We have setup a test bed having 3 clusters on different logical networks modeling three different geographical regions. Each cluster has one front node and two servers connected to front node for that cluster. Servers are configured to have aliased secondary IP address same as cluster IP and have local private IP address that is used for IP packet forwarding by the front node.
We have used ten clients to generate requests to web server system. Clients were also assigned IP addresses in such a way that clients in same geographical region had higher order seven bits as mentioned in RFC 1466 [19] describing guidelines for management of IP address space. Using this RFC, we modeled three geographical regions for clusters - region1 as Europe (machines had IP addresses in 194.*), region2 as North America (machines had IP addresses in 198.*) and region3 as Pacific Rim (address with 202.*). Similarly three clients each were present in region1 and region2 and two clients in region3. We also had three more clients in other regions which represent mix of client not falling in either of three regions. A DNS was also setup to resolve IP addresses of clusters. Actual test bed setup used for performing experiments is shown in Figure 4.2.
To model WAN effects, artificial delays and packet losses were introduced using Nistnet software. Half of delay (in specified range) and losses occurred in one direction and half in the reverse direction. Front nodes introduced delays and packet losses for packets transmitted by clients and clients introduced similar delays and losses after receiving packets from servers but before giving it to the higher protocol layers.
We have configured lower delays for IP packets sent and received between clients and servers in the same geographical region and relatively higher delays for packets between clients and servers in different geographical regions. These delays were generated randomly within specified range (say, 10-50 ms round trip delay in the same region and 50-250 ms delay across the regions).
Similarly we configured lower packet losses with higher correlation between drop of packets to model bursty lower packet losses in small distance links for links in same geographical region and higher packet losses with high correlation between successive packet drops for links across different geographical regions (e.g. 5% loss with .9 correlation on links in same region and 10% loss with .85 correlation on links connecting different regions).
More details about experiment are discussed in the next chapter containing results.
