Beyond Machine Boundaries: Experience with Memory Disaggregation
summary
Prototype uses ThymesisFlow on two IC922 Power9 servers with AlphaData ADM9V3 FPGAs connected back-to-back, exposing remote memory via either a kernel mappable device or firmware hot‑plug NUMA node. Measured pointer‑chase latency around 850 ns (stable up to ~10 GB); TPC‑H on Hyrise shows wide variance—some queries ≈4.5× slower (Q9) while others are barely affected. Implementations include modified firmware, kernel modules, Pegasus heap partitioning, and a VM live‑migration demo that resumes execution on a borrower without copying all memory.
AI-written summary · about
chapters (10)
- 0:00Introduction and motivation
- 4:54Coherent interconnects for disaggregation
- 8:44Terms of flow prototype setup
- 11:54Latency measurements and results
- 19:27Current prototype status and challenges
- 25:27Open challenges: performance and reliability
- 31:09Architecture implications for software
- 37:32Pegasus memory management system
- 38:13VM migration with terms of flow
- 41:30Workshop advertisement and Q&A
related talks
16mA Tale of Two Paths: Toward a Hybrid Data Plane for Efficient Far-Memory Applications
16mDisaggregation and the Application
20mUniMem: Redesigning Disaggregated Memory within A Unified Local-Remote Memory...
23mBreaking the limits of x86
16mOneSidedMW: Managing Disaggregated Memory Efficiently, Flexibly, and Securely with RNIC
12mDirect Access, High-Performance Memory Disaggregation with DirectCXL
31mUnshackling Memory!
23mTeRM: Extending RDMA-Attached Memory with SSD