Internet Experiment Note: 191 C. Sunshine
D. Cohen
J. Postel
ISI
July 1981
Comments on Rosen's Memos
INTRODUCTION
This memo comments on recent IEN's by Eric Rosen of BBN (numbers 182,
183, 184, 187, 188, 189) [1,2,3,4,5,6]. We think these notes raise
some important and interesting issues which require further
discussion. In the following we focus on the points of disagreement
(but don't assume that we agree with something simply because we
don't mention it here). After a brief general comment we discuss
each note in turn.
There are some good points raised in this series. Unfortunately the
presentation is both verbose and incomplete. There is nothing wrong
with taking a certain aspect of a topic and exploring it at length,
but these memos seemingly present all available alternatives and
select the "best" for further development. Our concern is that, in
fact, not all alternatives are studied, and not all evaluation
criteria are given the proper weight in selecting the "best"
alternative. A minor problem is the informality of the references.
It is unclear exactly which earlier memos, reports, and papers the
author has in mind in some of the discussion, and it is unclear if
the author is aware of some very relevant material. In some sections
it appears that the author is unfamiliar with much of the relevant
material, and hence fails to include important points in his
presentation.
IEN 182
This note on "Issues in Buffer Management" is, in the main, a
description of buffer management in the ARPANET IMPs. This is quite
useful and should be food for thought for gateway designers and
implementers since gateways may have some of the same constraints and
concerns in buffer management as IMPs. However, the differences that
do exist in the goals for gateways and IMPs are not taken into
account, so the policies adopted for IMPs are not necessarily
appropriate for gateways. Differences in the level of reliability of
delivery, and the end-to-end virtual circuit vs. the datagram style
of service can lead to substantial differences in the requirements
for buffer management.
This is a useful memo in that it exposes a good deal about the buffer
management polices used in the ARPANET IMPs, information that is not
easily found elsewhere. But it is contains some weakly supported
Sunshine & Cohen & Postel Page [1]
July 1981 IEN 191
Comments on Rosen's Memos
overly broad conclusions that seem to ignore and sometimes contradict
existing results in this area.
IEN 183
This memo presents a proposal for a logical addressing mechanism in
the ARPANET, and includes a good deal of discussion of alternatives.
Interested readers should see earlier IEN's on the subject from MIT,
ISI, and Xerox, plus the classic paper by Shoch, and recent work on
"naming authorities" at Xerox, which the author fails to credit or
reference [7,8,9].
We prefer the more commonly used term "name" to the phrase "logical
address" which the author uses.
The key proposal is to include a name-to-address lookup function in
the source switches of a network so that the "user" will not have to
supply ("physical") addresses. This seems a worthwhile goal, but the
meaning of "user" seems confused between (1) people or application
programs using the network, and (2) network access software (such as
NCP or TCP) supporting (1) in the hosts. The author seems oblivious
of this distinction.
Everyone agrees that category (1) "users" should be able to use
names. Of course, most ARPANET hosts' category (2) software already
provides this function (the host table) for category (1) "users".
The proper discussion should be whether this function is best located
in the switches, or in the network support software of the hosts, but
this is not explicitly addressed by the author.
The author presents a reasonable approach to implementing a name
lookup function without requiring broadcast of dynamic changes to all
participants. A basic table of all potentially usable addresses for
each name must be distributed to all parties (the "authorized"
table), but this is expected to change relatively slowly. Entries in
this table are assumed usable ("effective") until an explicit
exception message ("destination not accessable") results from using
them. The unusable markings are reset after a time interval.
We agree that this is a worthwhile proposal, but the placement of
this function in the hosts, the switches, or a separate name lookup
service needs further discussion. Since most hosts are already
performing this function as noted above, it is clearly within their
capabilities. An advantage of placement in the switches seems to be
prevention of "spoofing" since hosts can only send/receive messages
from/for a specified name if that name is "authorized" for the
Page [2] Sunshine & Cohen & Postel
IEN 191 July 1981
Comments on Rosen's Memos
addresses they are physically attached to. Of course this requires
source and destination switches to check messages in a "trusted"
fashion.
