James Hu et al. [17] perform an analysis of Web server optimizations. They consider two different architectures, the multi-threaded architecture and one that employs a pool of threads, and evaluate their performance on UNIX systems as well as Windows NT using the WebStone benchmark.
Various researchers have analyzed the processing costs of the different steps of HTTP request serving and have proposed improvements. Nahum et al. [25] compare existing high-performance approaches with new socket APIs and evaluate their work on both single-file tests and other benchmarks. Yiming Hu et al. [18] extensively analyze an earlier version of Apache and implement a number of optimizations, improving performance especially for smaller requests. Yates et al. [31] measure the demands a server places on the operating system for various workloads types and service rates. Banga et al. [5] examine operating system support for event-driven servers and propose new APIs to remove bottlenecks observed with large numbers of concurrent connections.
The Flash server and its AMPED architecture bear some resemblance to Thoth [9], a portable operating system and environment built using ``multi-process structuring.'' This model of programming uses groups of processes called ``teams'' which cooperate by passing messages to indicate activity. Parallelism and asynchronous operation can be handled by having one process synchronously wait for an activity and then communicate its occurrence to an event-driven server. In this model, Flash's disk helper processes can be seen as waiting for asynchronous events (completion of a disk access) and relaying that information to the main server process.
The Harvest/Squid project [8] also uses the model of an event-driven server combined with helper processes waiting on slow actions. In that case, the server keeps its own DNS cache and uses a set of ``dnsserver'' processes to perform calls to the gethostbyname() library routine. Since the DNS lookup can cause the library routine to block, only the dnsserver process is affected. Whereas Flash uses the helper mechanism for blocking disk accesses, Harvest attempts to use the select() call to perform non-blocking file accesses. As explained earlier, most UNIX systems do not support this use of select() and falsely indicate that the disk access will not block. Harvest also attempts to reduce the number of disk metadata operations.
Given the impact of disk accesses on Web servers, new caching policies have been proposed in other work. Arlitt et al. [2] propose new caching policies by analyzing server access logs and looking for similarities across servers. Cao et al. [7] introduce the Greedy DualSize caching policy which uses both access frequency and file size in making cache replacement decisions. Other work has also analyzed various aspects of Web server workloads [11,23].
Data copying within the operating system is a significant cost when processing large files, and several approaches have been proposed to alleviate the problem. Thadani et al. [30] introduce a new API to read and send memory-mapped files without copying. IO-Lite [29] extends the fbufs [14] model to integrate filesystem, networking, interprocess communication, and application-level buffers using a set of uniform interfaces. Engler et al. [20] use low-level interaction between the Cheetah Web server and their exokernel to eliminate copying and streamline small-request handling. The Lava project uses similar techniques in a microkernel environment [22].
Other approaches for increasing Web server performance employ multiple machines. In this area, some work has focused on using multiple server nodes in parallel [6,10,13,16,19,28], or sharing memory across machines [12,15,21].