Exploring CXL Memory Disaggregation Use Cases and System Benefits

Jungmin Choi,Myoungseo (Matthew) Kim, SK hynix - In this session, we dive into powerful use cases for DCD (Dynamic Capacity Device) based memory pooling and sharing specified in CXL 3.1 that can reduce costs from a datacenter operations perspective. Each host connected to the memory pool can manage resources flexibly through dynamic memory allocation. We also address use cases for system benefits, such as optimal memory tiering and data placement that each host can achieve through telemetry including hotness tracking provided by the device. Regarding memory sharing, we analyze the characteristics of HPC-AI workloads and explore performance improvement through control and data path optimization between hosts. Specifically, we focus on distributed computing frameworks such as Spark and Ray, which suffer from heavy network traffic and memory duplication due to data object sharing. Through HW-SW system implementation and concrete performance studies, we provide insight into CXL memory pooling and sharing within a multi-host environment. Accelerating Analytics Processing and Ray AI Framework 12:59 Identifying hot and cold pages and automatic migration 14:55