There is a small inconsistency in the author's discussion of
source-only vs. intermediate ("tandem") node name lookup. At the top
of page 11, it is stated that the tandem nodes will be "no more
likely" than the source node to have new information during a
transient update period. However, on page 12-13, it is pointed out
(correctly) that tandem nodes likely WILL produce a "better route
selection ... if delay changes or topology changes take place while a
packet is in transit."
There will be substantial modification needed to the host software in
order to implement this scheme. It is proposed (we think) that both
the current scheme and the logical address scheme be available at the
same time. The details of the logical address are not very clear,
but a 16-bit logical address is suggested, which would require a
character string to number lookup in the hosts to make it convenient.
IEN 184
This memo claims that the previous work on the Internet is deficient
due to reliance on an inadequate model of the structure of the
Internet. IEN 184 claims to present a new model of the Internet that
does provide a basis for future work.
The proposed model of internetwork operation views the gateways more
explicitly as switching nodes, with the hosts attached to these
nodes. In particular, each host is multi-hommed on all the gateways
on the same network as the host.
There is some merit to this model and the questions it raises, but
the author is not the first to think of this viewpoint (see for
example IEN-135 [10]). There are also some problems with this model
that the author seems unaware of.
This new model might be acceptable if one wanted to build a super
ARPANET based solely on lines and super-IMPs, but if one is planning
to include other technologies such as broadcast satellite and
broadcast local networks, the proposed model has serious flaws.
For example, two hosts on the same net may still wish to use Internet
protocols to communicate. In the author's model, they would have to
do so by going through an intermediate gateway on their net, since by
definition, no hosts can communicate directly over a "Pathway" with
Sunshine & Cohen & Postel Page [3]
July 1981 IEN 191
Comments on Rosen's Memos
no intervening "Switch." This is clearly inefficient in the intranet
case, and one way in which it differs from the ARPANET. This would
also be true in many single broadcast nets where there are no
intervening switches between hosts even at the single network level
of "Network Structure."
This memo fails to consider the impact on the host systems. Host
will be designed to use a common approach to communication with other
hosts whether they be across the room or across the world. With the
existing model and Internet Protocol, the same procedures and formats
can be used between hosts on the same network and between hosts many
networks apart (though different performance parameters may be
necessary).
The model developed in the Internet Working Group and described by
Cerf (IEN-48 [11]) continues to be the most reasonable basis for
developing the Internet.
IEN 187
This memo assumes the model (of IEN 184) of hosts always sending and
receiving internet traffic via an "Internet Switch". It goes on to
describe the interactions of a host and an internet switch, and then
criticizes the existing Internet Protocol for not being a perfect
host-switch interface protocol.
We cannot possibly take on all of the topics and "lessons" presented,
but Section 2.4 of IEN-187 on fragmentation provides a good example
of what is wrong with these reports. Again, the author seems unaware
of previous important work on this subject, for example IEN-20 by
Shoch (expanded and published in Computer Networks in 1979) [13], or
the paper by Sunshine on interconnection of networks published in
Computer Networks in 1977 [14]. If the author had read these, he
might have avoided several serious deficiencies in his presentation:
1. After a long discussion of the evils of final destination (or
internet) fragmentation, the author reveals his preferred approach
of hop-by-hop (or intranet) fragmentation as if he invented the
idea.
2. There is an important goal that internet fragmentation
supports, but intranet fragmentation does not: independent and
possibly different routing of each fragment through different exit
gateways from a "small packet" net (and subsequently). The author
fails to consider this point.
Page [4] Sunshine & Cohen & Postel
IEN 191 July 1981
Comments on Rosen's Memos
3. In presenting scenarios (page 58) showing the evils of internet
fragmentation, the author omits the important scenario of several
small packet nets in a row, where repeated intranet fragmentation
is just the WRONG thing to do.
