HMSDK: Enhancing Software Support for CXL Memory
summary
HMSDK provides two concrete kernel features now upstream: weighted interleaving (merged around Linux 6.9) and DAMON-based tiering (merged around 6.11). Weighted interleaving shows measured bandwidth gains of about 21–31% in single-socket tests and a socket-aware variant (in HMSDK 4.0 as a local patch) avoids UPI cross-socket penalties. DAMON runs at 100 ms sampling intervals, migrates hot regions (up to ~180 GB observed) and cuts worst-case Llama runtime ~23% versus CXL-only; a separate benchmark shows up to ~34% runtime reduction. HMSDK also ships a BW-Profiler (bwprof) tool.
AI-written summary · about
chapters (7)
- 0:00HMSDK and CXL memory intro
- 3:11Bandwidth extension mode
- 7:25Tiered memory capacity
- 18:33BWProbe bandwidth tool
- 22:03HMSDK open-source timeline
- 25:16Summary
- 26:42Q&A on upstreaming
related talks
38mMaximizing Heterogeneous Memory Bandwidth in Multi-Socket Systems
42mDAMON Recipes: Ways to Save Memory Using a Linux Kernel Subsystem in the Real World
26mCXL Memory Tiering for heterogenous computing- 42mDAMON Recipes: Ways to Save Memory Using a Linux Kernel Subsystem in the Real World
43mExpanding the Memory Limits with CXL Software Stack
37mMeta’s CXL Journey and Learnings in Linux
29mCXL HDM-DB support for Linux
42mMessing up Your NUMA Topology with CXL