On Thu, 4 Feb 2021 16:31:36 +0100 SeongJae Park <sjpark@xxxxxxxxxx> wrote:
> From: SeongJae Park <sjpark@xxxxxxxxx>
>
[...]
>
> Introduction
> ============
>
> DAMON is a data access monitoring framework for the Linux kernel. The core
> mechanisms of DAMON called 'region based sampling' and 'adaptive regions
> adjustment' (refer to 'mechanisms.rst' in the 11th patch of this patchset for
> the detail) make it
>
> - accurate (The monitored information is useful for DRAM level memory
> management. It might not appropriate for Cache-level accuracy, though.),
> - light-weight (The monitoring overhead is low enough to be applied online
> while making no impact on the performance of the target workloads.), and
> - scalable (the upper-bound of the instrumentation overhead is controllable
> regardless of the size of target workloads.).
>
> Using this framework, therefore, several memory management mechanisms such as
> reclamation and THP can be optimized to aware real data access patterns.
> Experimental access pattern aware memory management optimization works that
> incurring high instrumentation overhead will be able to have another try.
>
> Though DAMON is for kernel subsystems, it can be easily exposed to the user
> space by writing a DAMON-wrapper kernel subsystem. Then, user space users who
> have some special workloads will be able to write personalized tools or
> applications for deeper understanding and specialized optimizations of their
> systems.
>
I realized I didn't introduce a good, intuitive example use case of DAMON for
profiling so far, though DAMON is not for only profiling. One straightforward
and realistic usage of DAMON as a profiling tool would be recording the
monitoring results with callstack and visualize those by timeline together.
For example, below link shows that visualization for a realistic workload,
namely 'fft' in SPLASH-2X benchmark suite. From that, you can know there are
three memory access bursting phases in the workload and
'FFT1DOnce.cons::prop.2()' looks responsible for the first and second hot
phase, while 'Transpose()' is responsible for the last one. Now the programmer
can take a deep look in the functions and optimize the code (e.g., adding
madvise() or mlock() calls).
https://damonitor.github.io/temporal/damon_callstack.png
We used the approach for 'mlock()'-based optimization of a range of other
realistic benchmark workloads. The optimized versions achieved up to about
2.5x performance improvement under memory pressure[1].
Note: I made the uppermost two figures in above 'fft' visualization (working
set size and access frequency of each memory region by time) via the DAMON user
space tool[2], while the lowermost one (callstack by time) is made using perf
and speedscope[3]. We have no descent and totally automated tool for that yet
(will be implemented soon, maybe under perf as a perf-script[4]), but you could
reproduce that with below commands.
$ # run the workload
$ sudo damo record $(pidof <your_workload>) &
$ sudo perf record -g $(pidof <your_workload>)
$ # after your workload finished (you should also finish perf on your own)
$ damo report wss --sortby time --plot wss.pdf
$ damo report heats --heatmap freq.pdf
$ sudo perf script | speedscope -
$ # open wss.pdf and freq.pdf with our favorite pdf viewer
[1] https://linuxplumbersconf.org/event/4/contributions/548/attachments/311/590/damon_ksummit19.pdf
[2] https://lore.kernel.org/linux-mm/20201215115448.25633-8-sjpark@xxxxxxxxxx/
[3] https://www.speedscope.app/
[4] https://lore.kernel.org/linux-mm/20210107120729.22328-1-sjpark@xxxxxxxxxx/