4. Packets with the Don't Fragment flag on are not "simply lost in
transit" (page 53) if they cannot be forwarded without
fragmentation. Specific error packets are returned to the source
host, which may try to resend smaller packets.
5. After all his discussion, the author admits in the final
paragraph that destination host fragmentation is necessary anyway
if the final network gets too large a packet. The author claims
this will be necessary only for hosts on nets with "unusually
small" maximum packet sizes, but in fact it will be necessary on
all nets with less than the maximum maximum packet size of any net
in the system if they wish to receive packets from the largest
packet size nets.
The net effect of this sort of incomplete presentation is a step
backward from the current imperfect level of understanding of this
important issue.
The author also attacks the Type of Service (TOS), Time to Live
(TTL), Source Routing (SR), Flow Control (FC), and Fault Isolation
(FI) features of IP and ICMP.
On Type of Service the author tells us for ten pages all the bad
things about the Internet Protocol provision for TOS, while agreeing
it is an important concept, but has nothing different to offer,
except some vague notion that service catagories should correspond
more closely to application types.
On Time to Live the author complains that there is an inconsistency
since the TTL is stated to be in seconds, and that the gateways must
decrement the TTL by one, and that the gateways are expected to
process datagrams faster than one a second. If one assumes that the
intention is to guarantee that datagrams stay alive as long as the
TTL, he is right. But the intention is really to guarantee that they
disappear before TTL. So TTL is an upper bound on how long the
datagram may exist. Most reliable transport protocols assume a
maximum datagram lifetime (sometimes unknowningly) for the correct
operation of their reliability procedures [15].
On Source Routing the author suggests that this feature exists due
only to problems with existing routing procedures and for
Sunshine & Cohen & Postel Page [5]
July 1981 IEN 191
Comments on Rosen's Memos
experiments, and that any really adequate routing procedure in the
gateways will eliminate the need for source routing in normal
operations. We suggest that the Internet will be a much more dynamic
environment than the author has yet imagined and that source routing
will be essential to reach through the Internet to local environments
not fully integrated into the main Internet routing world.
On Flow Control and Fault Isolation the author indicates that the
current mechanisms are inadequate, but does not suggest workable
alternatives. On FC the ICMP "Source Quench" message is cited as a
case of "choke packet" flow control which the author does not believe
in (page 64). Earlier (page 63) the author complains that "source
quench" is only advisory, and later (page 66) the author makes vague
suggestions that a better flow control scheme would use advisory
messages to suggest that datagrams had been discarded (exactly what
source quench does).
All in all this memo comes across as an attack on the Internet
Protocol, with few suggestions for improvement. But it is based on
an assumption: that the Internet Protocol is a host-switch access
protocol. This assumption requires further discussion.
IEN 188
This memo describes logical addressing in the Internet, primarily by
recasting the method of IEN 183 in generalized terms. There are a
number of inaccuracies and omissions in the discussion. One serious
limitation is failure to consider the case of hosts sending Internet
datagrams to each other directly on a single net as discussed above.
On page 4 (middle), the author correctly states that IP addresses are
hierarchical, but incorrectly states that their second component is
necessarily a "physical address." In fact, it may be a name or
"logical address" in networks that provide that capability (but must
be carried in 24 bits).
On page 7, the author proposes using a "unique name which is
meaningful at each level of internet hierarchy." This seems to be a
strong violation of layering, and as the author admits, would require
the switches in every constituent network to "understand" and be able
to lookup the names, probably an intolerable demand on individual
network autonomy.
On page 34, the author's claim that hierarchical addressing requires
less table space than flat addressing is false. His justification is
incomprehensible to us, particularly since he has just finished
Page [6] Sunshine & Cohen & Postel
IEN 191 July 1981
Comments on Rosen's Memos
proposing an "area" addressing scheme similar to hierarchical schemes
in order to reduce table sizes!
