NUMA vs PostgreSQL
summary
Shows experimental patches that add NUMA visibility, force or interleave shared memory, and change buffer replacement and locking to reduce cross-node contention; prototypes report big wins (one case reaches ~550k TPS, another halves jitter from 382 ms to ~? using interleave) and a 7x scalability jump when eliminating shared-buffer hot spots. Uses Linux numactl and kernel stats, perf cache-to-cache tracing and ugly prototypes to prove interleave shared buffers and a local fast-path locking idea work; proposes adding NUMA-aware diagnostics to pg catalog.
AI-written summary · about
chapters (11)
- 0:00Introduction to NUMA
- 2:09Modern hardware challenges
- 6:30Impact of NUMA on latency
- 12:29Official vs. inofficial NUMA
- 15:43NUMA behavior on Linux
- 18:49Visibility and diagnostics
- 23:45Solving memory imbalances
- 30:07Optimizing buffer replacement
- 35:19Fast path buffer locks
- 39:53Future hardware trends
- 44:00Q&A and closing
related talks
26mRethinking PostgreSQL Performance in the Age of Monster Hardware
45mEnhancing PostgreSQL Plasticity: New Frontiers in Memory Management
28mFast-path locking improvements in PG18
56mPerformance Improvements in PostgreSQL9 2
30mPostgreSQL and Linux Kernel interactions
26mLinux and PostgreSQL in the Multiverse of Connections
49mAnalyzing cacheline contention using perf c2c
56mPostgreSQL at low Level: stay curious!