In this web page we attempt to summarize what we collectively know about the characteristics of aggregate traffic, and what we know about realistic or interesting models for generating aggregate traffic for experiments and simulations. Good models of aggregate traffic are needed for analysis as well (so that the analysis is not restricted to models of one-way traffic with long-lived flows with a common round-trip time). In addition, we attempt to identify key open issues. One end goal is a set of reference scenarios for generating aggregate traffic, for use in simulations, experiments, and analysis.
We are not in this web page investigating models for end-to-end paths. Such models would be needed in simulations, experiments, and analysis investigating the end-to-end performance of transport protocols. Key properties of end-to-end paths include loss rates and per-packet delay, as well as reordering, corrupted packets from noisy links, reverse-path congestion, asymmetric bandwidth, and the like. The performance of transport protocols is of course also affected by the router mechanisms and competing traffic experienced along the path.
* Is it helpful to have separate models for different types of congested links: e.g., access links, campus links, transoceanic links, links to public or private peering points, etc.?
* What are "typical" patterns for the level of congestion?
Answer: We don't know. The "level of congestion" can be quantified by the packet drop rates at the queue. The Internet Traffic and Weather section of the web page on Measurement Studies of End-to-End Congestion Control in the Internet includes pointers to the Internet Traffic Report, the Internet Weather Report, the Internet End-to-end Performance Monitoring Group, and other sites that have long had measurements of the packet loss rates of pings to various routers in the Internet. Most of this measurement is necessarily about end-to-end path properties, rather than direct router measurements, but it gives an upper bound on packet loss rates for the links along that path.
* How often are there periods of extreme congestion, e.g.,
from flash crowds, DoS attacks, link failures, or other causes?
Answers. The summary, so far, is that a range of more than 10:1 in round-trip times seems common, but with most (85%, in one case) of the connections having round-trip times between 15 and 500 ms.
* For round-trip time measurements greater than 500 ms, to what extent is this delay due to queueing delay? to routing problems? to delay at the end-node?
Answer: The section on Mice and Elephants on the web page on Measurement Studies of End-to-End Congestion Control in the Internet includes pointers to a range of measurements for the distribution of flow sizes. The distribution of connection sizes is usually modeled by a log-normal distribution for the body of the distribution, and with a heavy tail (e.g., a Pareto distribution with shape parameter less than two).
The web page on Self-Similarity and Long Range Dependence in Networks also has some pointers to work in this area.
The web page on Modeling Peer-to-Peer Traffic.
Answers: Recent measurements tend to show 90-95% of the bytes on a link from TCP. The section on Bandwidth used by Different Traffic Types on the web page on Measurement Studies of End-to-End Congestion Control in the Internet includes pointers to a range of studies showing the traffic breakdown on various links by protocol and by application.
* How can this non-TCP traffic be characterized, in terms of the applications and other characteristics.