Designing a Flash-Aware Two-Level Cache

The random read efficiency of flash memory, combined with its growing density and dropping price, make it well-suited for use as a read cache. We explore how a system can use flash memory as a cache layer between the main memory buffer pool and the magnetic disk. We study the problem of deciding which data pages to cache on flash and propose alternatives that serve different purposes. We give an analytical model to decide the optimal caching scheme for any workload, taking into account the physical properties of the flash disk used. We discuss implementation issues such as the effect of the flash cache block size on performance. Our experimental evaluation shows that questions on systems with flash-resident caches cannot be given universal answers that hold across all flash disks and workloads. Rather, our cost model should be applied per case to provide an optimal setup with confidence.