Before we can choose a solution for the network, we need to understand what types of workloads are currently being injected into the Internet and how reliable different elements in the network are. Different workloads will stress the system in different ways and will have different performance requirements and notions of quality of service. For example, a workload with traffic sent periodically in fixed-sized bursts will behave quite differently than one that has a lot of variation in terms of interarrival times, flow durations and flow rates. The first workload would be best served by a slotted network, while the second one would not.
This Thesis provides a short analysis and discussion of the traffic and failures that we see in real networks, especially in or near the backbone. Some of these results are based on my own analysis of traffic traces [131,170], other results have been reported elsewhere [31,30,102,107,106,105]. In general, one is interested in knowing both the type of application (to prioritize the performance metrics) and the distributions and correlations of:
Based on those observations, we can make some assumptions of the system workload. The most fundamental one is that flow durations in the Internet have long and heavy tails [146,56,183,23], as shown in Figure 1.6. It shows how fewer than 10% of the flows in a backbone link carry over 90% of the bytes transported in the link. There are, thus, two types of user flows: most flows are short; and then a few are very, very long and carry most of the bytes. These long flows may hog the system for extended periods of time and degrade the overall system performance significantly. All these flow characteristics will be incorporated in the performance study of packet and circuit switching in the core of the network.
![\resizebox*{0.8\columnwidth}{!}{\includegraphics[clip]{fig/HeavyTailedTraffic}}](pmf_thesis_img9.png)
|