Direct answer
On multi-socket or NUMA servers, memory access latency depends on where processes and memory pages are allocated; this is different from simply having many CPU cores.
How to check it
- Use
numactl --hardware(when installed) to inspect topology, then check actual CPU scheduling/affinity and memory pressure on the host. Compare controlled runs rather than pinning production Java blindly. - Record the original configuration or test conditions, capture a profile or a repeatable observation, and identify the affected world or player group.
- Make the measurement under the same player count and workload you will later use to verify the change.
What to change
Potential optimizations depend on memory placement, GC behavior and whether the Minecraft workload truly spans NUMA nodes. For ordinary single-socket VPS hosting, NUMA tuning is unlikely to be the first useful fix.
Aggressive affinity can reduce available CPU and worsen pauses. Document and revert any host-level changes if measurements do not improve.
Software and version notes
Check the active version and generated configuration before copying paths or values. Different forks and mod loaders may implement this behavior differently.
This article is written for Paper / compatible servers. Configuration keys and defaults can move between versions. The source material checked for this page includes PaperMC profiling guide. If your server is Folia, Leaf, Purpur, Fabric, Forge, NeoForge, or another fork/loader, do not assume a Paper-only setting exists or behaves identically.
Verify the result
Repeat the same test with the same software build, player count and world activity. Compare the targeted symptom and any spark measurements, then check that intended game mechanics still work. If the change does not produce a reproducible improvement, revert to the recorded baseline.