We described the setup of test bed used for performing experiment in the last chapter. To generate load and measure performance we have used Webstone originally developed by Silicon Graphics and is now maintained by mindcraft.com. This is standard software used to benchmark commercial web servers. Different schemes were tested with everything kept identical except policy for cluster selection at DNS. Webstone software's master process controlling clients was run on one of client PCs.
Experiments were conducted by varying number of client processes from 20 to 120 in steps of 10. Webstone tries to execute roughly equal number of processes at each client. To generate uneven distribution of requests, we wrote the ID of same client machines multiple times in its configuration file. These machines then generated more load than others. When number of client processes were 20,40,..,120, clients in all geographical regions generated almost equal load (per client load i.e. number of processes running were still different). When number of client processes were 30,50,..,110 clients in region1 and region2 were running twice as many client processes as they were running with 10 less client processes (i.e. at 20,40...), while other clients were still running same number of processes, so load was highly uneven.
We have run at least ten iterations of one minute duration each for each data point and taken average of them for plotting. Each Webstone client processes made just single query to DNS before sending requests to servers (clusters) and used resolved mapping for whole testing period of one minute. So due to application level caching by webstone clients, requests from same client process reached to same cluster for one minute duration regardless of TTL value provided by DNS.