Memory Database A Number Of Key Technologies,

As semiconductor memory becomes cheaper and chip densities increase, memory resident database system—MMDB comes true. MMDB provides much better response times and much more transaction throughputs than disk resident database (DRDB), which make real-time applications with restrict time response and high performance transaction processing systems perfect candidates for MMDB. This paper deeply analyzes the difference between MMDB and DRDB, and gives prominence to the characteristic of MMDB. It reconsiders how MMDB is implemented from points of data structures and algorithms, and puts its emphases on the techniques of three aspects: optimization of cache utilization, recovery management and storage model.In the first chapter of this paper, we introduce MMDB concepts and differences between MMDB and DRDB, and put our emphases on the development status of MMDB according to its speciality. In the 2~nd Chapter, we study the structure of index of MMDB, and enumerate several tree like indices, which are used in traditional MMDB systems. We demonstrate that in main-memory databases, the number of processor cache misses has a critical impact on the performance of the system. Based on the study of those indices' cache performance, we proposed a cache sensitive index called "Improved Cache Sensitive B+-Tree"(MCSB+-Tree). It gives attention to the effects of cache misses and TLB misses, and can gain better performance than conventional cache conscious indices. The 3~rd Chapter gives research on recovery technology of MMDB. The performance of recovery manager of MMDB is key to whole system. Based on the former techniques of recovery, we propose the "Fast Recovery Model", which has special architecture. With support of log driven checkpointing schema and reloading algorithm, "Fast Recovery Model" not only improves the system's stabilization, but also provides the method of fast restart when MMDB fails. In the 4th Chapter, we study the storage technology of MMDB. We enumerate several storage models which are used in MMDB, and describe the disadvantage of those models when they utilize the processor cache to improve performance. Then, we propose a data organization model called "Partition Attributes Across Storage Model", that improves cache performance by grouping together all values of each attribute within each page. The last chapter is overview of this paper and show future research direction of MMDB.