Check out the new USENIX Web site. next up previous
Next: Conclusions Up: A Flexible and Efficient Previous: Module Performance


Related Work

Sections 1 and 5 discuss the Internet Content Adaptation Protocol (ICAP) and compare it with the API presented here. This section discusses other related work.

The event-aware nature of our API is clearly motivated by previous research on event-driven servers [3,19], particularly by work showing their scalability benefits versus traditional multi-threaded or multi-process servers [13]. Through the use of dynamic loading of modules coupled with an event-driven proxy core, our implementation achieves performance comparable to adding existing states into a event-driven server.

The TransSend/TACC proxy [6] performs content adaptation using a system akin to Unix pipes, where thread-based modules receive a stream of bytes from the main proxy. In comparing the relevant section of that work, we find differences in architecture and coverage. By exposing an event-aware API, modules can choose to avoid the overhead of threads or processes, yielding higher scalability. In terms of coverage, since our API is specifically designed for caching proxy servers, it contains content management and utility functions not present in other APIs.

Commercial proxy caches by Inktomi and Novell have previously announced APIs. No public documentation of functionality or performance is available for the Inktomi API. The Novell Filter Framework provides a content adaptation system for Novell Border Manager and Volera Excelerator [12]. Filter modules are supported using only a callback model. Additionally, the system appears to be tightly integrated with the operating system kernel because standard libraries for memory management such as malloc are not available; instead, all memory allocation and management must take place through kernel-style memory chains. Filter Framework was never fully implemented, and has now been discontinued.

Many academic studies and commercial products have been based on modifying the source code of the Harvest cache and its successors such as Squid [3,18]. However, if these source code changes are not integrated into the public releases of the proxy, the groups maintaining the modified proxy must track the public releases to incorporate bug fixes, performance improvements, and new features. In contrast, changes to an API-enabled proxy server only affect modules if the API specification changes.

Research in content adaptation has often shown the difficulty in modifying proxy behavior. For example, Chi et al. describe a proxy server that modifies Squid to compress incoming data objects, but keeps the original content-length header intact [4]. That work tests the proxy with a modified client that ignores the content-length header. In an API-based solution, deleting or changing headers is a simple task since the API provides the needed infrastructure.

The ad insertion proxy developed by Gupta and Baehr uses special header lines that provide information about what parts of an HTML document are ads that can be replaced by the proxy in cooperation with the origin server [7]. Their non-caching proxy was developed specifically for this purpose. The same system could be developed with an API-enabled proxy with much less effort, as the ad replacement module could use the same special headers to communicate with cooperating servers without modifying the infrastructure for managing other HTTP headers.

Various researchers have examined the issue of content management, often to address the limitations of the HTTP protocol's handling of object expiration/staleness or to take advantage of regional proxies. The PoliSquid server develops a domain-specific language to allow customization of object expiration behavior [1]. The Adaptive Web Caching project uses proxies in overlapped multicast groups to push content and perform other optimizations related to object placement [5]. Likewise, the approach proposed by Rabinovich et al. uses routing/distance information to determine when proxies should contact neighbors versus when they should request objects directly [14]. In all of these cases, a content-management API would reduce development work of the customizers and would allow them to focus on their policies and improvements rather than the underlying mechanism.

Researchers and companies have also examined mechanisms for extending proxy server functionality using Java. The Active Cache project associates with each cacheable document a Java applet that is invoked when a proxy accesses the document [2]. Likewise, the JProxyma proxy uses Java plug-ins for performing content adaptation [9]. We believe that the API we propose can enable either approach; in particular, the use of helper processes in sample modules such as the transcoder shows that extended services can effectively launch external programs for their API interactions.

Component-based software architectures are rapidly gaining popularity in various domains of the computer industry. For instance, the applications in office productivity suites, such as Microsoft Office or the public-domain Koffice, all follow the component-based paradigm, exporting a set of APIs to other applications [10,11]. The reason for this growing popularity is identical to the one that caused us to develop an API for proxy caches: providing the ability to control an application without having to modify it.


next up previous
Next: Conclusions Up: A Flexible and Efficient Previous: Module Performance
Vivek Sadananda Pai 2003-01-17