Feature Request: Add NUMA-Aware CPU Core Assignment for LXC Containers in Proxmox VE
Summary
Proxmox currently allows administrators to specify a number of CPU cores for LXC containers (e.g., “5 cores”), but the system does not assign those cores in a NUMA-aware or topology-aware manner. The chosen CPUs are distributed arbitrarily across all available NUMA nodes, leading to performance degradation for many latency-sensitive or memory-locality-sensitive workloads.
This proposal introduces an optional, advanced “NUMA-optimized CPU assignment” mode that ensures LXC containers are assigned logical CPU cores intelligently, grouped by physical NUMA node, improving locality and reducing cross-node latency.
Problem Description
When an LXC container is assigned a specific number of cores (for example: 5, 8, or 10), Proxmox:
- Allocates exactly that many logical CPUs
- But distributes them randomly across the system’s entire CPU topology
- With no awareness of NUMA boundaries
- With no grouping by socket
- Without pairing hyperthreads
- Without respecting locality
Example from a real system with 4 NUMA nodes:
Assigning 5 cores results in CPU selection such as:
These cores span multiple NUMA nodes, often all of them.
Impact
For real-time or latency-sensitive workloads (e.g., Rust game servers), this behavior causes:
- High CPU Pressure Stall (PSI)
- NUMA thrashing
- Cross-socket memory access latency
- Poor tick stability
- Performance inconsistencies
- Thread migration penalties
- Cache invalidation and reduced efficiency
Conversely, when administrators manually pin containers to cores within a single NUMA node and assign matching memory affinity, PSI drops dramatically (e.g., from 10% to nearly 0%), and workload performance becomes stable and predictable.
Why This Matters
NUMA locality is essential for modern systems:
- Multi-socket Xeon and EPYC hosts
- High-density container clusters
- Memory-intensive workloads
- Game servers
- Databases
- Web hosting stacks
- Virtualized or containerized HPC workloads
As hardware grows more NUMA-complex, topology-aware CPU assignment becomes increasingly important.
Proposed Solution: Add a NUMA-Optimized CPU Assignment Option in Advanced Settings
UI Proposal (Advanced tab):
Behavior:
If “NUMA-optimized CPU assignment” is enabled:
- Proxmox automatically chooses cores within the specified NUMA node(s).
- The selected CPUs are contiguous, topology-aware, and local.
- Memory affinity (cpuset.mems) is set automatically to match.
- No random scattering across NUMA nodes occurs.
- Linux scheduler migrations stay within the NUMA-local core group.
Optional:
If multiple NUMA nodes are selected, Proxmox divides cores proportionally.
Implementation Outline
The system would modify LXC config using:
Proxmox already supports cpuset generation internally but does not expose NUMA grouping.
Implementing this feature only requires:
- Reading host CPU topology
- Selecting NUMA-local CPUs
- Writing cpuset + mems entries
This is a lightweight enhancement with a large performance impact.
Real-World Example Result
On a 4-socket, 4-NUMA node machine:
Before NUMA pinning (using default Proxmox assignment):
- 10-15% CPU some pressure stall
- Rust game server lagging/stuttering
- High latency spikes
After manual NUMA pinning:
- 0–1% PSI
- Completely stable tickrate
- Significantly lower load
- Better performance and player experience
- No cross-socket thrashing
This is a dramatic improvement enabled purely by proper CPU locality.
Why This Should Be Added to Proxmox
- Easier for users than manual config editing
- Provides correct behavior on multi-socket systems
- Brings LXC CPU assignment to parity with VM NUMA features
- Reduces support burden (users often misinterpret “cores”)
- Improves performance for a huge class of workloads
- Makes Proxmox competitive with NUMA-aware orchestrators
This feature would be optional, non-breaking, and highly beneficial.
Conclusion
Proxmox currently assigns LXC CPUs without considering NUMA topology, which hinders performance on modern multi-socket systems. Adding an advanced setting for NUMA-optimized CPU assignment would:
- Improve performance
- Reduce scheduler overhead
- Lower PSI
- Make workloads more predictable
- Empower administrators
- Modernize Proxmox for NUMA-heavy hardware
This proposal is both technically feasible and highly valuable for real users.