A Study Of Optimization Of Performance For Memcached In Cloud Computing

Memcached is a distributed cached system of high performance, which is widely used in cloud computing. It establishes caches in memory to reduce the number of accessing data, and improve the speed of reading and writing data. However, with the rapid development of the large-scale clouding computing and Web applications, the volume and speed of data storage are growing exponentially, Memcached has been unable to meet the growing demand for reading and writing fast in the efficiency of space utilization and concurrent execution. So, there is an urgent need to address the problem for the performance optimization of Memcached.Firstly, the paper compares the accessed features of data in different memory resource pools for Memcached, then result shows that the operation of Memcached is dominated by reading. However, the mechanism of locking is not necessary in read operations, and the small key/value object is stored in Memcached as the cached data, making Memcached improve highly in terms of space utilization and performance.Then, the paper compared chained, linear detection, Cuckoo hashing and Hopscotch hashing, and presented three methods for Memcached system: 1)Hopscotch Algorithm replaced chained in Memcached, so as to reduce the overhead of thread synchronization and improve concurrent performance and space utilization; 2)CLOCK Algorithm replaced LRU Algorithm in Memcached, so as to optimize locking mechanism, and improve the performance of reading; 3)a Cached-Partitioned hash-table is presented. It partitions the hash-table into partitions and assign a partition to a particular core, which owns its hash-table algorithm and page-replacement algorithm, and thread level parallel processing for each hast-table, so as to increase the efficiency of concurrent accessing for the system.Finally, this paper implements these three methods based on Open Stack, The experimental results show that, under the same workload, the optimized caching system has been improved the parallelism and utilization, reduced the number of accessing memory of finding the key-value, and increased to 1.5 times around as compared withthe original one in performance.