Server Load

Throughout this chapter, we assumed that the cost of installing and operating a separate server far out weights the cost of buying more bandwidth at a location. We are also interested in the distribution of load among the servers, particularly the minimum load on a server, since it is not very cost-effective to install a server only to operate under very small load.

Figure 5.6 shows the average and minimum server load distribution under the IR algorithm with co-location. As expected, the average server load increases with the increase of service range. The load distribution is more even on the random networks than on the geographic networks due to the different node distribution: the geographic networks have significantly higher population density on both coastal areas which creates higher load for servers.

Figure 5.6: Characteristics of Server Load

[Random Network, 5 nets, 100 nodes per network, peering density = 0.4.] [width=0.45]figure/chap5/server_random_range.eps [Geographic Networks, 5-1 nets, total 96 nodes, peering density = 0.4.] [width=0.45]figure/chap5/server_geo_range.eps

However, the minimal load does not lift as much for both networks. It is easier to understand this effect in the geographic networks, since a few metropolitan areas, such as Seattle and Portland, are more segregated from the rest of the areas. Consequently, it takes at least one server to cover and only cover these two areas. Although it is less obvious in the random networks, it seems that there are always a small number of nodes that are particularly far away from the rest of the group. The curves labeled $l = 5$ indicate the average and minimal fraction of clients served by a single server when we enforce each node to be included in at least 5 server sets. The increase of the minimal server load from the relaxation is more visible in the geographic networks than in the random networks.