Coursera Spring Sale
40% Off Coursera Plus Annual!
Grab it
Explore a technical conference presentation that delves into optimizing Non-Volatile Memory Express over Fabric (NVMe-oF) TCP for disaggregated storage on Data Processing Units (DPUs). Learn how DPUs provide local NVMe controller offload to remote storage, enabling VMs and containers to access NVMe devices as local PCIe controllers while connecting to namespaces over fabric targets. Discover the standardization of this approach in AI/ML storage workloads and its adoption by newer transport protocols like Falcon/UEC. Examine specific challenges of using DPUs for disaggregated storage with NVMe over TCP, including latency and bandwidth issues caused by DPU memory bounce buffers. Master advanced optimization techniques using recent Linux networking stack developments such as dmabuf and TCPdevicemem to overcome performance limitations. Gain insights into how DPU-based disaggregated storage is evolving through industry-standard transport protocols and understand the role of DPUs in providing streamlined infrastructure control for hyperscalers and datacenter operators.
