Figure 9. End-to-end Latency of the 8×8 Mesh-of-Torus network in different flow patterns.
Simulation results for the 8×8 Mesh-of-Torus network with a 256 byte packets size and different flow patterns are shown in Figure 9. We can conclude from the figure that end-to-end latency can remain at an extremely low level before the network reaches its saturation point. But if the infusion rate exceeds the saturation point, end-to-end latency are expected to soar. And if we use all-to-all traffic patterns, the network will reach its saturation point in a short time, but if we use the local traffic patterns, the network can get a higher network saturation point. That is because that in the local traffic patterns, the network can fully utilize the long torus network links. And this property is a firm proof for the idea that Mesh-of-Torus network can be more suitable for data center network with more local traffics. We use Figure 10 (a),(b),(c),(d) to show different flow patterns of the network link load distribution in a same network model, respectively.(a) all-to-all