Re: [PATCH v9 00/15] Introduce Data Access MONitor (DAMON)

[Jonathan Cameron <Jonathan.Cameron@Huawei.com>]

On Wed, 29 Apr 2020 09:49:54 +0200
SeongJae Park <sjpark@amazon.com> wrote:

> On Tue, 28 Apr 2020 17:17:13 +0100 Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:
> 
> > On Tue, 28 Apr 2020 15:23:42 +0200
> > SeongJae Park <sjpark@amazon.com> wrote:
> >   
> > > On Tue, 28 Apr 2020 13:27:04 +0100 Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:
> > >   
> > > > On Mon, 27 Apr 2020 14:04:27 +0200
> > > > SeongJae Park <sjpark@amazon.com> wrote:
> > > >     
> > > > > From: SeongJae Park <sjpark@amazon.de>
> > > > > 
> > > > > Introduction
> > > > > ============
> > > > > 
> > > > > Memory management decisions can be improved if finer data access information is
> > > > > available.  However, because such finer information usually comes with higher
> > > > > overhead, most systems including Linux forgives the potential benefit and rely
> > > > > on only coarse information or some light-weight heuristics.  The pseudo-LRU and
> > > > > the aggressive THP promotions are such examples.
> > > > > 
> > > > > A number of data access pattern awared memory management optimizations (refer
> > > > > to 'Appendix A' for more details) consistently say the potential benefit is not
> > > > > small.  However, none of those has successfully merged to the mainline Linux
> > > > > kernel mainly due to the absence of a scalable and efficient data access
> > > > > monitoring mechanism.  Refer to 'Appendix B' to see the limitations of existing
> > > > > memory monitoring mechanisms.
> > > > > 
> > > > > DAMON is a data access monitoring subsystem for the problem.  It is 1) accurate
> > > > > enough to be used for the DRAM level memory management (a straightforward
> > > > > DAMON-based optimization achieved up to 2.55x speedup), 2) light-weight enough
> > > > > to be applied online (compared to a straightforward access monitoring scheme,
> > > > > DAMON is up to 94,242.42x lighter) and 3) keeps predefined upper-bound overhead
> > > > > regardless of the size of target workloads (thus scalable).  Refer to 'Appendix
> > > > > C' if you interested in how it is possible, and 'Appendix F' to know how the
> > > > > numbers collected.
> > > > > 
> > > > > DAMON has mainly designed for the kernel's memory management mechanisms.
> > > > > However, because it is implemented as a standalone kernel module and provides
> > > > > several interfaces, it can be used by a wide range of users including kernel
> > > > > space programs, user space programs, programmers, and administrators.  DAMON
> > > > > is now supporting the monitoring only, but it will also provide simple and
> > > > > convenient data access pattern awared memory managements by itself.  Refer to
> > > > > 'Appendix D' for more detailed expected usages of DAMON.
> > > > >     
> > > [...]  
> > > > > 
> > > > > Future Plans
> > > > > ============
> > > > > 
> > > > > This patchset is only for the first stage of DAMON.  As soon as this patchset
> > > > > is merged, official patchsets for below future plans will be posted.
> > > > >     
> > > [...]  
> > > > > 
> > > > > Support Various Address Spaces
> > > > > ------------------------------
> > > > > 
> > > > > Currently, DAMON supports virtual memory address spaces using PTE Accessed bits
> > > > > as its access checking primitive.  However, the core design of DAMON is not
> > > > > dependent to such implementation details.  In a future, DAMON will decouple
> > > > > those and support various address spaces including physical memory.  It will
> > > > > further allow users to configure and even implement the primitives by
> > > > > themselves for their special usecase.  Monitoring of page cache, NUMA nodes,
> > > > > specific files, or block devices would be examples of such usecases.
> > > > > 
> > > > > An RFC patchset for this plan is already available
> > > > > (https://lore.kernel.org/linux-mm/20200409094232.29680-1-sjpark@amazon.com/).
> > > > >     
> > > [...]  
> > > > > 
> > > > > Patch History
> > > > > =============
> > > > > 
> > > > > The most biggest change in this version is support of minimal region size,
> > > > > which defaults to 'PAGE_SIZE'.  This change will reduce unnecessary region
> > > > > splits and thus improve the quality of the output.  In a future, we will be
> > > > > able to make this configurable for support of various access check primitives
> > > > > such as PMUs.    
> > > > 
