Re: [PATCH 0/2] Add a new scheme to support demotion on tiered memory system

[SeongJae Park <sj@kernel.org>]

On Tue, 21 Dec 2021 22:32:24 +0800 Baolin Wang <baolin.wang@linux.alibaba.com> wrote:

> 
> 
> On 12/21/2021 9:26 PM, SeongJae Park wrote:
> > Hi Baolin,
> > 
> > On Tue, 21 Dec 2021 17:18:02 +0800 Baolin Wang <baolin.wang@linux.alibaba.com> wrote:
> > 
> >> Hi,
> >>
> >> Now on tiered memory system with different memory types, the reclaim path in
> >> shrink_page_list() already support demoting pages to slow memory node instead
> >> of discarding the pages. However, at that time the fast memory node memory
> >> wartermark is already tense, which will increase the memory allocation latency
> >> during page demotion. So a new method from user space demoting cold pages
> >> proactively will be more helpful.
> >>
> >> We can rely on the DAMON in user space to help to monitor the cold memory on
> >> fast memory node, and demote the cold pages to slow memory node proactively to
> >> keep the fast memory node in a healthy state.
> >>
> >> This patch set introduces a new scheme named DAMOS_DEMOTE to support this feature,
> >> and works well from my testing. Any comments are welcome. Thanks.
> > 
> > I like the idea, thank you for these patches!  If possible, could you share
> > some details about your tests?
> 
> Sure, sorry for not adding more information about my tests.

No problem!

> 
> My machine contains 64G DRAM + 256G AEP(persistent memory), and you 
> should enable the demotion firstly by:
> echo "true" > /sys/kernel/mm/numa/demotion_enabled
> 
> Then I just write a simple test case like below to mmap some anon 
> memory, and then just read and write half of the mmap buffer to let 
> another half to be cold enough to demote.
> 
> int main()
> {
>          int len = 50 * 1024 * 1024;
>          int scan_len = len / 2;
>          int i, ret, j;
>          unsigned long *p;
> 
>          p = mmap(NULL, len, PROT_READ | PROT_WRITE,
>                   MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
>          if (p == MAP_FAILED) {
>                  printf("failed to get memory\n");
>                  return -1;
>          }
> 
>          for (i = 0; i < len / sizeof(*p); i++)
>                  p[i] = 0x55aa;
> 
>          /* Let another half of buffer to be cold */
>          do {
>                  for (i = 0; i < scan_len / sizeof(*p); i++)
>                          p[i] = 0x55aa;
> 
>                  sleep(2);
> 
>                  for (i = 0; i < scan_len / sizeof(*p); i++)
>                          j +=  p[i] >> 2;
>          } while (1);
> 
>          munmap(p, len);
>          return 0;
> }
> 
> After setting the atts/schemes/target_ids, then start monitoring:
> echo 100000 1000000 1000000 10 1000 > /sys/kernel/debug/damon/attrs
> echo 4096 8192000 0 5 10 2000 5 1000 2097152 5000 0 0 0 0 0 3 2 1 > 
> /sys/kernel/debug/damon/schemes
> 
> After a while, you can check the demote statictics by below command, and 
> you can find the demote scheme is applied by demoting some cold pages to 
> slow memory (AEP) node.
> 
> cat /proc/vmstat | grep "demote"
> pgdemote_direct 6881

Thank you for sharing this great details!

I was just wondering if you have tested and measured the effects of the memory
allocation latency increase during the page demotion, which invoked by
shrink_page_list(), and also if you have measured how much improvement can be
achieved with DAMON-based demotion in the scenario.  Seems that's not the case,
and I personally think that information is not essential for this patch, so I
see no problem here.  But, if you have tested or have a plan to do that, and if
you could, I think sharing the results on this cover letter would make this
even greater.


Thanks,
SJ