From lixia@arisia.Xerox.COM Tue Oct 24 09:24:07 1989
Date: Tue, 24 Oct 89 09:21:45 -0700
From: Lixia Zhang <lixia@arisia.Xerox.COM>
To: estrin@usc.edu, mckenney@spam.istc.sri.com
Subject: Re: Proposed network experiments
Cc: internet-testbed@venera.isi.edu
Status: R

		3. Congestion control....what about experimenting with Lixia
		Zhang's flows...
	
	As I understand them, Lixia's flows use a bandwidth reservation scheme
	with rate-based flow control.  The congestion avoidance mechanisms I
	listed are 
	targeted more towards pure datagram networks.  However, it would seem
	that Lixia's flows would require a mechanism resembling pure datagram
	congestion avoidance to handle traffic consisting mostly of
	transactions (or flow setups!) 
	in a fair manner.  So, it looks to me like there might be some overlap.
	

One shouldn't forget that there are long control delays involved in those
feedback algorithms mentioned (DEC-bit, random drop, slow-start, etc).
Therefore although they are used in datagram networks, they DO NOT work if
traffic consists mostly of transactions.

Those feedback algorithms show any effect ONLY when traffic consists of
long-lasting connections, in which case the flow model can do a much
better job.  It doesn't seem likely that transaction applications would
compose of the dominant portion of the load in high-speed networks, and
they will receive good service if the network can have the utilization
under control -- that's the goal of a Flow Net.

Lixia