FileMR: Rethinking RDMA Networking for Scalable Persistent Memory
summary
Combines a new RDMA abstraction, FileMR, with range-based translation and an append verb so RDMA addresses files by offset instead of virtual pages; implements the prototype on SoftRoCE and a file-system binding that propagates extent updates to the NIC. Claims a 3.8x–340x increase in NIC translation cache hit rate and 1.8x–2.0x application speedup by eliminating expensive 1:1 page registrations (mapping 1 TB previously took ~25 minutes), though the SoftRoCE prototype has higher latency than real NICs.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:32Persistent memory and RDMA
- 2:43PM: memory management
- 4:49RDMA on PM
- 7:35Alternative approaches to VM Issues on RDMA
- 8:49Conflicting roles of metadata management
- 9:32FileMR: File-based memory region
- 11:35Range-based memory translation table
- 13:35FileMR: Implementation
- 14:49FileMR: Evaluation
- 16:22Conclusion
related talks
28mPersistent Memory plus RDMA, new age remote device
12mPreview: Networking
18mRDMA is Turing complete, we just did not know it yet!
14mODRP: On-Demand Remote Paging with Programmable RDMA
17mSRNIC: A Scalable Architecture for RDMA NICs
19mFineMem: Breaking the Allocation Overhead vs. Memory Waste Dilemma in Fine-Grained...
25mRDMA @Scale
20mEmpowering Azure Storage with RDMA