2.4 Evaluation
2.4.2 Effectiveness with k Trees
Here we consider the topological aspects of routing, and more specifically we evaluate the ability of our scheme to use the underlying network effectively. We use two measures of cost: stretch and congestion.
Stretch is the factor by which the distance between two nodes is extended by the routing scheme. Since our scheme routes each packet on a tree, this is the ratio between the distance on the tree and the distance on the full graph. For example, in Figure 2.5 the distance between node i and node j is 4 on the green spanning tree while these nodes are adjacent in the original graph. This means that the path between node i and node j is stretched 4 times. Given a set of k trees, the stretch for the path between two nodes is the expected
1Internet AS-level topology archive. Data retrieved 29/06/2012. http://irl.cs.ucla.edu/
37 2.4 Evaluation
a b c
d e f g h
i j k
Figure 2.5. Stretch problem: stretchT(i, j) =
distanceT(i, j) distanceG(i, j) = 4 a b c d e f g h i j k
Figure 2.6. Congestion problem: congestionT(e, b) = loadT(e, b) = 3
stretch; since we choose trees uniformly at random, it is simply the average stretch.
Congestion is the factor by which the usage of a link would grow using the routing scheme as compared to an optimal usage of the full network graph. The optimal usage here refers to the link usage with a distribution of traffic that achieves the best possible throughput. In practice, for each tree T , given a link (u, v) in T , we compute the cut defined by that link on T , meaning the partition of the nodes that are on the two sides of the link on T . We then compute the number of links that cross the cut on the original graph, which is the total capacity of the network over that cut. For example in Figure 2.6, the edge (e, b) on the green spanning tree defines a cut that splits the nodes of
38 2.4 Evaluation
the tree into two partitions. On the original graph, three edges of (e, b), (e, f) and (i, j) connect these two partitions together, hence the loadT(e, b) = 3. We
assume that, for the portion of traffic routed on T (1/k of the total traffic for k trees), the link (u, v) would need to carry the traffic that could instead go over all the links that cross the cut. So, for a cut of size sT,u,v on a tree
T out of k trees, link (u, v) is given a congestion of sT,u,v/k, and the total
congestion of that link is the sum of its congestion for all the k trees. Notice that this congestion factor is a very conservative measure, since it uses the globally optimal allocation of flow for all network cuts as a baseline.
1 3 5 7 9 11 2 4 8 16 32 Stretch Number of Trees min,1%,50%,99%,max L LC
Figure 2.7. Path Stretch.
In Figure 2.7 we show the expected stretch for various sets of global (AS- level) trees. We generate sets of 2, 4, 8, 16, and 32 trees using the heuristics H1 and H2 discussed in Section 2.3.4. The label L in the plots refers to the latency- only heuristic, H1, while LC refers to the latency-and-congestion heuristic, H2. Each box plot in the chart shows the minimum, the 1-percentile, the median, the 99-percentile, and the maximum. The plot shows that the maximum ex- pected stretch decreases with more trees, while more trees lead to a minor increase of the median (expected) stretch. Despite the growth, we can see that the stretch is low: the median always remains under 2 and the 99-percentile under 3. There is also a clear difference between the two heuristics: heuristic L achieves better results than LC.
Our experimental analysis is consistent with another study on the approx- imability of the AS-level topology with trees. Krioukov et al. [49] studied var- ious compact routing schemes with a theoretical expected stretch of 3 [28, 79], and found that in practice, on two AS-level topologies measured by the Skitter
39 2.4 Evaluation
and DIMES tools, respectively, their average stretch is instead very close to 1. The result of their study is shown in Table 2.1.
average stretch
Skitter DIMES
scheme (9204 nodes) (13931 nodes) number of trees
TZ 1.08 1.13 O(n˜ 1/2)
BC 1.06 1.03 O(n˜ 2/3)
TZ/BC hybrid 1.02 1.01 O(n˜ 2/3)
Abraham 1.35 1.45 O(n˜ 1/2)
Table 2.1. Stretch of Compact Routing Schemes in Practice.
These routing schemes in practice provide near optimal routing paths by using quite a large number of trees. The routing scheme we described in this chapter also achieves an average stretch very close to 1, but with significantly smaller sets of trees. Note that the compact routing schemes that were the subject of Krioukov’s study, guarantee a theoretical upper bound for the stretch of paths. Our routing scheme does not provide this guarantee and hence it is able to achieve near optimal average stretch with only a constant number of trees for the AS-level topology network.
0 10 20 30 40 50 60 2 4 8 16 32 Congestion Number of Trees 1%,5%,50%,95%,99% L LC
Figure 2.8. Link Congestion.
Figure 2.8 shows the congestion for the same set of trees of Figure 2.7. This plot shows the 1-percentile, 5-percentile, median, 95-percentile, and 99- percentile of the distribution. The salient result is that most links experience no congestion penalty at all, experiencing a congestion factor of 1, and further that
40 2.4 Evaluation
extreme levels of congestion are reduced when using more trees. As expected, the congestion factors for the L heuristic are higher as compared with the LC heuristic.
The analysis of stretch and congestion shows that different tree-building strategies may be used to achieve different design goals. More importantly, the general conclusion we can draw from this analysis is that even small sets of trees can cover the Internet at the AS-level topology quite well, with only minimal cost in terms of path-length stretch and link congestion.