Performance enhancement of a large-scale web caching system

Abstract

The need for performance enhancement of a large scale Web caching system is becomingly increasing, since the World Wide Web proliferates as the unique channel of information communication. We have identified the I/O bottleneck of a storage subsystem as the major performance limiting factor for a large scale web caching system. Since I/O operations are caused by both metadata operations and data operations, we have proposed new data structures and algorithms that can reduce I/O overhead caused by each of them. First, we have proposed metadata structures for a large scale Web caching system, which utilizes the benefits of NAND flash memory. We have proposed an efficient index structure for NAND flash memory, called the Dynamic Forest. By accumulating the changes to the index in RAM and writing them together sequentially at the end of the index, Dynamic Forest can significantly reduce write operations. Consequently, a large scale Web caching system adopting Dynamic Forest can achieve lower latency for metadata access. We also have proposed an efficient NAND flash-optimized hash index structure, called Multi Level Dynamic Hashing. Like Dynamic Forest, the proposed hash index structure completely eliminates in place writes for index updates. Consequently, the proposed index structure can minimize the number of write operations and improve the performance for metadata access. Second, we have proposed an effective admission control algorithm, called Adapitive Frequency-based Admission Control (AFAC), for a large scale Web caching system, which can reduce the I/O load imposed upon storage subsystem. By enforcing the hotness of admitted objects adaptively, AFAC can significantly reduce disk I/O while maintaining a high hit ratio. The main benefits of our proposed methods are as follows: First, our proposed index structures, specially optimized for NAND flash memory, can make a Web caching system operate fast. Adopting NAND flash memory for storing metadata can make t...