Octopus Design for CXL Memory Pods

This technical report presents the Octopus design, a novel approach to CXL (Compute Express Link) memory pods that enhances memory pooling and communication efficiency across servers. The report outlines the limitations of existing CXL pod designs, which typically require full connectivity between servers and memory devices, leading to inefficiencies. Octopus introduces a sparse topology that connects servers to a limited number of memory devices, allowing for scalability without the need for costly switches. The design balances the need for low-latency communication and effective memory pooling by organizing servers into islands that optimize interconnectivity. The report details the performance evaluation of Octopus through a prototype and simulations, demonstrating significant improvements in communication latency and cost savings compared to traditional CXL switch designs. The findings indicate that Octopus can achieve both high performance and cost efficiency, making it a viable solution for addressing the growing memory demands in cloud computing environments.