DOGI: Data Placement with Oracle-Guided Insights for Log-Structured Systems

Log-structured systems have become the backbone of modern data-intensive applications thanks to their high write throughput. Their efficiency, however, is deteriorated by the write amplification factor (WAF) induced by garbage collection. Despite extensive studies, there still exists a wide gap between practice and optimality. In this paper, we bridge this gap with two key contributions. We first design NoDaP, a near-optimal oracle baseline that sets the upper bound for WAF reduction. Then, guided by insights from NoDaP, we propose DOGI, an oracle-inspired data placement technique that combines simple yet effective heuristics with lightweight machine learning. DOGI predicts invalidation times for data blocks with high accuracy, dynamically tunes group configurations, and finds the sweet spot between fine-grained data placement and misprediction penalty. Our experiments, using simulations and a prototype on a zoned device, show that DOGI reduces WAF by up to 23.2% while improving write throughput by up to 13.3% over the best-performing baseline.