Re: [RFC v11 3/8] mm/damon: Implement data access monitoring-based operation schemes

[David Hildenbrand <david@redhat.com>]

On 09.06.20 11:17, SeongJae Park wrote:
> On Tue, 9 Jun 2020 10:47:45 +0200 David Hildenbrand <david@redhat.com> wrote:
> 
>> On 09.06.20 08:53, SeongJae Park wrote:
>>> From: SeongJae Park <sjpark@amazon.de>
>>>
>>> In many cases, users might use DAMON for simple data access aware
>>> memory management optimizations such as applying an operation scheme to
>>> a memory region of a specific size having a specific access frequency
>>> for a specific time.  For example, "page out a memory region larger than
>>> 100 MiB but having a low access frequency more than 10 minutes", or "Use
>>> THP for a memory region larger than 2 MiB having a high access frequency
>>> for more than 2 seconds".
>>>
>>> To minimize users from spending their time for implementation of such
>>> simple data access monitoring-based operation schemes, this commit makes
>>> DAMON to handle such schemes directly.  With this commit, users can
>>> simply specify their desired schemes to DAMON.
>>
>> What would be the alternative? How would a solution where these policies
>> are handled by user space (or inside an application?) look like?
> 
> Most simple form of the altermative solution would be doing offline data access
> pattern profiling using DAMON and modifying the application source code or
> system configuration based on the profiling results.
> 
> More automated alternative solution would be a daemon constructed with two
> modules:
> 
>  - monitor: monitors the data access pattern of the workload via the DAMON
>    debugfs interface
>  - memory manager: based on the monitoring result, make appropriate memory
>    management changes via mlock(), madvise(), sysctl, etc.
> 
> The daemon would be able to run inside the application process as a thread, or
> outside as a standalone process.  If the daemon could not run inside the
> application process, the memory management changes it could make would be
> further limited, though, as mlock() and madvise() would not be available.  The
> madvise_process(), which is already merged in the next tree, would be helpful
> in this case.
> 
>>>
>>> Each of the schemes is composed with conditions for filtering of the
>>> target memory regions and desired memory management action for the
>>> target.  Specifically, the format is::
>>>
>>>     <min/max size> <min/max access frequency> <min/max age> <action>
>>>
>>> The filtering conditions are size of memory region, number of accesses
>>> to the region monitored by DAMON, and the age of the region.  The age of
>>> region is incremented periodically but reset when its addresses or
>>> access frequency has significantly changed or the action of a scheme was
>>> applied.  For the action, current implementation supports only a few of
>>> madvise() hints, ``MADV_WILLNEED``, ``MADV_COLD``, ``MADV_PAGEOUT``,
>>> ``MADV_HUGEPAGE``, and ``MADV_NOHUGEPAGE``.
>>
>> I am missing some important information. Is this specified for *all*
>> user space processes? Or how is this configured? What are examples?
>>
>> E.g., messing with ``MADV_HUGEPAGE`` vs. ``MADV_NOHUGEPAGE`` of random
>> applications can change the behavior/break these applications. (e.g., if
>> userfaultfd is getting used and the applciation explicitly sets
>> MADV_NOHUGEPAGE).
> 
> Only monitoring target processes will be applied.  The monitoring target
> processes can be specified by writing the process ids to 'pids' debugfs file or
> constructing the 'struct damon_ctx' via the programming interface.
> 
> I will refine the commit message to make the points clearer, in the next spin.

Understood, so a process configures damon to only modify its mappings.
thanks for clarifying! This makes exposing the do_madvise() look less
dangerous.


-- 
Thanks,

David / dhildenb