Ceph RBD Integration with DPUs
summary
Prototype runs a mini librbd on DPU that exposes NVMe namespaces to hosts and converts IO to RADOS calls, using AMD/Pensando ARM cores plus MPUs for encryption/compression and network offload. Configuration and metrics travel over gRPC-configurable agents as JSON. With a fast SPDK backend the DPU path hits about 2 million IOPS on 4K vs Linux kernel ~800–850k; performance falls with slow backends. The CRUSH/librbd mapping stays on ARM and large-cluster scaling remains unresolved.
AI-written summary · about
chapters (12)
- 0:00Introduction to DPUs
- 2:27Why DPUs in data centers
- 6:02NVMe over Fabrics overview
- 9:11Current Ceph RBD deployments
- 12:01DPU-based RBD integration
- 15:53Optimizing RBD on DPUs
- 18:09Benefits of DPU offloading
- 20:19Observability and metrics
- 21:01Technical implementation details
- 24:15Challenges with libRBD
- 26:45Platform openness and flexibility
- 27:33Q&A and summary
related talks
14mODRP: On-Demand Remote Paging with Programmable RDMA
30mZXDH DPU Adapter and It's Application
39mFunction Dedicated Network Deployment Model in DPU/IPU-Like Device Acro...
06mMaximizing Programmable DPUs in the Network
25mDisaggregating Persistent Memory and Controlling Them Remotely
28mPersistent Memory plus RDMA, new age remote device
22mKernel-Native Packet Processing on AMD GPUs: XDP Offload without Userspace Runtimes
20mPeRF: Preemption-enabled RDMA Framework