Clean first or dirty first?

Abstract

Flash SSDs originate a disruptive change concerning storage technology and become a competitor for conventional magnetic disks in the area of persistent database stores. Compared to them, they provide a dramatic speed-up for random reads, but exhibit a distinct read/write (R/W) asymmetry, i.e., updates are more expensive than reads. Existing buffer management algorithms for those devices usually trade physical reads for physical writes to some extent. But they ignore the actual R/W cost ratio of the underlying device and the update intensity of the workload. Therefore, their performance advantage is sensitive to device and workload changes. We propose CASA (Cost-Aware Self-Adaptive), a novel buffer replacement policy, which makes the trade-off between physical reads and physical writes in a controlled fashion, depending on the R/W cost ratio, and automatically adapts itself to changing update intensities in workloads. Our experiments show that CASA outperforms previous proposals in a variety of cost settings and under changing workloads.