* Ingo Molnar <[email protected]> wrote:
> So my thinking here is: if the NUMA balancing code (which is node granular at
> the moment and uses node masks, etc.) is extended to be CPU granular (which
> is a
> big task in itself), then the two problems can be 'unified':
>
> - the NUMA balancing code inputs arbitrarly CPU (node) affinity masks from
> the
> MM code into the scheduler.
>
> - the scheduler syscall ABI (and other configuration sources) inputs
> arbitrary
> CPU affinity masks into the scheduler.
>
> it's a similar problem, with two (minor looking) complication:
btw., this highlights how hard the optimization problem is: the NUMA balancing
code is (at least ...) O(nr_nodes^2) complex - but we had O(nr_nodes^3) passes
too
in some of the NUMA balancing submissions...
We'd upgrade that to O(nr_cpus^2), which is totally unrealistic with 16,000
CPUs
even in a slowpath - but it would probably cause problems even with 120 CPUs.
It
will get quadratically worse as the number of CPUs in a system increases on its
current exponential trajectory ...
So the safest bet would be to restrict any 'perfect' balancing attempts to node
boundaries. Which won't solve the problem you outlined to begin with.
Thanks,
Ingo
