At the 2024 ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), Nils Asmussen and Michael Roitzsch presented their paper, “Towards Disaggregation-Native Data Streaming between Devices,” during the 3rd Workshop on Heterogeneous Composable and Disaggregated Systems (HCDS).
Their research explores the emerging trend of disaggregation in datacenters, a method aimed at enhancing flexibility. Disaggregation involves using technologies like CXL to connect pools of CPUs, accelerators, and memory through a datacenter fabric. This setup allows applications to select the specific resources they need from these pools, optimizing performance and efficiency.
However, a challenge arises with data movement. Typically, data needs to be streamed through multiple devices, but instead of flowing directly from one device to another, it often gets staged in memory by a CPU. This staging can create delays and inefficiencies.
The researchers propose a solution: a disaggregation-native and device-independent data streaming facility. This innovation enables data to flow directly between devices without the need for intermediary staging in memory. The result is improved processing speeds and reduced latencies, making datacenters more efficient.
Their full paper can be accessed here.
Check out all list of COREnext scientific publications here.