Deprecating zone_reclaim_mode
summary
zone_reclaim_mode, added in 2005, forces local direct reclaim when a local zone runs out instead of falling back to remote allocation; on a two‑node 125 GB system with 60 GB anon and 100 GB pagecache it produces “abysmal” performance for mixed workloads, while pure latency‑sensitive random memory workloads show gains that membind matches or exceeds. RFC proposes deprecating the sysctl (make it noop the sysctl or remove) and using NUMA balancing, memory.reclaim, membind, or tiering instead, which also lets allocator code remove many watermark checks.
AI-written summary · about
chapters (11)
- 0:00Introduction to zone_reclaim_mode
- 1:22How zone_reclaim_mode works
- 2:24When zone_reclaim_mode harms performance
- 3:06When zone_reclaim_mode helps
- 4:09Reasons to deprecate zone_reclaim_mode
- 5:10Deprecation strategies for user-facing features
- 6:35Existing use cases and alternatives
- 8:20Historical context of zone_reclaim_mode
- 9:26Tiered memory and NUMA systems
- 10:30Potential improvements and future steps
- 12:01Closing remarks and Q&A
