Re: [PATCH v3] mm: migrate: Support multiple target nodes demotion

[Baolin Wang <baolin.wang@linux.alibaba.com>]

On 2021/11/12 11:02, Huang, Ying wrote:
> Baolin Wang <baolin.wang@linux.alibaba.com> writes:
> 
>> On 2021/11/12 10:44, Huang, Ying wrote:
>>> Baolin Wang <baolin.wang@linux.alibaba.com> writes:
>>>
>>>> We have some machines with multiple memory types like below, which
>>>> have one fast (DRAM) memory node and two slow (persistent memory) memory
>>>> nodes. According to current node demotion policy, if node 0 fills up,
>>>> its memory should be migrated to node 1, when node 1 fills up, its
>>>> memory will be migrated to node 2: node 0 -> node 1 -> node 2 ->stop.
>>>>
>>>> But this is not efficient and suitbale memory migration route
>>>> for our machine with multiple slow memory nodes. Since the distance
>>>> between node 0 to node 1 and node 0 to node 2 is equal, and memory
>>>> migration between slow memory nodes will increase persistent memory
>>>> bandwidth greatly, which will hurt the whole system's performance.
>>>>
>>>> Thus for this case, we can treat the slow memory node 1 and node 2
>>>> as a whole slow memory region, and we should migrate memory from
>>>> node 0 to node 1 and node 2 if node 0 fills up.
>>>>
>>>> This patch changes the node_demotion data structure to support multiple
>>>> target nodes, and establishes the migration path to support multiple
>>>> target nodes with validating if the node distance is the best or not.
>>>>
>>>> available: 3 nodes (0-2)
>>>> node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
>>>> node 0 size: 62153 MB
>>>> node 0 free: 55135 MB
>>>> node 1 cpus:
>>>> node 1 size: 127007 MB
>>>> node 1 free: 126930 MB
>>>> node 2 cpus:
>>>> node 2 size: 126968 MB
>>>> node 2 free: 126878 MB
>>>> node distances:
>>>> node   0   1   2
>>>>     0:  10  20  20
>>>>     1:  20  10  20
>>>>     2:  20  20  10
>>>>
>>>> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
>>
>> snip
>>
>>>>    	/*
>>>>    	 * 'next_pass' contains nodes which became migration
>>>> @@ -3192,6 +3281,14 @@ static int __init migrate_on_reclaim_init(void)
>>>>    {
>>>>    	int ret;
>>>>    +	/*
>>>> +	 * Ignore allocation failure, if this kmalloc fails
>>>> +	 * at boot time, we are likely in bigger trouble.
>>>> +	 */
>>>> +	node_demotion = kmalloc_array(nr_node_ids,
>>>> +				      sizeof(struct demotion_nodes),
>>>> +				      GFP_KERNEL);
>>>> +
>>> I think we should WARN_ON() here.
>>
>> In this unlikey case, I think the mm core will print more information,
>> IMHO WARN_ON() will help little. Anyway no strong opinion on
>> this. Other than that, can I get your reviewed-by tag with this nit
>> fixed? Thanks.
> 
> Yes.  Please add my "reviewed-by" after changing this.

OK. Thanks for your reviewing.