Disaggregating Persistent Memory and Controlling Them Remotely
summary
Explores a passive disaggregated persistent memory model (pDPM), implements three remote key-value systems—pDPM-direct, pDPM-central, and Clover—and measures them over 100Gbps RDMA. Clover separates a distributed data plane from a centralized metadata plane, uses one-sided RDMA for data and out-of-place writes with multiversion chains and shortcut reads, achieves reads at ~1 RTT and writes ~2 RTTs (close to raw verbs), matches active disaggregation performance in many workloads, reduces CapEx by 1.4x and reduces OpEx by 3.9x, but suffers under heavy write contention.
AI-written summary · about
chapters (11)
- 0:00Start
- 0:26Resource Disaggregation
- 0:44Disaggregated Storage
- 1:08Disaggregated Memory
- 1:25Disaggregated Persistent Memory?
- 12:15Clover: Combining Distributed and Centralized Approaches
- 19:19Reliability and Load Balancing
- 20:17Evaluation Setup
- 21:04Microbenchmark - Latency
- 23:25Conclusion
- 24:11Final Thoughts
related talks
28mPersistent Memory plus RDMA, new age remote device
17mPreview: Persistent Memory
15mReplicating Persistent Memory Key-Value Stores with Efficient RDMA Abstraction
16mFORD: Fast One-sided RDMA-based Distributed Transactions for Disaggregated Persistent...
14mODRP: On-Demand Remote Paging with Programmable RDMA
35mDeveloping Software for Persistent Memory
35mDeveloping Software for Persistent Memory
15mMore Than Capacity: Performance-oriented Evolution of Pangu in Alibaba