In the detailed model of operation given in Section 3.4, an important
step is omitted when the first sentence states, "Let's assume that a
source Host has given a message to a source Switch ..." How does the
source host pick the source switch? In fact, it must pick both a
network level (e.g., IMP) and internet level (gateway) switch,
assuming it is multi-homed, which at least at the internet level is
quite likely. In order to make this selection, the host will have to
have a table giving the best switch (at each level) for each possible
destination name. But these are precisely the sort of tables the
author's scheme is meant to avoid having in the hosts. In light of
the comment above about hosts talking to each other directly on the
same net, the hosts must at least know the names and addresses of
every other host on their own net.
The treatment of mobile hosts is quite brief and offers no
improvement over previously proposed solutions.
IEN 189
This memo discusses routing in the Internet, and proposes that the
existing gateway routing procedure be replaced by the SFP procedure
now used in the ARPANET. This is surely a useful suggestion. The
note does however raise a number of issues in its examples of routing
problems that indicate an incomplete understanding of the whole area.
The note proposes a "gateway discovery protocol" that could be
provided by individual nets. This idea seems worthwhile, although it
is not clear how many individual nets would be willing to make such
additions. We should like to point out that it is also possible to
perform this function directly among gateways in networks which
support broadcast or group addressing.
The discussion of routing alternatives makes generally sound if
qualitative conclusions, but a few details are confused. The
discussion of throughput performance on page 41 assumes TCP will
operate with a small enough window over a high delay path so that
throughput is reduced, but this is precisely the situation in which
proper "tuning" requires a large window, which would allow high
throughput.
The analogy with "whole picture" algorithms on pages 44-45 fails to
mention that in the whole picture scenario, each person would have to
get a piece of paper 100 times bigger than with the local scheme, and
Sunshine & Cohen & Postel Page [7]
July 1981 IEN 191
Comments on Rosen's Memos
hence this approach has an information distribution requirement that
is much higher.
This memo contains several informal citations that could be usefully
spelled out for the IEN audience. The author mentions algorithms by
Gallager (page 17), Dijkstra (page 20), and Floyd (page 20), all
without references. It is safe to say that any list of references
containing only the author and his coworkers (as consistently done in
this series) cannot be adequate.
One particular example provokes the following response:
Please replace the second paragraph of page 49 of IEN-189 with the
following paragraph:
"In fact the situation could be even worse. If Switches in
Boston know nothing about what happening inside the building on
4676 Admiralty Way then data for the North section of the 11th
floor which arrives at the South section of the 11th floor is
then sent back from the South section to Boston for alternate
routing will just loop back to the South section. The data
will be stuck in an infinite loop, never reaching its
destination. In IEN 179 [12] Danny Cohen proposed a regional
scheme like this, apparently not realizing that it suffers from
loops. His proposal also includes a form of hierarchical
addressing which is closely bound up with routing, so that a
Switch is Boston might not even be able to distinguish data for
the South section from data for the North section. That is, in
Cohen's scheme, data for the South section and data for the
North section would be indistinguishable at the Boston
Switches; all such data would appear to be addressed to the
South section. Only the Switches at the South section would
look further down the address hierarchy to determine whether
the data needs further forwarding to the North section. Any
such scheme is hopelessly loop-prone, except in a Network
Structure whose connectivity is extraordinarily rich, much more
so than the Catenet's will ever be."
Since the above suggestion was merely to follow the routing
strategy used by the phone companies, TELENET and others, you
should warn them immediately about this hopelessly loop-prone
situation.
I believe that if the Boston Switch has ALL the information about
EVERYthing, EVERYwhere it would be in position to make better
decisions, ALWAYS, especially if that information is updated with
Page [8] Sunshine & Cohen & Postel
IEN 191 July 1981
Comments on Rosen's Memos
absolutely ZERO time delay. If this information is absolutely
free (in terms of communication, storage and processing) it may be
dumb not to make every Switch always know everything about
everything, down to (or "up to"?) the finest granularity
(location? site? process? file? register? bit?). However, if this
is not absolutely free, some compromises may have to take place.
