H-Store, one of the main memory database management system, uses anti-caching to evict data from primary storage (main memory) to secondary storage (disk) where the size of the data exceeds memory threshold. Anti-caching uses hot-cold separate policy to manage data, and moves cold data to disk in order to reduce number of long-latency transactions caused by accessing to secondary storage. Current methods (e.g. LRU, Timestamp) use ;;recency;; (how recently data is accessed) information to separate data. These methods do not keep past information.In this paper, I present low overhead, frequency-based anti-caching method which keeps the past information. This method adds frequency value to each tuple, and increases the value when corresponding tuple is accessed. Performance was evaluated using the TPC-C benchmark. Through accurate hot-cold separation, overall transaction throughput is increased by reducing secondary storage access.
PDF
download
878987797
028-85220240
