Building high-performance web-caching servers

World-Wide Web is one of the primary applications of Internet today. Web-caches can decrease bandwidth consumed by HTIP traffic and improve user experience by decreasing Web object retrieval latencies. Transparent web-caches can be used by organisations to intercept and cache all HTIP traffic without significant administrative expenses and therefore minimize traffic costs and improve filtering and monitoring capabilities. Internet Service Providers use en-route transparent Web-caching on their backbone communication links to decrease amount of HTIP traffic, which currently represents a major part of overall traffic. Web-caches are used in Content Delivery Networks to push content closer to end user, greatly improving latencies of object retrieval and reducing overall Internet traffic, at the same time offloading original Webserver. Web-caches are used in accelerator mode in high-volume Web-sites, decreasing overall cost of the Web-site and/or improving its scalability and performance. Though web-caches are often invisible at first glance, modem Internet in the form we see it today would be probably not possible without wide deployment of Web-caching technology. Deployment of web-caching technology gives immediate effect; performance of properly designed web-caching system can be improved step-by-step as needed. Building high-performance web-caches capable of serving multigigabit links is a challenging task. Web-caching system must be able to handle up to 16 000 user requests per second per every gigabit of traffic (for average object size of 8KB, 1 Gbps = 2°/(8'8'21°) = 214 requests per second). Total cache size must allow to store up to several days of HTIP traffic, which is impossible to store in RAM (due to the overall price considerations), and therefore the performance and capacity of persistent storage (typically hard drives) becomes a crucial issue, because hard drives are mechanical devices and are not capable of sustained rate of random read-write operations exceeding a few hundred operations per second. Often web-caching system is the only possibility for end user to fetch web-object (even though user may not be aware of it, because all HTIP traffic will be diverted at router to transparent web-cache), and therefore reliability and failover capabilities of web-caching systems must meet strictest requirements. Last but not least web-caches must be easy to use and administer, ideally not requiring any actions on behalf of end user, and require a minimum effort from network administrators. One of the most promising solutions to the problem of web-cache scalability is webcache clustering. Clustering is a technology of building from a few building blocks (e.g. low-cost PCs) a single virtual object visible from outside as single entity. Clustering is widely used in high-performance computations, high-performance web-servers and databases, and everywhere where a single big computational or processing task can be parallelized and run simultaneously at several computers. Web caching is no exceptionsingle image web-caching clusters have the best price/performance ratio possible and provide almost linear scalability, easy extendibility, good reliability, seamless failover capabilities, and require little maintenance. In this thesis a particular implementation of web-caching cluster is proposed which is capable of handling 500Mbps of HTTP traffic at the cost less than $20000 using cheap PC hardware, high-quality open-source software. As shown below, proposed solution has a number of advantages to other methods of solving web-cache scalability problems. First, the proposed cluster scheme is designed from very beginning to take into account the properties and characteristics of web-caching application as distinct from universal approach to clustering. One of the most important differences between cached content and, say, database records is the fact that cached data can be discarded at any moment, and this will not have any grave consequences in contrast to lost database records. Second, proposed approach to cluster building does not require any additional hardware (well, probably additional network switch), changes to application web-caching software or changes to operating system. Proposed cluster architecture is entirely software-based. Third, this particular implementation of cluster uses transparent web-caches (therefore any configuration at user's side is not needed), and cluster control software is easy to install and operate, require only a minimal effort from network administrator, at the same time providing good performance, excellent scalability and automatic fault detection and failover capabilities.