Optimizing Checkpoints with Built-in Memory Page Compression
summary
Integrates LZ4 in-flight compression into CRIU so pages are compressed before being written, eliminating the extra write-read-compress cycle and cutting I/O and CPU waste. A memhog demo compresses ~12 GB of RAM down to 12 GB → ~80 MB; per-page compressed sizes are stored enabling random access and sparse-file tricks via a per-page compressed-size index. Preliminary evaluation shows linear scaling and much lower overhead compared to the naive post-checkpoint compression fails with OOS approach; best for highly redundant memory, less effective for typical LLM inference.
AI-written summary · about
chapters (14)
- 0:00Introduction to checkpoint optimization
- 1:27Compression algorithms comparison
- 3:39Naive checkpoint compression issues
- 5:48Built-in memory page compression design
- 6:30Live demo of compression in action
- 8:21Performance evaluation results
- 9:44Summary and integration benefits
- 10:05Q&A on language model compression
- 13:15LZ4 fast mode and optimizations
- 15:07Dictionary usage in LZ4
- 16:29Iterative checkpointing limitations
- 16:50Multi-threaded restore discussion
- 17:51Live migration and async restore
- 19:16Use cases for compression
related talks
42mAllocators for Compressed Pages: In-depth Comparison with z3fold in Focus
40mInspecting and Optimizing Memory Usage in Linux
30mMemory tiering
22mEMD: Fair and Efficient Dynamic Memory De-bloating of Transparent Huge Pages
26mLet's demystify memory
41mLies Programmers Believe about Memory
44mLarge Pages in Linux
27mLive Migration With Hardware Acceleration