Towards a Tool for Access-Affinity Based Structure Reordering in the Linux Kernel
Talkat Open Source Summit India 2026 · on Wednesday, 2026-06-17· at 15:35· Lotus Ballroom 3 (Level 3)·
summary
A prototype tool traces field-level loads/stores on struct rq, builds an access-affinity graph by counting temporal co‑access within a short sliding window, and uses hierarchical clustering to propose cache-line-aware reorderings. Data collection explores KVM/QEMU TCG traces, HDM hardware traces, HTM ideas and MemFriend; perf sampling is rejected as too lossy. Applied with a waitstressor benchmark, cache misses drop from 36.2B (13.9%) to 25.1B (9.4%) and idle_cpu() misses fall 6.40% → 3.11%.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:25Problem with scheduler stats
- 3:52Cache misses in structures
- 7:31Field reordering rules
- 9:39Existing reordering approaches
- 12:11Tool pipeline overview
- 15:01Data collection methods
- 18:33Access affinity graph
- 21:04Field clustering technique
- 23:11Results and improvements
- 27:54Future challenges and plans
- 31:27Q&A session
related talks
48mProfiling data structures
21mPreview: Operating Systems
28mRefining Data Structure Implementations in the Linux Kernel for Improve..
27mTuning in-kernel routines on RISC-V
35mOpening the Box: Diagnosing Operating-System Task-Scheduler...
20mPercpu hashtab traversal measurement study
01h 11mPerformance engineering - being friendly to your hardware
52mEffectively Measure and Reduce Kernel Latencies for Real-time Constraints