> > > > That is a good improvement.  Might be interesting to consider taking
> > > > hugepages into account as well.    
> > > 
> > > Thanks!  Kudos to Stefan and you for giving me the comments for the change.
> > > 
> > > As abovely mentioned in 'Future Plans' section, DAMON will be highly
> > > configurable.  You can see the plan in more detail via the RFC patchset[1].
> > > Thus, the minimal region size will also be able to configured as users want,
> > > including the size of the hugepage.
> > > 
> > > [1] https://lore.kernel.org/linux-mm/20200409094232.29680-1-sjpark@amazon.com/
> > >   
> > > > 
> > > > One issue I've noted is that we have a degeneracy problem with the current
> > > > region merging and splitting that perhaps could do with a small tweak.
> > > > 
> > > > Currently we can end with a very small number of regions because there
> > > > is no limit on how many regions can be merged in a give pass for merging.
> > > > However, splitting only doubles the number of regions.
> > > > 
> > > > I've been experimenting with a few loops of the splitting algorithm to ensure
> > > > we don't end up stuck with limited regions.  I think the problem we are working
> > > > around can be roughly described as:
> > > > 
> > > > 1) Program allocates a lot of memory - not really touching much of it.
> > > > 2) Damon fuses the large memory allocations in to one region because the
> > > >    access counts are always near 0. 
> > > > 3) Program finishes setup.
> > > > 4) Program accesses a few pages in the huge reason a lot, but not that much
> > > >    for most of the rest.  Taking an extreme option, the page in the middle
> > > >    gets all the accesses and the other 1G on either side gets none.
> > > > 5) As a split always breaks the page in two, the chances of significantly
> > > >    different values for the two resulting regions is low (as we only sample
> > > >    the hot page occasionally).
> > > > 
> > > > If we just run the splits twice if the number of regions < max regions / 4
> > > > then over time we should eventually get a region with the single hot page in it.
> > > > We will get there faster if we split more (keeping below max regions).
> > > > 
> > > > As we always remain below max regions, we are still obeying the fixed
> > > > maximum overhead and actually monitoring at closer to the desired granularity.    
> > > 
> > > Good point.  However, as you also mentioned, DAMON will slowly, but eventually
> > > adjust the regions appropriately.
> > > 
> > > And yes, your suggested solution will work pretty well.  Indeed, my one
> > > previous colleague found this problem on a few of special workloads and tried
> > > the solution you suggested.  The improvement was clear.
> > > 
> > > However, I didn't adopt the solution due to below reasons.
> > > 
> > > First, IMHO, this is an accuracy improvement, rather than bug fix.  But the
> > > extent of the enhancement didn't seem very critical to me.  Most of other
> > > workloads didn't show such problem (and thus improvement).  Even with the
> > > workloads showing the problem, the problem was not seem so critical.
> > > 
> > > Second, if the low accuracy is problem, users could get higher accuracy by
> > > simply adjusting the sampling interval and/or aggregation interval to lower
> > > value.  This is the supposed way to trade the accuracy with the overhead.  
> > 
> > I disagree.  There is very little chance of getting out of this situation with the
> > current splitting.  Changing sampling and aggregation intervals doesn't actually help.
> > 
> > Let's draw out an example to discuss.
> > 
> > Toy state - taking just one block of memory.
> > 
> > 0 = not accessed page (very cold)
> > X = accessed page (extremely hot)
> > 
> > First few cycles - no accesses
> > 
> > in X.Regions list average value estimated by damon.
> > 
> > Region C is needed to set the max and will never be aggregated.
> > 
> > aggregation cycle then state.
> > 0.start
> > 0.accessed          0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 X X X
> > 0.regions (percent)|  A (0)          |   B (0)                         | C(1)|
> > 0.merge            |   A                                               | C   |
> > 0.split            |  A                                |     B         | C   |
> > 
> > After a few cycles, hot page
> > 1.start
> > 1.accessed          0 0 0 0 0 0 0 0 0 0 0 X 0 0 0 0 0 0 0 0 0 0 0 0 0 0
> > 1.regions (acc_cnt)|  A (1/18)                         |   B (0)       | C(1)|  
> 
>              ^ not count but ratio, right?

