Re: [RFC PATCH] mm: migrate: Add new node demotion strategy

[Huang, Ying]

Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx> writes:

> On 2021/11/8 14:48, Huang, Ying writes:
>> Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx> writes:
>> 
>>> On 2021/11/7 23:20, Dave Hansen wrote:
>>>> On 11/7/21 1:33 AM, Baolin Wang wrote:
>>>>> Thanks for your suggestion. After some thinking, can we change the
>>>>> node_demotion[] structure like below? Which means one source node can be
>>>>> demoted to mutiple target node, and we can set up the target node mask
>>>>> according to the node distance. How do you think? Thanks.
>>>>>
>>>>> static nodemask_t node_demotion[MAX_NUMNODES] __read_mostly =
>>>>>       {[0 ... MAX_NUMNODES - 1] = NODE_MASK_NONE};
>>>> How large is that in the worst case?
>>>
>>> For the worst case (MAX_NUMNODES=1024), the size of the node_demotion
>>> is 131072 bytes, while the size of original data structure is 4096
>>> bytes. Maybe we can allocate the node_demotion dynamically?
>> Per my understanding, in most cases, the number of demotion target
>> nodes
>> should be quite small.  So why not restrict the number of demotion
>> target nodes to make it some kind of simple array?
>
> Yes, agree. Something like below is reasonable for you?
>
> #define DEMOTION_TARGET_NODES 16
> typedef struct { DECLARE_BITMAP(bits, DEMOTION_TARGET_NODES); }
> demotemask_t;
>
> static demotemask_t node_demotion[MAX_NUMNODES];

I don't think we need a bitmap.  May be something as following,

#define DEMOTION_TARGET_NODES 15
struct demotion_nodes {
  unsigned short nr;
  unsigned short nodes[DEMOTION_TARGET_NODES];
};

Best Regards,
Huang, Ying