OneSidedMW: Managing Disaggregated Memory Efficiently, Flexibly, and Securely with RNIC
summary
System implements RNIC offloading plus memory windows to perform remote bind/unbind on the NIC so allocation and deallocation happen with zero memory-node CPU involvement; prototype runs on ConnectX-5 100Gbps RNICs with one memory-core. It uses type-2 memory windows to prevent R_Key enumeration, adds QP and memory-window grouping and management/access QP separation to avoid contention, and evaluates on a KV store and swap system, reporting up to 10.6× throughput in KV and up to 32.3% swap improvement while saving 43% memory at 8KB granularity.
AI-written summary · about
chapters (15)
- 0:00Disaggregated memory intro
- 1:21Challenges in memory management
- 2:04Limitations of allocation
- 4:30OneSidedMW design insight
- 5:11RDMA offloading with NIC
- 5:52OneSidedMW architecture
- 7:35Security risks of type one windows
- 8:38Type two windows for isolation
- 9:20Performance with type two windows
- 10:25QP and window grouping
- 11:27Separating management access QP
- 12:10Evaluation setup
- 13:15Redis performance results
- 13:58FastSwap cloud tests
- 15:00Conclusion
related talks
14mODRP: On-Demand Remote Paging with Programmable RDMA
17mSRNIC: A Scalable Architecture for RDMA NICs
17mDRack: A CXL-Disaggregated Rack Architecture to Boost Inter-Rack Communication
18mSTYX: Exploiting SmartNIC Capability to Reduce Datacenter Memory Tax
14mUnderstanding RDMA Microarchitecture Resources for Performance Isolation
18mFileMR: Rethinking RDMA Networking for Scalable Persistent Memory
45mBeyond Machine Boundaries: Experience with Memory Disaggregation
20mMemory Recycling for Network Interfaces