Hi all,
My truenas host has 2 sockets, 2 NUMA nodes.
I created a vm with 2 vCPUs, pinned to the cores such that each vCPU gets cores from the corisponding socket.
The guest shows all the cores, /proc/cpuinfo shows their ‘physical id’ correctly but they are all on the same NUMA node in the guest OS, not split as i would expect
How do i configure a vm such that the guest OS has internal NUMA nodes that reflect that of the host?
(if it matters, this is for CPU inference on large LLMs so it’s i need a single instance of ollama to distribute layers across the numa nodes to maximize memory bandwidth)
thanks