【摘要】：隨著大規(guī)模互聯(lián)網(wǎng)應用的高速發(fā)展，給存儲系統(tǒng)的可擴展性提出了更高的要求，鍵值數(shù)據(jù)庫由于其簡單高效的數(shù)據(jù)模型在可擴展性方面較傳統(tǒng)的數(shù)據(jù)庫系統(tǒng)有很大的優(yōu)勢。另一方面，隨著硬件的發(fā)展，固態(tài)盤的性價比逐步提高，已經(jīng)成為越來越多系統(tǒng)的首選，其隨機讀的優(yōu)勢與互聯(lián)網(wǎng)應用的特征高度符合，但是其寫性能的相對低下以及擦除次數(shù)的限制影響了它的應用場景，于是將固態(tài)盤與磁盤相結(jié)合的融合存儲系統(tǒng)得到了大家的廣泛關(guān)注。因此針對鍵值數(shù)據(jù)庫而設計的融合存儲系統(tǒng)是一個有價值的結(jié)合點。 針對Web應用的負載特點，通過日志的方式順序的記錄對鍵值數(shù)據(jù)庫的操作，所有的操作都僅僅是順序的寫入內(nèi)存，當內(nèi)存中的數(shù)據(jù)達到了一定閾值以后一次刷寫到后端存儲上，日志的方式一定程度上是通過犧牲讀性能來優(yōu)化寫性能，所以和固態(tài)盤有天然的互補關(guān)系。由于后端固態(tài)盤與磁盤的不同特性，使用層次化的方法將寫的特點進行分類后寫入不同的設備上，并設計實現(xiàn)了一個可定制文件放置與遷移策略的文件系統(tǒng)HybridFS，，通過監(jiān)控分析腳本可以對同一文件系統(tǒng)中不同特點的文件選擇不同的放置與遷移策略。針對鍵值數(shù)據(jù)庫的文件訪問特點，日志類文件多是一次寫很少讀，所以直接寫入磁盤。元數(shù)據(jù)類文件讀寫都很頻繁但是文件大小和數(shù)量偏少，因此寫入固態(tài)盤。對數(shù)據(jù)文件由于其海量、定長、一次寫多次讀的特點，根據(jù)負載特點選擇性的寫入磁盤或者固態(tài)盤。針對純寫入型負載使用概率選擇的方式比Flashcache提升了5%-56%，而對于讀寫混合行負載使用LRU的遷移方式相比Flashcache有4%-14%的性能提升。
[Abstract]:With the rapid development of large-scale Internet applications, the scalability of storage system is higher. Because of its simple and efficient data model, key-valued database has great advantages over traditional database system in scalability. On the other hand, with the development of hardware, the performance-price ratio of solid disk has been gradually improved, which has become the first choice of more and more systems, and its advantages of random reading are highly consistent with the characteristics of Internet applications. However, the relatively low writing performance and the limitation of erasure times affect its application scenario, so the fusion storage system which combines solid-state disk and disk has been paid more and more attention. Therefore, the fusion storage system designed for key-valued database is a valuable combination point. According to the load characteristics of Web application, the operation of key value database is recorded sequentially by logging, and all the operations are only written to memory sequencely. when the data in memory reaches a certain threshold, it is brushed to the back end storage once. The way of logging is to optimize writing performance at the expense of reading performance, so it is naturally complementary to solid-state disk. Because of the different characteristics of back-end solid-state disk and disk, the characteristics of writing are classified and written to different devices by hierarchical method, and a file system HybridFS, with customizable file placement and migration strategy is designed and implemented. Different placement and migration strategies can be selected for files with different characteristics in the same file system by monitoring and analyzing scripts. According to the file access characteristics of key-valued database, log class files are mostly written and rarely read at a time, so they are written directly to disk. Metadata files are read and written frequently, but the size and number of files are small, so they are written to solid-state disks. The data file is selectively written to disk or solid state disk according to the characteristics of load because of its mass, fixed length and multiple reading at a time. The choice of usage probability for pure write load is 5% 鈮

本文編號：2492413