Archer: Adaptive Memory Compression with Page-Association-Rule Awareness for High-Speed..
summary
Archer implements an association-rule-aware memory compression framework integrated into the Linux kernel and ZRAM, with a prototype on Pixel and Huawei phones. It finds ~26.3% of anonymous pages are strongly associated, mines them online with a FSG sliding-window footprint stream and an in-kernel FP-list, and groups associated pages into an ACR adaptive compression region indexed by virtual handles. Result: average app launch 1.55× faster, photographic speed 1.42× and frame rate 1.31× versus state-of-the-art, with much lower tail latency.
AI-written summary · about
chapters (8)
- 0:00Archer introduction
- 1:43Page compression problems
- 3:34Performance degradation cause
- 5:20Large-grain compression insight
- 6:50Archer's three components
- 9:34Adaptive compression regions
- 11:00Mobile device evaluation
- 14:55Summary
related talks
01h 11mMachine Learning Acceleration in an Embedded World
58mKeynote: Enabling the AI Revolution: Next-Generation Memory Solutions for AI Computing(…)
22mCXL-ANNS: Software-Hardware Collaborative Memory Disaggregation and Computation...
34mLeveraging Apache Arrow for ML Workflows
28mPagedAttention: Revolutionizing LLM Inference with Efficient Memory Management
01mAdaSpring: Context-adaptive and Runtime-evolutionary Deep Model Compression for Mobile Applications
06mTensorPRAM: Designing a Scalable Heterogeneous Deep Learning Accelerator
14mSYMPHONY: Enabling Compute-Memory Disaggregation in LLM Serving Systems