To make our evaluation less dependent on a particular hit rate, we evaluate cache performance separately on hits and on misses for a given list of URLs. Sites that know their hit rate can use these measurements to evaluate the gain themselves. Alternatively, the reader can compute expected savings based on hit rates provided by earlier wide-area network traffic measurement studies [7][10].
Figures 3 and 4 show the cumulative distribution of response times for hits and misses respectively. Figure 3 also reports both the median and average response times. Note that CERN's response time tail extends out to several seconds, so its average is three times its median. In the discussion below, we give CERN the benefit of the doubt and discuss median rather than average response times.
To compute Figure 4, ten clients concurrently referenced 200 unique objects of various sizes, types, and Internet locations against an initially empty cache. A total of 2,000 objects were faulted into the cache this way. Once the cache was warm, all ten clients concurrently referenced all 2,000 objects, in random order, to compute Figure 3. No cache hierarchy was used.
These figures show that the Harvest cache is an order of magnitude faster than the CERN cache on hits and on average about twice as fast on misses. We discuss the reasons for this performance difference in Section 3.4. We chose ten concurrent clients to represent a heavily accessed Internet server. For example, the JPL server holding the Shoemaker-Levy 9 comet images received a peak of approximately 4 requests per second, and the objects retrieved ranged from 50-150 kbytes.
For misses there is less difference between the Harvest and CERN caches because response time is dominated by remote retrieval costs. However, note the bump at the upper right corner of Figure 4. This bump comes about because approximately 3%of the objects we attempted to retrieve timed out (causing a response time of 75 seconds)-either due to unreachable remote DNS servers or unreachable remote object servers. While both the Harvest and the CERN caches will experience this long timeout the first time an object retrieval is requested, the Harvest cache's negative DNS and object caching mechanisms will avoid repeated timeouts issued within 5 minutes of the failed request. This can be important for caches high up in a hierarchy because long timeouts will tie up file descriptors and other limited system resources needed to serve the many concurrent clients.