Oh, one point which I did not quite follow: why if the
Nevada/California lines are broken forever, Boston is never told
about it - as described by you? By the way, what made you
understand that the "The Switch at Nevada would look further down
the address hierarchy to determine whether the data needs further
forwarding to California" ?
I highly recommend that you get hold of any telephone directory
and read the area-codes tables. This may help you understanding
that the California area codes are neither above, nor below, nor
further on any hierarchy than the Nevada ones, and vice versa.
This is a very subtle point which may escape the casual reader.
Mastering this idea may help you understand what IEN-179 is all
about. In short, IEN-179 is not an attempt to describe the ideas
which you have in mind by using the telephone scenario, but an
attempt (which obviously failed, at least in your case) to
introduced old well-proven ideas from other communication arenas
into ours.
SUMMARY
In summary we are glad to have this information and these opinions
presented for discussion in the Internet Working Group, and we hope
that others will speak up with their opinions too. We are concerned
that too many will be so overwhelmed by the wide ranging arguments to
notice that some important considerations were not mentioned.
We especially want to make clear that a fundamentally different model
of the Internet architecture is proposed by Rosen, and that we have
serious reservations about aspects of that model.
Sunshine & Cohen & Postel Page [9]
July 1981 IEN 191
Comments on Rosen's Memos
REFERENCES
[1] Rosen, E., "Issues in Buffer Management", IEN 182, Bolt Beranek
and Newman, May 1981.
[2] Rosen, E., "Logical Addressing", IEN 183, Bolt Beranek and
Newman, May 1981.
[3] Rosen, E., "Issues in Internetting Part 1: Modelling the
Internet", IEN 184, Bolt Beranek and Newman, May 1981.
[4] Rosen, E., "Issues in Internetting Part 2: Accessing the
Internet", IEN 187, Bolt Beranek and Newman, June 1981.
[5] Rosen, E., "Issues in Internetting Part 3: Addressing", IEN 188,
Bolt Beranek and Newman, June 1981.
[6] Rosen, E., "Issues in Internetting Part 4: Routing", IEN 189,
Bolt Beranek and Newman, June 1981.
[7] Clark, D., "A Proposal for Addressing and Routing in the
Internet", IEN 46, MIT/Laboratory for Computer Science, June
1978.
[8] Cerf, V., "Internet Addressing and Naming in a Tactical
Environment", IEN 110, Information Processing Techniques Office,
Defense Advanced Research Projects Agency, August 1979.
[9] Shoch, J., "Inter-Network Naming, Addressing, and Routing",
Proceedings 17th IEEE Computer Society International Conference,
pp72-79, September 1978.
[10] Sunshine, C., "Addressing Mobile Hosts in the ARPA Internet
Environment", IEN 135, USC/Information Sciences Institute, March
1980.
[11] Cerf, V., "The Catenet Model for Internetworking", IEN 48,
Information Processing Techniques Office, Defense Advanced
Research Projects Agency, July 1978.
[12] Cohen, D., "Addressing and Routing", IEN 179, USC/Information
Sciences Institute, March 1981.
[13] Shoch, J., "Packet Fragmentation in Inter-Network Protocols",
Computer Networks, V.3, N.1, pp3-8, February 1979.
Page [10] Sunshine & Cohen & Postel
IEN 191 July 1981
Comments on Rosen's Memos
[14] Sunshine, C., "Interconnection of Computer Networks", Computer
Networks, V.1, N.3, pp175-195, January 1977.
[15] Watson, R., "Timer-Based Mechanisms in Reliable Transport
Protocol Connection Management", Computer Networks, V.5, N.1,
pp47-56, February 1981.
Sunshine & Cohen & Postel Page [11]