Live in a world with multiple memory types
summary
Kernel support lands for NUMA balancing-based promotion and explicit mm parameters, with demotion implemented as page reclaim migration to PMEM and promotion driven by access hotness plus locality migration. A new abstract distance metric combines latency and bandwidth (workload-dependent, user-overridable) and a user-visible memory-type devices / memory-tier model exposes types, tiers, and knobs. Evaluation on a two-socket DRAM+DCPM server shows large gains on microbenchmarks and about a 12% throughput boost for a Redis-like in-memory DB versus a base placement.
AI-written summary · about
chapters (12)
- 0:00Introduction to multiple memory
- 1:51Memory management mechanisms
- 5:29Emerging memory types
- 6:11Abstract distance concept
- 10:09User space memory types interface
- 14:31Memory tiers and grouping
- 19:24Open discussion and feedback
- 24:11Page placement optimization
- 27:45Demotion and cross-socket control
- 31:22Performance evaluation results
- 34:17Current gaps and future work
- 36:31Q&A session
related talks
48mPersistent Memory Usage within Linux Environment
30mdrm and cgroup memory management - two great tastes
15mNomad: Non-Exclusive Memory Tiering via Transactional Page Migration
11mAdvanced Parallel Memory Virtualization
26mLet's demystify memory
26mCXL Memory Tiering for heterogenous computing
42mDAMON Recipes: Ways to Save Memory Using a Linux Kernel Subsystem in the Real World
21mPreview: Operating Systems