Again it argues that TCP is no longer mainly controlled by loss and congestion but instead by algorithms and settings under the control of the sender or receiver deliberately or accidentally designed to restrict throughput for a variety of reasons (for example limiting video sending to the rate at which the viewer is watching).
It contains extended discussion of the methodology and in particular how flight and RTT data was extracted from passive traces.
This paper used a likelihood based framework to create a rigorous way to assess models of networks. Network evolution is broken down into an operation model (it decides the 'type' of change to be made to the network, e.g. "add node" "add link" "remove node" "remove link") and an object model (that decides the exact change -- which node/link to add).
The system is shown to be able to recover known parameters on artificial models and to be useful in analysis of real data.
The essential findings are that TCP is not working as we expect. The expected correlation between throughput and packet loss is not found. The correlation with delay (RTT) is as expected -- throughput proportional to 1/delay. A high correlation with flow length is found -- longer flows have higher throughput. However, this may be a sampling error due to the restricted length of the samples used.
The TCP flows studied are broken down by assumed cause where TCP mechanisms are thought not to be the primary cause for throughput:
1) Application limited -- an application decides to reduce its own flow by deliberately not sending data.
2) Host Window limited -- one or other host has a low maximum window size that restricts flow.
3) Advertised window limitation -- a middlebox or the receiver manipulates their advertised window size to reduce the flow.
More than half of TCP flows (and more than 80% of long flows) are limited by these mechanisms and not by traditional TCP mechanisms.
This talk is an updated version of this talk at QMUL. The difference is two slides at the end which provide insight into the sampling issues related to the data.
The key message of this paper is that TCP/IP does not work in the real world as it is generally taught. The idea of a connection when one side sends data as fast as possible controlled by loss to fill a pipe is not what happens in the real world.
The talk analyses passive traces with the aim of explaining what are the root causes of bandwidth on a connection. Theoretical results show that in equilibrium an unconstrained TCP flow has a bandwidth proportional to 1/RTT and 1/sqrt(p) where p is probability of packet loss. The experimental results here show different results, however. In particular, while the relationship with RTT is upheld, the relationship with loss is not found. A strong relationship with the length of flow is found. Longer flows have faster throughput in proportion to sqrt(L) where L is the length of the flow in packets.
A follow up analysis looks at the causes of throughput. It is found that less than half of flows are governed by loss. Flow bandwidth is very often governed by applications -- for example you tube deliberately throttles traffic so that users do not download too far ahead. Some flows are governed by operating system restrictions which do not scale window sizes. Some flows are governed by middleboxes which manipulate the window size. It is these restrictions which, the network studied, are the primary method which restricts bandwidth on connections.