Figure 5 demonstrates the performance of a Harvest cache configured as an httpd-accelerator. In this experiment, we faulted several thousand objects into a Harvest cache and then measured the response time of the Harvest cache versus the NCSA and Netsite httpd. Notice how Harvest serves documents that hit the httpd-accelerator with a median of 20 milliseconds, while the medians of Netscape's Netsite and NCSA's 1.4 httpd are each about 300 ms. On objects that miss, Harvest adds only about 20 ms to NCSA's and Netscape's 300 ms median access times.
Note that on restart, the httpd-accelerator simply refaults objects from the companion Web server. For this reason, for added performance, the accellerator's disk store can be a memory based filesystem (TMPFS).
The httpd-accelerator can serve 200, small objects per second. We measured this by having 10 clients concurrently fetching two hundred different URL's, in random order, from a warm cache. Note that this is faster than implied by the average response time; we explain this below.
