[e2e] tcp in high rate network

[Joe Touch]

> On Mon, 17 Jun 2002, Jonathan M. Smith wrote:
> 
>>In the absence of errors and congestion TCP can run as fast as IP (minus
>>some processing overhead) if properly tuned and presuming the end-point
>>device has the horesepower to drive it.

As Ted Faber pointed out:

	- in the presence of loss, congestion avoidance plays a
	big factor. Sally Floyd has more recent results on this that
	she posted recently.

Further:

	- in the absence of loss, slow-start is a factor, but only
	where the file size is smaller than the RTT * bandwidth
	(within a factor of 10 or so). for very large files,
	slow start isn't an issue.

	- slow-start restart is an issue for transfers without
	continuous sources. The restart process is slightly different
	from the slow-start (notably if there is return traffic)

Other internal factors, which Johnathan referred to, include:

	- data touching overheads (encryption, checksum if software...)

	- copying overheads (programmed I/O, multiple copies over the
	same bus or memory interface, etc.)

	- proper buffering, which can (when improper) lead to
	excessive context switching, livelock, etc. (see esp.
	www.psc.edu)

	- processing overheads (too much 'integrated layer
	processing' can lead, e.g., to thrashing of the CPU
	instruction cache

	- internal system bandwidth

As well as:

	- packet processing synchronization (if multithreaded or using
	parallel processing)

	- per-packet overhead, including processing and bandwidth
	overhead (mitigated by using larger packets, but
	unavoidably high for small packets)

Many of these are covered in various chapters of the following, 
including references to individual issues:

"High-Speed Networking: A Systematic Approach to High-Bandwidth 
Low-Latency Communication,"
J. Sterbenz, J. Touch, John Wiley & Sons, April 2001, ISBN: 0471330361
http://catalog.wiley.com/remtitle.cgi?isbn=0471330361&country=826

Joe