Authors: Peiran Du, Weishu Deng, Yang Song, Chenfei Liao, and Yujie Yang (The University of Texas at Arlington); Ren Wang (Intel Corporation); and Hui Lu and Jia Rao (The University of Texas at Arlington)
Abstract: Current multi-host CXL platforms provide shared, byte-addressable memory, but no cross-host cache coherence or atomic operations. As a result, the main challenge is not remote-memory latency itself, but the software overhead required to ensure visibility and synchronization.
We quantify this overhead on a real four-host CXL shared-memory system, covering raw memory performance, software-managed publish/observe visibility, and synchronization built from explicit flushes and polling. On our platform, one-way cross-host visibility is about 1 µs, and each lock acquisition costs 3–5 µs, roughly an order of magnitude higher than local hardware spin locks. These overheads dominate synchronization-heavy workloads: BFS gains nothing from adding a second host; it requires frequent barriers and cross-host exchange. In contrast, triangle counting, which is read-dominated with little synchronization, does speed up when scaling from 32 to 32+32 threads across two hosts. Overall, suitability depends on remote-write intensity, synchronization frequency, and whether hot metadata stays in local DRAM.
Presentation: file