Multiple Networks

In this section, we study the relationships between server placement and the density of network peering links. By ``peering links'', we mean both the peering and transit relationship between two ISPs. As these links aggregate and transport traffic from one domain to another, their limited capacities contribute significantly to the user experienced network congestion. Additionally, these network exchange points maybe located off the optimal path, resulting in longer and more circuitous routes. One way to circumvent these congestion points is to use co-location services, where servers can access multiple networks and can route traffic directly to these networks without going through the exchange points. We demonstrate the relative performance with and without server co-location in Figure 5.4.

Figure 5.4: Variation on Network Peering Density

[Random Network: 5 networks, 100 nodes per network, range = 20, $TX_{scope} = 10$, vicinity = 2]

[width=0.8]figure/chap5/random_peer_density.eps

[Geographic Networks: 5 regional networks, 1 national networks, total nodes = 96, range = 800 km]

[width=0.8]figure/chap5/geo_peer_density.eps

4pt

Table 5.3: Number of Peering Links In Use

	Random Network			Geographic Network
Density	5 networkss, 100 nodes per network			5 regional, 1 national network, total 96 nodes
	total links	co-location	no co-location	total links	co-location	no co-location
0.2	83.05	22.16%	38.77%	11.6	17.24%	80.17%
0.4	158.05	23.28%	37.52%	18.9	10.05%	82.54%
0.6	238.60	21.42%	38.98%	27.4	6.20%	82.48%
0.8	317.75	22.19%	35.22%	36.3	8.82%	90.36%
1.0	410.1	21.24%	35.94%	45.2	7.30%	96.90%

For this simulation, we use two network configurations: one is constructed from 5 random graphs, the other is constructed from 5 regional networks and 1 national network in the geographic model. Hereafter, we will use the term m-n to denote geographic networks consisting of $m$ regional networks and $n$ national networks. We use $\tx_p = 1$, so every pair of networks is always interconnected, unless they do not have any common presences in any of the peering regions. The peering density in Figure 5.4 determines the number of peering points of two networks.

Figure 5.4 shows the number of required servers for the lower bound with co-location and the IR algorithm with and without co-location. The co-location service reduces the number of required servers by as much as 50% when peering is sparse. The geographic graph appears to be more sensitive to the peering density, as the performance of IR with no collocation approaches the lower bound when the peering density approaches 1. This is because the peering link is ``cheaper'' in a geographic network than in a random network. In the geographic network, two networks peer only if they both have presences in the same metropolitan area, which results in the delay on the peering link being zero. Contrarily, the peering link has a positive delay in the random graphs which adds to the server to client delay. Additionally, the ``hot-potato'' routing policy always selects the closest peering link rather than the one with the lowest delay. These combined effect determines that the increased peering density in random networks does not help much in reducing the server cost. Table 5.3 summarizes the number of links used in the two scenarios. In the random graph, the percentage of peering links in use stays almost constant; even more of them are available with increasing peering density. On the other hand, large percentages of peering links are in use in the geographic network when no co-location is allowed. The results for the case of co-location show that only a very small number of peering links are needed if servers are able to access multiple metropolitan area networks.

Figure 5.5: Relative Performance Ratio Against Lower Bound

[Random Network: range = 20, vicinity = 2, 100 nodes per network.] [width=0.45]figure/chap5/peer_random_ratio.eps [Geographic Networks: range = 800 km.] [width=0.45]figure/chap5/peer_geo_ratio.eps

Figure 5.5 shows the relative performance ratio of the IR algorithm, with and without co-location, against the lower bound. This is the same measure as those in Figure 5.4 but with more varieties on the network configuratons. We varied the number of random networks as 2, 5 and 10 and used two configurations for the geographic networks: 5-1 and 0-2. The results are mostly consistent with that in Figure 5.4. Additionally, it suggests that the performance of non co-located servers can degrade greatly with the larger number of networks.