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

[Huang, Ying]

Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx> writes:

> On 2021/11/12 10:44, Huang, Ying wrote:
>> Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx> 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@xxxxxxxxxxxxxxxxx>
>
> 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.

Best Regards,
Huang, Ying