FOSSWire
databases-storage

PostgreSQL 18 introduces libnuma support for better NUMA performance

Brief written by the FOSSWire AI editor from Planet PostgreSQL

1 min read Impact: 6/10

Key Takeaway

PostgreSQL 18 includes patches to micromanage NUMA policies, interleaving shared resources while localizing backend memory to improve performance on large multi-processor systems.

Summary

PostgreSQL 18 will feature a significant update to how the database handles Non-Uniform Memory Access (NUMA) architectures via libnuma. The changes, contributed by Tomas Vondra, allow the system to interleave shared resources like shared buffers and the ProcArray while ensuring backend-local allocations remain on the local NUMA node. This approach aims to reduce memory access penalties and lock contention on large-scale hardware.

Technical improvements in PostgreSQL 18 NUMA handling

  • Shared Resource Interleaving: Shared buffers and freelists are explicitly interleaved across all available nodes to maximize throughput.
  • ProcArray Transformation: The ProcArray has been converted from an array to a linked list (renamed to PGProc) to allow it to be spread across different NUMA nodes.
  • Backend Memory Localization: Local allocations for backends are now set to a local allocation policy, preventing the performance penalty of distributed local memory.
  • New System View: A new view, pg_shmem_allocations_numa, has been added to monitor where memory is stored.
  • Huge Pages Requirement: The use of huge pages becomes even more critical to prevent buffer pages from being split across different memory controllers.

Why it matters

These changes significantly optimize PostgreSQL for high-end enterprise servers and large-scale database deployments running on multi-socket Linux systems.

Original reference: postgr.es →