GAIA: An OS Page Cache for Heterogeneous Systems
summary
GAIA integrates GPU memory into the OS page cache by implementing a weakly-consistent page cache and a lazy release consistency protocol inside Linux, prototyped on NVIDIA Pascal GPUs. It lets applications mmap into the GPU address space, keeps CPU processes and unmodified GPU kernels compatible, and removes false sharing that causes page ping‑pong. Evaluation uses Gunrock, image collage and microscopy stitching and shows up to 3× CPU file I/O speedup and up to 8× end-to-end gains.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:45Core services that require OS management of GPU memory
- 1:15Example: Road navigation service
- 4:28Consistency model considerations
- 5:04A bit of background: NVIDIA unified memory
- 5:26NVIDIA unified memory - Page level strict coherence
- 6:44False sharing due to strict coherence
- 7:43False sharing impact on the system
- 8:49Lazy Release Consistency (LRC)
- 11:44Evaluation
- 12:00A Hybrid CPU-GPU System for Stitching Large Scale Optical Microscopy Images
- 14:03Dynamic graph processing with Gunrock
- 14:35GAIA - Conclusion
related talks
33mGPU Para-Virtualization on Xen
19mP2CACHE: Exploring Tiered Memory for In-Kernel File Systems Caching
44mLarge Pages in Linux
27mLive Migration With Hardware Acceleration
10mKeynote: An Optimized Linux Stack for GenAI Workloads
29mXenGT: A software based Intel Graphics Virtualization Solution
47mEverything Awesome about GPU Drivers
30mdrm and cgroup memory management - two great tastes