oops. Absolutely.

> 
> > 1.merge            |             A                                     | C   |
> > 1.split            |  A                    |                 B         | C   |
> > 2.start
> > 2.accessed          0 0 0 0 0 0 0 0 0 0 0 X 0 0 0 0 0 0 0 0 0 0 0 0 0 0
> > 2.regions (acc_cnt)|  A (1/12)             |               B (0)       | C(1)|
> > 2.merge            |             A                                     | C   |
> > 2.split            |  A      |                               B         | C   |
> > 3.start
> > 3.accessed          0 0 0 0 0 0 0 0 0 0 0 X 0 0 0 0 0 0 0 0 0 0 0 0 0 0
> > 3.regions (acc_cnt)|  A (0)  |               B (1/21)                  | C(1)|
> > 3.merge            |             A                                     | C   |
> > 3.split            |  A                |                     B         | C   |
> > 
> > Now make that 1000 pages long with the hot page at page 500.
> > So the average best case we will ever get is a 1/500 * number of sample period
> > between aggregations.  
> 
> So nice example, thank you!  Now I understand the point.
> 
> So, the problem is that we cannot find the small hot region near the _middle_
> because we split each region into only two subregions.

Exactly.  Pathological case :(

> 
> > 
> > So what are the chances of failing to aggregate on the sample after we split
> > at that optimal point? We need to successfully sample that one page enough that
> > we get it 10% of the time.
> > 
> > I 'think' this a case of where the 10% point is on the CDF of a binomial
> > f(1/N, M) where N is number of bins and Mis number of samples.
> > 
> > Using matlab online I think the best chance you ever get is when you take 10 samples
> > and need just one of them to be in the region.
> > 
> > p = 1 - binocdf(0,10,1/N)
> > For N = 500, p = 0.0198
> > For N = 1000, p = 0.0099
> > 
> > Someone with better maths than me can check.
> > 
> > Now this just got us to the point where we won't aggregate the region for one
> > round of aggregation.  We may split it again and if the resulting region is small
> > enough might not merge it the next aggregation cycle.
> > 
> > So I'd argue that allowing at least 2 repeats of splitting is well worth while.
> > It is just a couple of additional lines of code.  
> 
> Nice suggestion, I will apply this suggestion in the next spin.  It might be as
> below:
> 
>     if (nr_regions() < nr_max_regions / 4)
>             split_into_4_regions();
>     else if (nr_regions() < nr_max_regions / 2)
>             split_into_2_regions();
> 
> If this pseudo-code is missing some of your point, please let me know.

That's it.  My prototype was less efficient in that it just ran the
2 way split twice if we still had too few regions, but result is very
nearly the same (potentially some changes in the location of the split as
10-90% both times vs whatever limits you put in the 4 region version).

> 
> >   
> > > 
> > > Finally, I would like to keep code as simple as it can.
> > > 
> > > For same reasons, I would like to keep the code as currently is until real user
> > > problem is reported.  If you have different opinions, please feel free to yell
> > > at me.  
> > 
> > :)   
> 
> Appreciate your explanations and suggestions.

You are welcome.

Out of interest, do you have any comparative data on how 'accurate' the resulting
estimates are vs a more precise heatmap from a memory trace?

I'm looking at gathering such data but much happier to leverage your work if
you've already done it!

Thanks,

Jonathan

> 
> 
> Thanks,
> SeongJae Park