ODRP: On-Demand Remote Paging with Programmable RDMA
summary
ODRP offloads fine‑grain remote paging onto the RNIC by chaining RDMA work requests and using event‑triggered worker queues to implement allocation, translation and access entirely on the NIC. Prototype on ConnectX‑5 with a single‑core memory node and 8 compute nodes shows zero memory‑node CPU usage, up to 12× better memory utilization, and only 0.8%–14.6% runtime overhead versus pre‑registered one‑sided RDMA. Key techniques: a client‑assisted principle to shorten work‑request chains and RNIC tricks (masking for ring buffers, scatter/gather endian conversion) to make modular queues and pointer arithmetic work.
AI-written summary · about
chapters (10)
- 0:00ODRP introduction
- 2:44Disaggregated memory limits
- 3:48ODRP design goals
- 5:54ODRP architecture
- 6:58Swap request workflow
- 8:03Efficiency challenges
- 9:28Functionality challenges
- 11:17Evaluation setup
- 12:20Performance results
- 13:02Conclusion
related talks
16mOneSidedMW: Managing Disaggregated Memory Efficiently, Flexibly, and Securely with RNIC
18mRDMA is Turing complete, we just did not know it yet!
17mSRNIC: A Scalable Architecture for RDMA NICs
12mPreview: Networking
20mRDMA at Cloud Scale: The OCI Experience
28mPersistent Memory plus RDMA, new age remote device
20mPeRF: Preemption-enabled RDMA Framework
06mRDMA Programming Design and Case Studies - For Better Performance Distributed