Data Placement In Hybrid Storage Systems

The increase of CPU speed has far exceeded the increase of I/O speed ofthe magnetic disk (Hard Disk Drive, namely HDD), and the gap is widening.Storage systems become the bottleneck of database systems.The emergence of flash disks (Flash-based Solid State Drive, i.e., FlashSSD, commonly referred to as SSD) has made a revolutionary change topermanent storage systems, bringing the hope to enhance the performance of thedatabase systems. Since a SSD is fully built on the semiconductor chips withoutmechanical operations, the fundamental different feature makes it the idealstorage medium, thereby improving the performance of the storage system.Although flash disks have an incomparable superiority to magnetic disks,for the following reasons, in the foreseeable future, it is impossible tocompletely replace magnetic disks with flash disks:(1) the manufacture costs offlash disks are still much higher than that of magnetic disks.(2) data structuresand algorithms in database systems are based on a magnetic disks. Databasesystems need adjust them to make full use of the advantages of flash disks.(3)The superior features of magnetic disks, including low cost, large capacity and fast sequential access speed, make magnetic disks irreplaceable in the storagehierarchy. Therefore, for most storage systems, flash disks are not considered thereplacement of magnetic disks, rather, a way to improve the performance of thestorage systems. It is necessary to find a suitable place for flash disks, so that theright balance is formed between performance and cost. At the present stage,database researchers pay their attention to the hybrid storage architectures usingboth magnetic disks and flash disks, which take full advantage of respectivestrengths of magnetic disks and flash disks, reduce purchasing costs andoperating costs, and thereby improve the overall performance of the storagesystems.In a hybrid storage system, the four access operations, namely sequentialread, sequential write, random read and random write, have different behaviorson the two storage medias, and hence different contributions to the databasesystem performance. It is essential to consider the differences when making dataplacement plans. The existing methods take into account the differences betweenstorage media, but ignore the differences among four operations. Based onexisting methods, this paper proposes a weighted placement method for databaseobjects in hybrid storage systems at the granularity level of tables and indexes.The method makes a distinction between the different operations on differentstorage media, according to the different loads, assigns four operations differentweights on the database system performance contributions, to make full use ofperformance advantages of storage devices. Then when the storage decision is made on tables and indexes，the global optimal placement of database objects isrealized by taking advantage of backtracking algorithm, thereby to maximize theperformance increase of database systems. Finally, experiments are tested whereTPC-C is acted as the load of the database and the correctness of method isverified. The experimental results show that the method, compared with theexisting method, can better enhance the performance of database systems andprolong the lifetime of the flash disks.