Check out the new USENIX Web site. next up previous
Next: Implementation Up: IO-Lite Design Previous: Cross-Subsystem Optimizations

Operation in a Web Server

In this section, we describe the operation of IO-Lite in a Web server as an example. We start with an overview of the basic operation of a Web server on a conventional UNIX system.

A Web server repeatedly accepts TCP connections from clients, reads the client's HTTP request, and transmits the requested content data with an HTTP response header. If the requested content is static, the corresponding document is read from the file system. If the document is not found in the filesystem's cache, a disk read is necessary.

Copying occurs as part of the reading of data from the filesystem, and when the data is written to the socket attached to the client's TCP connection. High-performance Web servers avoid the first copy by using the UNIX mmap interface to read files, but the second copy remains. Multiple buffering occurs because a given document may simultaneously be stored in the file cache and in the TCP retransmission buffers of potentially multiple client connections.

With IO-Lite, all data copying and multiple buffering is eliminated. Once a document is in main memory, it can be served repeatedly by passing buffer aggregates between the file cache, the server application, and the network subsystem. The server obtains a buffer aggregate using the IOL_read operation on the appropriate file descriptor, concatenates a response header, and transmits the resulting aggregate using IOL_write on the TCP socket. If a document is served repeatedly from the file cache, the TCP checksum need not be recalculated except for the buffer containing the response header.

Dynamic content is typically generated by an auxiliary third-party CGI program that runs as a separate process. The data is sent from the CGI process to the server process via a UNIX pipe. In conventional systems, sending data across the pipe involves at least one data copy. In addition, many CGI programs read primary files that they use to synthesize dynamic content from the filesystem, causing more data copying when that data is read. Caching of dynamic content in a CGI program can aggravate the multiple buffering problem: Primary files used to synthesize dynamic content may now be stored in the file cache, in the CGI program's cache as part of a dynamic page, in the server's holding buffers, and in the TCP retransmission buffers.

With IO-Lite, sending data over a pipe involves no copying. CGI programs can synthesize dynamic content by manipulating buffer aggregates containing data from primary files and newly generated data. Again, IO-Lite eliminates all copying and multiple buffering, even in the presence of caching CGI programs. TCP checksums need not be recomputed for portions of dynamically generated content that are repeatedly transmitted.

IO-Lite's ability to eliminate data copying and multiple buffering can dramatically reduce the cost of serving static and dynamic content. The impact is particularly strong in the case when a cached copy (static or dynamic) of the requested content exists, since copying costs can dominate the service time in this case. Moreover, the elimination of multiple buffering frees up valuable memory resources, permitting a larger file cache size and hit rate, thus further increasing server performance.

Finally, a Web server can use the IO-Lite facilities to customize the replacement policy used in the file cache to derive further performance benefits. To use IO-Lite, an existing Web server need only be modified to use the IO-Lite API. CGI programs must likewise use buffer aggregates to synthesize dynamic content.

A quantitative evaluation of IO-Lite in the context of a Web server follows in Section 4.

next up previous
Next: Implementation Up: IO-Lite Design Previous: Cross-Subsystem Optimizations
Peter Druschel