DOGI: Data Placement with Oracle-Guided Insights for Log-Structured Systems
summary
Presents DOGI, a data-placement system that combines a hybrid heuristic+ML predictor (hot and frozen filters plus an MLP with six features and 10 outputs) with dynamic group tuning via a Markov-chain WAF estimator and a prediction-log-driven GC relocation policy. Implements a prototype on a ZNS SSD and validates in simulation; compares against a NoDaP near-optimal oracle and four prior schemes. Achieves up to 23.2% WAF reduction and up to 13.3% higher write throughput (15% average WAF win vs best baseline), while keeping prediction latency near heuristic levels.
AI-written summary · about
chapters (9)
- 0:00DOGI introduction
- 1:20Log-structured systems & GC
- 2:43Invalidation time prediction
- 4:09Existing placement limits
- 8:23DOGI design goals
- 9:26DOGI hybrid prediction
- 12:16DOGI GC relocation
- 13:39DOGI on ZNS SSD
- 15:27Conclusion
related talks
17mDOTE: Rethinking (Predictive) WAN Traffic Engineering
18mScaling RGW D4N Caches With Experimental Offloading
15mDistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language...
18mDayu: Fast and Low-interference Data Recovery in Very-large Storage Systems
12mCharacterizing and Optimizing Remote Persistent Memory with RDMA and NVM
40mDistributed, Modular Grid Software for Management and ...
01h 07mIndustry Presentations
22mPOSH: A Data-Aware Shell