An Efficient Web Cluster with Content-Aware Request Distribution on Linux Kernel 2.6

• Main Authors: Shang-Yi Zhuang, 莊尚益
• Other Authors: Mei-Ling Chiang
• Format: Others
• Language: en_US
• Published: 2008
• Online Access: http://ndltd.ncl.edu.tw/handle/45583625164876908772

碩士 === 國立暨南國際大學 === 資訊管理學系 === 96 === The web cluster has been commonly used in popular web sites because of its high availability and scalability. The web cluster can be content-aware or content-blind. Many researches have focused on building the content-aware platform because dispatching requests from clients to servers according to the content of request (i.e. URI) can achieve high locality in back-end servers’ main memory caches and get high system throughput. In our previous work, we have proposed a content-aware platform named LVS-CAD which uses TCP Rebuilding mechanism in back-end servers to rebuild the connection with the client and the fast handshaking mechanism in the front-end server to establish connection with the client. Both mechanisms were implemented on Linux kernel 2.4.18. In recent years, Linux kernel 2.6 has been released and added in many new features to make a web server more stable and more efficient. In this thesis, in order to gain performance improvement from these new features in kernel 2.6, we have transplanted the LVS-CAD on Linux kernel 2.6.18. In other words, we have implemented the TCP Rebuilding mechanism and the fast handshaking mechanism in Linux kernel 2.6.18. Besides, in order to make the LVS-CAD platform more flexible and easier to maintain and debug, we have also implemented these two mechanisms as loadable kernel modules that can be dynamically loaded/unloaded into/from Linux kernel. In addition to the transplant of the LVS-CAD platform, we have also proposed and implemented two content-aware request dispatching policies named HyLARD and LARD/FR policies on the LVS-CAD platform. These two policies enhance the well-known LARD/R policy. Moreover, in our experiments, we found that reducing multiple handoffs can efficiently improve performance of a web cluster. Therefore, we have implemented and integrated the idea of reducing multiple handoffs of the proposed policies into the LVS-CAD platform named Linux Virtual Server with efficient Content-Aware Dispatching platform (LVS-eCAD). Experimental results show that the LVS-CAD platform on Linux kernel 2.6 can perform 239.12% better than the LVS-CAD on Linux kernel 2.4. Besides, our proposed policies can achieve 60.24% better performance than the LARD/R policy on LVS-CAD with Linux kernel 2.4. The LVS-eCAD platform can outperform LVS-CAD and LVS by about 24.62% and 24.55% respectively under Weighted Round-Robin policy with Linux kernel 2.6.