Optimizing NoSQL DB on Flash: A Case Study of RocksDB

A solid-state drive (SSD) gains fast I/O speed and is becoming an ideal replacement for traditional rotating storage. However, its speed and responsiveness heavily depend on internal fragmentation. With a high degree of fragmentation, an SSD may experience sharp performance degradation. Hence, minimizing fragmentation in the SSD is an effective way to sustain its high performance. In this paper, we propose an innovative file data placement strategy for Rocks DB, a widely used embedded NoSQL database. The proposed strategy steers data to a write unit exposed by an SSD according to predicted data lifetime. By placing data with similar lifetime in the same write unit, fragmentation in the SSD is controlled at the time of data write. We evaluate our proposed strategy using the Yahoo! Cloud Serving Benchmark. Our experimental results demonstrate that the proposed strategy improves the Rocks DB performance significantly: the throughput can be increased by up to 41%, 99.99%ile latency reduced by 59%, and SSD lifetime extended by up to 18%.