Proceedings of INFOCOM
This paper looks at when TCP is "not" TCP by analysis of five years of data on a Japanese data set. That is to say, when TCP throughput is limited by mechanisms other than traditional TCP rate control (loss or delay in the network feedback causing a reduction in window size).
Other mechanisms are important:
1) Application limiting where the sender "dribbles" out data more slowly, for example in the way that you tube does, to reduce their bandwidth.
2) Window size limitations -- where hosts have an OS built in limitation on how large the TCP window can be.
3) Middle box/receiver window tweaking -- where the receiver or (more likely) a middle box tweaks the advertised window size to reduce throughput.
It is found that in the traces studied these three mechanisms account for more than half the packets. The traces include data from well known sites such as YouTube and it seems likely that the findings are more general than just applicability to this particular trace set.
In general this paper finds that TCP in the wild is not behaving in the way it is traditionally taught... by a variety of mechanisms, TCP is not "filling a pipe" and "controlled by loss"... other mechanisms are at play beyond traditional TCP congestion control.
This talk is essentially the same as that delivered in Cambridge two months earlier (alas no progress on this research for that period).
The research is based on two papers:
A longitudinal analysis of Internet rate limitations -- http://www.richardclegg.org/tcp_rate_infocom_2014
and
On the relationship between fundamental measurements in TCP flows -- http://www.richardclegg.org/tcp_limitations_icc_2013
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.