Check out the new USENIX Web site. next up previous
Next: Persistent Connections Up: IO-Lite: A Unified I/O Previous: Implementation


In this section, we evaluate the performance of a prototype IO-Lite implementation. All experiments use a server system based on a 333MHz Pentium II PC equipped with 128MB of main memory and five network adaptors to a switched 100Mbps Fast Ethernet.

To fully expose the performance bottlenecks in the operating system, we use a high-performance in-house Web server called Flash [19]. Flash is an event-driven HTTP server with support for CGI. To the best of our knowledge, Flash is among the fastest HTTP servers currently available. Flash-Lite is a slightly modified version of Flash that uses the IO-Lite API.

While Flash uses memory-mapped files to read disk data, Flash-Lite uses the IO-Lite read/write interface to access disk files. In addition, Flash-Lite uses the IO-Lite support for customization of the file caching policy to implement Greedy Dual Size (GDS), a policy that performs well on Web workloads [10].

For comparison, we also present performance results with Apache version 1.3.1, a widely used Web server [3]. This version uses mmap to read files and performs substantially better than earlier versions. Apache's performance reflects what can be expected of a widely used Web server today.

Flash is an aggressively optimized, experimental Web server; it reflects the best in Web server performance that can be achieved using the standard facilities available in a modern operating system. Flash-Lite's performance reflects the additional benefits that result from IO-Lite. All Web servers were configured to use a TCP socket send buffer size of 64KBytes; access logging was disabled.

In the first experiment, 40 HTTP clients running on five machines repeatedly request the same document of a given size from the server. A client issues a new request as soon as a response is received for the previous request [4]. The file size requested varies from 500 bytes to 200KBytes (the data points below 20KB are 500 bytes, 1KB, 2KB, 3KB, 5KB, 7KB, 10KB and 15 KB). In all cases, the files are cached in the server's file cache after the first request, so no physical disk I/O occurs in the common case.

Figure 3: HTTP

Figure 3 shows the output bandwidth of the various Web servers as a function of request file size. Results are shown for Flash-Lite, Flash and Apache. Flash performs consistently better than Apache, with bandwidth improvements up to 71% at a file size of 20KBytes. This result confirms that our aggressive Flash server outperforms the already fast Apache server.

Flash using IO-Lite (Flash-Lite) delivers a bandwidth increase of up to 43% over Flash and up to 137% over Apache. For file sizes of 5KBytes or less, Flash and Flash-Lite perform equally well. The reason is that at these small sizes, control overheads, rather than data dependent costs, dominate the cost of serving a request.

The throughput advantage obtained with IO-Lite in this experiment reflects only the savings due to copy-avoidance and checksum caching. Potential benefits resulting from the elimination of multiple buffering and the customized file cache replacement are not realized, because this experiment does not stress the file cache (i.e., a single document is repeatedly requested).

next up previous
Next: Persistent Connections Up: IO-Lite: A Unified I/O Previous: Implementation
Peter Druschel