TeRM: Extending RDMA-Attached Memory with SSD
summary
Extends RDMA-attached memory by mapping SSD-backed virtual space (64 GB VM with 32 GB RAM in tests) and replacing NIC page-fault handling with user-space checks that return a magic physical page pattern so missing data triggers a software RPC to fetch SSD pages; uses a hybrid T IO (buffered if cached, direct IO otherwise), hotspot promotion, and an LD_PRELOAD shared library to avoid app changes. Implements and measures a cluster where TeRM runs close to the pinned upper bound, outperforms ODP by up to 600× and RPC by ~7–12×, with <7% dip and 1 s recovery after hotspot shifts.
AI-written summary · about
chapters (8)
- 0:00Introduction to RDMA
- 2:36ODP-MR mechanism
- 6:12Performance analysis
- 8:49The TERM design
- 11:21TERM operation flows
- 14:01Optimizing I/O and layout
- 16:49Experimental results
- 22:01Conclusion
related talks
12mCharacterizing and Optimizing Remote Persistent Memory with RDMA and NVM
14mUnderstanding RDMA Microarchitecture Resources for Performance Isolation
18mRDMA is Turing complete, we just did not know it yet!
19mFast Distributed Transactions for RDMA-based Disaggregated Memory
45mBeyond Machine Boundaries: Experience with Memory Disaggregation
25mStretch: A Fault-Driven DSM Runtime for Distributed Multithreaded Applications
16mOneSidedMW: Managing Disaggregated Memory Efficiently, Flexibly, and Securely with RNIC
12mDirect Access, High-Performance Memory Disaggregation with DirectCXL