Technical Papers Archive

OCEAN: Open-Source CXL Emulation for Hyperscale Architecture and Networking


Authors: Yiwei Yang (University of California, Santa Cruz); Mujahid Al Rafi (University of California, Merced); Jesun Sahariar Firoz and Zhen Peng (Pacific Northwest National Laboratory (PNNL)); Xi Wang (University of California, Merced); Kenza Bouasker (University of Washington, Seattle); Sayan Ghosh (Pacific Northwest National Laboratory (PNNL)); Jie Li (Texas Tech University); Daniel Wong (University of California, Riverside); Dong Li and Hyeran Jeon (University of California, Merced); and Kevin Barker, Nathan Tallent, and Luanzheng Guo (Pacific Northwest National Laboratory (PNNL))

Abstract: Compute Express Link (CXL) 3.0 introduces fabric-level memory pooling, yet the absence of production CXL 3.0 hardware and the complexity of multi-host configurations pose significant barriers to the research community. This paper presents OCEAN, a high-fidelity, open-source emulation framework that reproduces the fabric-level feature set that distinguishes CXL 3.0—including multi-host memory sharing, dynamic capacity management, and coherent pooling across hosts—using QEMU virtualization and custom Linux kernel modules. OCEAN makes two central technical contributions: active latency injection and a software cache coherence layer implementing MESI-state management with distributed snoop filters to maintain correctness across up to 16 hosts. We validate OCEAN against a real CXL Type-3 device. In addition, case studies on Tigon database workloads quantify the throughput impact of hardware cache coherence—an 11× improvement at full coverage—while a design-space sweep of 12 memory-placement policies on GROMACS demonstrates up to 7% performance variation from policy choice alone.


Presentation: file

Back to Technical Papers Archive Listing