Hermit: Low-Latency, High-Throughput, and Transparent Remote Memory...
summary
Hermit is a kernel swap redesign that implements feedback-directed asynchrony to move expensive, non-urgent work off the page-fault path, adaptively schedule background reclaim threads, and batch reclaim operations to save CPU. It ships as a Linux 5.14 implementation and evaluates six real cloud apps. Results show it is transparent to user code, reduces the 99th percentile latency of Memcached by 99.7% (36 ms → 91 µs) and raises batch throughput ~1.24× on average.
AI-written summary · about
chapters (15)
- 0:00Start
- 1:31Remote Memory Systems
- 2:08Kernel Swap Can Incur High Tail Latency
- 2:50Where Does the Latency Come From?
- 4:55Kernel Bypassing Is Not a Panacea
- 6:05Naive Asynchrony Is Not Enough
- 7:23Challenge #1: When To Start Reclamation
- 8:32Hermit Design: Feedback-Directed Asynchrony
- 8:49Hermit's Adaptive Reclaim Scheduling
- 11:54Hermit Achieves Low Latency
- 12:25How To Improve Throughput?
- 13:25Aggressive Batching For Async. Reclamation
- 14:09Evaluation
- 14:34Low Tail Latency
- 16:43Conclusion
related talks
20mUniMem: Redesigning Disaggregated Memory within A Unified Local-Remote Memory...
45mBeyond Machine Boundaries: Experience with Memory Disaggregation
15mHydra : Resilient and Highly Available Remote Memory
16mA Tale of Two Paths: Toward a Hybrid Data Plane for Efficient Far-Memory Applications
23mTeRM: Extending RDMA-Attached Memory with SSD
31mUnshackling Memory!
37mMemSweeper: Virtualizing Cluster Memory Management for High Memory Utilization
16mDisaggregation and the Application