Research And Implementation Of Strategies To Optimize The SSD-Based Key-Value System

The rapid development of Internet technology gave birth to massive data, more and more data-centric applications penetrated into every aspects of people’s life. These applications put forward higher requirements on storage systems. Among them, how to establish an efficient index for these data became a big challenge. Key-Value system is an effective solution to establish an index of massive data, it has features of low operational latency and high throughput. However, the access to the underlying storage device of Key-Value system is mainly random access, disk-based Key-Value system is sure to have poor performance. In contrast, flash-based SSD has good random access performance and is an ideal storage device to construct the Key-Value system. This paper mainly studies the lookup optimization strategies of SSD-based Key-Value system, the main work and innovations are as follows:First, we propose a read and write optimization mechanism that bases on chain compression, for the problem of the poor compression performance of the SSD-based Key-Value System called SkimpyStash. Compression activity can write a chain of Key-Value pairs that in different SSD pages to the same page, the strategy can reduce compression times by adding a compression counter in the hash table directory, so as to effectively reduce the impact of compression activity on the client. Experiments show that the total compression times reduce 10% to 34% on average. In addition, we test the effect of compression activity on the average lookup time, pointing out the importance of a reasonable compression parameter.Secondly, we propose a Get optimization strategy based on hot data identification and relevance detection, for the problem of massive Key-Value pairs have different access heat and some of the Key-Value pairs have relevance in the whole system. The strategy can reduce the read operations of SSD by putting the hot data in the head of the chain and putting the relevant data in the same page, so as to effectively reduce the time of Get on the hot Key-Value pairs and relevant Key-Value pairs. Experiments show that the hot data identification strategy can reduce the SSD read operations by 6%, the relevance detection strategy can reduce more than 10%.Finally, we propose using file mapping mechanism to reduce the access of SSD, for the problem of the high throughput requirement in the Key-Value system, so as to increase the system throughput and reduce access latency. The mechanism will map a certain size of file into the memory when there is a read or write operation. Experiments show that the optimization can reduce the average lookup time by 21% to 30%.