Enable NUMA and NUMA balancing

[rucoder]

• EVE runs on many big servers with multiple sockets these days. Let's enable NUMA and NUMA balancing in order to improve memory allocation strategy.
• it is also required for RT workloads and lets libpqos to detect NUMA topology, however NUMA balancing should be disable for PREEMPT_RT to avoid page table shuffling and this is done by default so no changes required in rt.fragment

Clarification on possible impact

• Linux CPU IDs (cpu0, cpu1, …) come from APIC IDs → logical CPU numbers assigned during SMP init, so EVE will not see any CPU re-numbering
• QEMU behavior remains identical.
• only one thing changes: where kernel allocates memory. with NUMA enabled it will try to avoid cross-socket memory allocation

[rucoder]

@OhmSpectator i hope it won't affect CPU pinning in EVE. Do we assume "flat" CPU core topology ?

[OhmSpectator]

We assume. But we operate only with CPU IDs. So, I don't think NUMA-enabling will change them.

On the other side, this change can affect a lot of other aspect of the system functionality.

[rene]

@rucoder , we don't have any NUMA configuration for QEMU and containerd, so these should be "not NUMA" aware, I'm a bit concerned this can decrease the performance instead of improve.... is this really needed? From how many cores should we see NUMA improvements?

[rucoder]

@rucoder , we don't have any NUMA configuration for QEMU and containerd, so these should be "not NUMA" aware, I'm a bit concerned this can decrease the performance instead of improve.... is this really needed? From how many cores should we see NUMA improvements?

NUMA is not about cores, it is about memory allocation. Kernel will allocate memory on the DIMM that is attached to the same CPU socket. Neither contained nor qemu care about it.

[rucoder]

@OhmSpectator @rene I updated PR description with possible impact section