FineMem: Breaking the Allocation Overhead vs. Memory Waste Dilemma in Fine-Grained...
summary
FineMem implements an RDMA-based remote allocator that removes MR registration from the allocation path via per-compute-node pre-registered MRs and isolates access with RDMA memory windows. It uses a lock-free, one-sided RDMA protocol and a two-layer bitmap (sections/spans) to allocate 4KB–2MB chunks with at most two round trips for up to 500 trunks. Evaluated on 16 compute and up to 6 memory nodes, it claims up to 95% reduction in allocation latency, ~40% lower memory cost, and up to 100% KV bandwidth improvement.
AI-written summary · about
chapters (7)
- 0:00FineMem introduction
- 1:44Remote memory allocation challenges
- 5:17FineMem design principles
- 9:16Efficient remote chunk allocation
- 12:08Crash consistency with logging
- 14:35Experimental evaluation
- 17:54Summary
related talks
18mFileMR: Rethinking RDMA Networking for Scalable Persistent Memory
28mPersistent Memory plus RDMA, new age remote device
16mOneSidedMW: Managing Disaggregated Memory Efficiently, Flexibly, and Securely with RNIC
20mUniMem: Redesigning Disaggregated Memory within A Unified Local-Remote Memory...
23mTeRM: Extending RDMA-Attached Memory with SSD
25mDisaggregating Persistent Memory and Controlling Them Remotely
14mODRP: On-Demand Remote Paging with Programmable RDMA
12mCharacterizing and Optimizing Remote Persistent Memory with RDMA and NVM