[Ground-station] MPTCP, network coding, ICN, and satellites!

[Phil Karn]

On 2/25/19 09:11, Michelle Thompson via Ground-Station wrote:
> One of the books I picked up at Information Theory and Applications
> Conference was Network and Protocol Architectures for Future Satellite
> Systems. 

When did it come out?

> The first problem you run into is that TCP views long latency links as
> congestion and does what it's good at and slows things down. 

This problem was addressed 20+ years ago with TCP header timestamps and
window scaling, both of which have been universally deployed for some
time now.

Header timestamps solves the problem I addressed my round trip timing
algorithm in the late 1980s. I chose to do it without changing TCP, but
given that the protocol could be changed, timestamps was a much better
solution.

Window scaling got around TCP's original 64KB window limit, meaning that
it couldn't send more than 64 kilobytes per round trip time. This
rapidly became a problem on terrestrial fiber, not just geostationary
satellites. The SYN (synchronize, or connection setup) packets convey a
"scale" parameter that is the base-2 log of the value to be multiplied
by all subsequent window advertisements in the session. E.g., if the
window scale parameter is 5, then all window announcements should be
left-shifted by 5 bits.

This completely solves the problem for bulk transfers over long delay
channels **provided** that the packet loss rate is kept small. TCP still
has problems when packets are lost more often than roughly once per
round trip time. There are additional enhancements (selective ACKs) to
handle this problem, but it is still best to design the subnet for a
much lower packet loss rate. How you do this is left up to the subnet
designer. See RFC 3819, Advice for Internet Subnetwork Designers.

Phil