[1423] in Commercialization & Privatization of the Internet
Re: Volume-sensitive charging
daemon@ATHENA.MIT.EDU (stev knowles)
Tue Oct 1 20:22:01 1991
Date: Tue, 1 Oct 91 14:03:53 -0400
To: vlad@erg.sri.com
From: stev@ftp.com (stev knowles)
Cc: com-priv@psi.com, vlad@erg.sri.com
internets. These results were published at INFOCOM '91. Here is a 2-page
summary in LATEX.
it would have been nice if you would have run this through latex and piped
the output into your mail.
These charging measures and algorithms have the property that,
by minimizing the individual domain's expenses, they also minimize the
global resource consumption over all possible paths and flows. All of the
above results and algorithms hold when we replace ``price'' by ``delay''.
i dont necessarily believe this is true. if you route purely based on cost,
you can end up with a T1 line that is constantly dropping packets, while the
T3 line is quiet, because it "costs more". or am i missing something?
We are concerned with the billing policies for important or necessary
packets, i.e., those packets whose loss requires retransmission. We
i believe this is a flawed assumption. i dont care how long it takes my mail
to transit the US, since i am not waiting for it to move. on the other
thand, i care a great deal about my supdup (or, for the rest of you, telnet)
response time. as such, retransmissions are alot worse with supdup (telnet)
than with mail, even though the mail may be more important.
as {\it hop-based}, assumes that a copy of the packet is
kept at each intermediate AD until that AD knows that the packet has
successfully cleared the next domain. The AD will
retransmit the packet if it gets lost in the next domain.
this is an interesting view, and will surely *screw* up most slow start and
congestion avoidance code that i have heard of. actually, it woudl be
strange to view IP in this light, since it would require interAD updates
about which packets have passed, but you can *very* easily map this on to
the current EMAIL system.
This paper concentrates on the {\it route-based}
policy because it involves more difficult mathematical issues.
and is much more appropiate for us, although i can consider that this paper
may have been originally targeted at another group with much different
needs.
We have developed
cost recovery schemes that assign fair wages to individual domains for
successful delivery of packets and fair penalties for losing packets.
These methods ensure that each domain breaks even in the long run.
The new prices (wages) reflect both the
delivery costs and the risks of losses in an economically optimal way.
you assume at this point that every AD is willing to be a transit network,
and can deal with congestion issues when they are making "money" by being a
transit network.
In particular, we define the notion of price fairness.
We consider a packet transiting through $k$ administrative domains
sorry, this was too hard to read, assuming there are alot of LaTeX
formatting stuff in it. if there are not, please choose simpler variables.
Fair pricing policy should be as follows:
If domain $N_i$ loses the packet, it should reimburse the client in the
amount $u_i$ equal to $E_{i-1}$.
If domain $N_i$ successfully transitions the packet, it
should be paid by the client in the
amount of \[e_i = \frac{p_i E_{i-1} \ + \ c_i}{1-p_i}.\]
this means that there is a penality for losing a packet, since you have a
loss of money. in theory, i could take in more money than i sepnt, as a
client, if i get alot of retransmissions. an interesting system . . . .
We also present efficient algorithms
for finding the least expensive (with respect to the above prices $e_i$)
paths between different pairs of domains in lossy internets.
it seems to me that the "least expensive path" woudl involve the lossiest
path, since they would be paying kickbacks and defraying the cost of
delivery.
rather than from all sources to a fixed destination $d$
The efficient DP algorithms are applicable to this approach because the
memoriless property holds in the backward direction, i.e., the portion of an
optimal path before some intermediate node $w$
is independent of the path after $w$.
does this mean that you are assuming all paths are 1) bidirectional, and 2)
price insensitive to direction?
