On Mon, May 30, 2016 at 10:53:22AM -0700, Andi Kleen wrote:
>
> It should work the same on larger systems, the solution scales
> naturally to lots of sockets. It's not clear it'll help enough on systems
> with a lot more cores per socket, like a Xeon Phi. But for now it should
> be good enough.
One change which I'm currently making is to use kmalloc_node() instead
of kmalloc() for the per-NUMA node, and I suspect *that* is going
to make a quite a lot of different on those systems where the ratio of
remote to local memory access times is larger (as I assume it probably
would be on really big systems).
- Ted
