Reproduce “Cohet: A CXL-Driven Coherent Heterogeneous Computing Framework with Hardware-Calibrated Full-System Simulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yanjing Wang, Lizhou Wu, Sunfeng Gao, Yibo Tang, Junhui Luo, Zicong Wang, Yang Ou, Dezun Dong and 2 more
Cohet, a CXL-driven coherent heterogeneous computing framework, and SimCXL, a hardware-calibrated full-system simulator with 3% average error, were developed to provide infrastructure for CXL-enabled heterogeneous computing. This work enables simplified programming and demonstrates speedups of up to 40.2x for Remote Atomic Operations and an average of 1.86x for Remote Procedure Calls by leveraging CXL's coherent memory for efficient CPU-XPU collaboration.
Key insights
- CXL.cache provides a substantial performance advantage, showing a 68% reduction in latency and a 14.4x increase in bandwidth over DMA for fine-grained 64B memory accesses.
- Cohet's CXL-driven approach simplifies heterogeneous programming by exposing standard memory allocation interfaces and enabling transparent OS-managed coherence and memory placement.
- Coherent CXL-NICs drastically accelerate distributed computing primitives, demonstrating efficiency for latency-sensitive workloads that struggle with traditional PCIe-based solutions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.