On 8/19/20 11:14 PM, David Gibson wrote:
On Fri, Aug 14, 2020 at 05:54:17PM -0300, Daniel Henrique Barboza wrote:
NVLink2 GPUs are allocated in their own NUMA node, at maximum
distance from every other resource in the board. The existing
logic makes some assumptions that don't scale well:
-
On Fri, Aug 14, 2020 at 05:54:17PM -0300, Daniel Henrique Barboza wrote:
> NVLink2 GPUs are allocated in their own NUMA node, at maximum
> distance from every other resource in the board. The existing
> logic makes some assumptions that don't scale well:
>
> - only NVLink2 GPUs will ever require
NVLink2 GPUs are allocated in their own NUMA node, at maximum
distance from every other resource in the board. The existing
logic makes some assumptions that don't scale well:
- only NVLink2 GPUs will ever require such mechanism, meaning
that the GPU logic is tightly coupled with the NUMA setup