RE: 5.11-rc4+git: Shortest NUMA path spans too many nodes

[Valentin Schneider]

On 22/01/21 11:09, Song Bao Hua (Barry Song) wrote:
>> -----Original Message-----
>> From: Dietmar Eggemann [mailto:dietmar.eggemann@xxxxxxx]
>> > For example, every cpu with the below numa_distance can have
>> > "groups don't span domain->span":
>> > node   0   1   2   3
>> >   0:  10  12  20  22
>> >   1:  12  10  22  24
>> >   2:  20  22  10  12
>> >   3:  22  24  12  10
>>                              2     20     2
>> So this should look like: 1 --- 0 ---- 2 --- 3
>
> Yes. So here we are facing another problem:
> kernel/sched/topology.c has an assumption that:
> node_distance(0,j) includes all distances in
> node_distance(i,j).
>
> void sched_init_numa(void)
> {
>       ...
>        *
>        * Assumes node_distance(0,j) includes all distances in
>        * node_distance(i,j) in order to avoid cubic time.
>        */
>       next_distance = curr_distance;
>       for (i = 0; i < nr_node_ids; i++) {
>               for (j = 0; j < nr_node_ids; j++) {
>                       for (k = 0; k < nr_node_ids; k++)
> }
>
> but obviously we are not this case. Right now, we are getting
> some performance decrease due to this, probably I'll start another
> thread for it.
>

It's not too difficult to solve that one; I must still have a patch laying
somewhere using a bitmap - this relies on the ACPI spec stating distance
values are 8bit, which gives us a reasonable bound for the bitmap size.

Let me fish this out.

> Thanks
> Barry