UniMem: Redesigning Disaggregated Memory within A Unified Local-Remote Memory...
summary
UniMem redesigns cache-coherent disaggregated memory by removing the accelerator address-translation hop and treating device-attached and host RAM as a unified local-remote hierarchy, using Shadow region and remote B to map fake physical space directly. It implements a small device-side filter cache with active/inactive lists to prevent thrash and pollution, plus a page promotion based on hotness fragmentation that promotes only hot subpages. Evaluation (simulation vs Kona/S1) reports up to 76.4% lower average memory access time and much lower data amplification.
AI-written summary · about
chapters (13)
- 0:00UniMem intro and disaggregated memory
- 1:27Disaggregated memory designs
- 2:52Cache-coherent system limits
- 6:27Cache pollution and thrashing
- 7:52Need for unified local-remote memory
- 8:15UniMem’s memory hierarchy design
- 10:44Shadow region and remote B isolation
- 11:06Filter cache to reduce pollution
- 13:14Fast host memory with UniMem
- 15:04Evaluation methodology and systems
- 16:51Performance results and amplification
- 18:17Concurrent workload benefits
- 19:01Conclusion and future
related talks
45mBeyond Machine Boundaries: Experience with Memory Disaggregation
16mOneSidedMW: Managing Disaggregated Memory Efficiently, Flexibly, and Securely with RNIC
18mHermit: Low-Latency, High-Throughput, and Transparent Remote Memory...
46mA Complete Re-Think for Memory Config...
16mA Tale of Two Paths: Toward a Hybrid Data Plane for Efficient Far-Memory Applications
19mFineMem: Breaking the Allocation Overhead vs. Memory Waste Dilemma in Fine-Grained...
37mMemSweeper: Virtualizing Cluster Memory Management for High Memory Utilization
31mUnshackling Memory!