We placed CRISP caches under worst-case load from synthetic clients to evaluate the scalability of CRISP's central mapping server. For these experiments, we ran the CRISP cache on a cluster of 266MHz AlphaStation 500 workstations linked by a switched 100 Mb/s ethernet. The clients continuously fetch distinct small objects, so no accesses are intercepted by proxy hits. This results in a probe to the mapping server for every request. The synthetic nature of this workload is irrelevant because the mapping server is not affected by any access properties other than the request arrival rate and URL length.
Table 3.2 shows the average reponse time for a mapping server query in a CRISP configuration with six proxies and varying request arrival rates. The first two columns represent the worst-case and average number of clients that would generate the query rate shown in the third column. The first column (conservative) assumes that all clients surf the Web as fast as they can position and click the mouse (five seconds). The second column (average) gives the number of clients needed to generate the given query rate, assuming each client emits requests at the average per-client request rate observed in the largest AT&T trace at the time of peak load.
Table 3.2 shows that the mapping server response time, including local network delay, is well below human perception even under the expected peak load from 85,000 average clients.
Table 1: Mapping server query latency.