
> > + *pmd = pmd_mkold(*pmd);
> > + }
> > +#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
>
> No need to flush the TLBs?

Good point!

I have intentionally skipped TLB flushing here to minimize the performance
effect to the target workload. I also thought this might not degrade the
monitoring accuracy so much because we are targetting for the DRAM level
accesses of memory-intensive workloads, which might make TLB flood frequently.

However, your comment makes me thinking differently now. By flushing the TLB
here, we will increase up to `number_of_regions` TLB misses for sampling
interval. This might be not a huge overhead. Also, improving the monitoring
accuracy makes no harm at all. I even didn't measured the overhead.

I will test the overhead and if it is not significant, I will make this code to
flush TLB, in the next spin.

>
> > +}
> > +
[...]
> > +/*
> > + * The monitoring daemon that runs as a kernel thread
> > + */
> > +static int kdamond_fn(void *data)
> > +{
> > + struct damon_ctx *ctx = data;
> > + struct damon_task *t;
> > + struct damon_region *r, *next;
> > + struct mm_struct *mm;
> > +
> > + pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
> > + kdamond_init_regions(ctx);
>
> We haven't called mkold on the initial regions so first check will
> get us fairly random state.

Yes, indeed. However, the early results will not be accurate anyway because
the adaptive regions adjustment algorithm will not take effect yet. I would
like to leave this part as is but add some comments about this point to keep
the code simple.

>
> > + while (!kdamond_need_stop(ctx)) {
> > + damon_for_each_task(ctx, t) {
> > + mm = damon_get_mm(t);
> > + if (!mm)
> > + continue;
> > + damon_for_each_region(r, t)
> > + kdamond_check_access(ctx, mm, r);
> > + mmput(mm);
> > + }
> > +
> > + if (kdamond_aggregate_interval_passed(ctx))
> > + kdamond_reset_aggregated(ctx);
> > +
> > + usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
> > + }
> > + damon_for_each_task(ctx, t) {
> > + damon_for_each_region_safe(r, next, t)
> > + damon_destroy_region(r);
> > + }
> > + pr_debug("kdamond (%d) finishes\n", ctx->kdamond->pid);
> > + mutex_lock(&ctx->kdamond_lock);
> > + ctx->kdamond = NULL;
> > + mutex_unlock(&ctx->kdamond_lock);
> > +
> > + return 0;
> > +}
> > +
[...]
> > +/*
> > + * Start or stop the kdamond
> > + *
> > + * Returns 0 if success, negative error code otherwise.
> > + */
> > +static int damon_turn_kdamond(struct damon_ctx *ctx, bool on)
> > +{
> > + int err = -EBUSY;
> > +
> > + mutex_lock(&ctx->kdamond_lock);
> > + if (!ctx->kdamond && on) {
>
> Given there is very little shared code between on and off, I would
> suggest just splitting it into two functions.

Good point, I will do so in next spin.

>
> > + err = 0;
> > + ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond");
> > + if (IS_ERR(ctx->kdamond))
> > + err = PTR_ERR(ctx->kdamond);
> > + } else if (ctx->kdamond && !on) {
> > + mutex_unlock(&ctx->kdamond_lock);
> > + kthread_stop(ctx->kdamond);
> > + while (damon_kdamond_running(ctx))
> > + usleep_range(ctx->sample_interval,
> > + ctx->sample_interval * 2);
> > + return 0;
> > + }
> > + mutex_unlock(&ctx->kdamond_lock);
> > +
> > + return err;
> > +}
> > +
[...]
> > +
> > +/*
>
> Why not make these actual kernel-doc? That way you can use the
> kernel-doc scripts to sanity check them.

Oops, I just forgot that it should start with '/**'. Will fix it in next spin.


Thanks,
SeongJae Park

>
> /**
>
> > + * damon_set_attrs() - Set attributes for the monitoring.
> > + * @ctx: monitoring context
> > + * @sample_int: time interval between samplings
> > + * @aggr_int: time interval between aggregations
> > + * @min_nr_reg: minimal 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)
> > +{
> > + if (min_nr_reg < 3) {
> > + pr_err("min_nr_regions (%lu) should be bigger than 2\n",
> > + min_nr_reg);
> > + return -EINVAL;
> > + }
> > +
> > + ctx->sample_interval = sample_int;
> > + ctx->aggr_interval = aggr_int;
> > + ctx->min_nr_regions = min_nr_reg;
> > +
> > + return 0;
> > +}
> > +
> > static int __init damon_init(void)
> > {
> > return 0;
>