Re: [PATCH RFC 3/4] mm/vmstat: rename pgdemote_* to pgdemote_dst_* and add pgdemote_src_*

[Huang, Ying]

"Yasunori Gotou (Fujitsu)" <y-goto@xxxxxxxxxxx> writes:

>> > Hello,
>> >
>> >> On 02/11/2023 13:45, Huang, Ying wrote:
>> >> > Li Zhijian <lizhijian@xxxxxxxxxxx> writes:
>> >> >
>> >> >> pgdemote_src_*: pages demoted from this node.
>> >> >> pgdemote_dst_*: pages demoted to this node.
>> >> >>
>> >> >> So that we are able to know their demotion per-node stats by checking
>> this.
>> >> >>
>> >> >> In the environment, node0 and node1 are DRAM, node3 is PMEM.
>> >> >>
>> >> >> Global stats:
>> >> >> $ grep -E 'demote' /proc/vmstat
>> >> >> pgdemote_src_kswapd 130155
>> >> >> pgdemote_src_direct 113497
>> >> >> pgdemote_src_khugepaged 0
>> >> >> pgdemote_dst_kswapd 130155
>> >> >> pgdemote_dst_direct 113497
>> >> >> pgdemote_dst_khugepaged 0
>> >> >>
>> >> >> Per-node stats:
>> >> >> $ grep demote /sys/devices/system/node/node0/vmstat
>> >> >> pgdemote_src_kswapd 68454
>> >> >> pgdemote_src_direct 83431
>> >> >> pgdemote_src_khugepaged 0
>> >> >> pgdemote_dst_kswapd 0
>> >> >> pgdemote_dst_direct 0
>> >> >> pgdemote_dst_khugepaged 0
>> >> >>
>> >> >> $ grep demote /sys/devices/system/node/node1/vmstat
>> >> >> pgdemote_src_kswapd 185834
>> >> >> pgdemote_src_direct 30066
>> >> >> pgdemote_src_khugepaged 0
>> >> >> pgdemote_dst_kswapd 0
>> >> >> pgdemote_dst_direct 0
>> >> >> pgdemote_dst_khugepaged 0
>> >> >>
>> >> >> $ grep demote /sys/devices/system/node/node3/vmstat
>> >> >> pgdemote_src_kswapd 0
>> >> >> pgdemote_src_direct 0
>> >> >> pgdemote_src_khugepaged 0
>> >> >> pgdemote_dst_kswapd 254288
>> >> >> pgdemote_dst_direct 113497
>> >> >> pgdemote_dst_khugepaged 0
>> >> >>
>> >> >>  From above stats, we know node3 is the demotion destination which
>> >> >> one the node0 and node1 will demote to.
>> >> >
>> >> > Why do we need these information?  Do you have some use case?
>> >>
>> >> I recall our customers have mentioned that they want to know how much
>> >> the memory is demoted to the CXL memory device in a specific period.
>> >
>> > I'll mention about it more.
>> >
>> > I had a conversation with one of our customers. He expressed a desire
>> > for more detailed profile information to analyze the behavior of
>> > demotion (and promotion) when his workloads are executed.
>> > If the results are not satisfactory for his workloads, he wants to
>> > tune his servers for his workloads with these profiles.
>> > Additionally, depending on the results, he may want to change his server
>> configuration.
>> > For example, he may want to buy more expensive DDR memories rather than
>> cheaper CXL memory.
>> >
>> > In my impression, our customers seems to think that CXL memory is NOT as
>> reliable as DDR memory yet.
>> > Therefore, they want to prepare for the new world that CXL will bring,
>> > and want to have a method for the preparation by profiling information as
>> much as possible.
>> >
>> > it this enough for your question?
>> 
>> I want some more detailed information about how these stats are used?
>> Why isn't per-node pgdemote_xxx counter enough?
>
> I rechecked the customer's original request.
>
> - If a memory area is demoted to a CXL memory node, he wanted to analyze how it affects performance
>  of their workload, such as latency. He wanted to use CXL Node memory usage as basic
>  information for the analysis.
>
> - If he notices that demotion occurs well on a server and CXL memories are used 85% constantly, he 
>   may want to add DDR DRAM or select some other ways to avoid demotion.
>   (His image is likely Swap free/used.)
>   IIRC, demotion target is not spread to all of the CXL memory node, right? 
>   Then, he needs to know how CXL memory is occupied by demoted memory.
>
> If I misunderstand something, or you have any better idea,
> please let us know. I'll talk with him again. (It will be next week...)


To check CXL memory usage, /proc/PID/numa_maps,
/sys/fs/cgroup/CGROUP/memory.numa_stat, and
/sys/devices/system/node/nodeN/meminfo can be used for process, cgroup,
and NUMA node respectively.  Is this enough?

--
Best Regards,
Huang, Ying

>> >
>> >>
>> >>
>> >> >>>   	mod_node_page_state(NODE_DATA(target_nid),
>> >> >>> -		    PGDEMOTE_KSWAPD + reclaimer_offset(),
>> >> nr_succeeded);
>> >> >>> +		    PGDEMOTE_DST_KSWAPD + reclaimer_offset(),
>> >> nr_succeeded);
>> >>
>> >> But if the *target_nid* is only indicate the preferred node, this
>> >> accounting maybe not accurate.
>> >>

[snip]