[Discuss] Next Generation Networks

[J. C. Jones]

Hello!

It is pleasant to see the New Year start with a new
network architecture proposal.  I am going to offer a
critique of Dr. Barroso's paper.  I acknowledge that
he did not post it to GENI, but since he undoubtedly
hopes it will be successful as a grand-scale Internet
architecture, I thought I would comment anyway.  Note
that my critique is, indeed, critical.  The goal is no
such much to trample the ideas of others, but to
illuminate quickly (what I perceived to be) flaws in
the design so that those who might be undecided can
quickly determine its relative virtue for themselves. 
Naturally, these are only my opinions.

With that said, I will offer a brief summary of his
paper for the benefits of others.

Summary of Proposal:
http://www.lmdata.es/uets/eChallenges-e2006-uets-paper.pdf

Dr. Barroso is essentially proposing to bypass Layer 3
IP addresses and use Layer-2 6-byte Ethernet addresses
directly to address devices (nodes) on a global scale.
 There are two bits in a first byte of each of the two
addresses (source and destination) in an Ethernet
frame that are reserved for "end user definition". 
Dr. Barroso is proposing that we take advantage of
these bits to define a new network addressing scheme.
Once one of these bits is set to 1, not counting the
other user-definable bit, the remaining 46 bits are
available for end-user definition.  He would like to
partition these bits as [world zone | country | ISP |
terminal].  The Ethernet hardware would inspect the
addresses to do incremental routing on the network. 
Ethernet "access points" would control the allocation
of addresses to attached devices.  The logical link
control (LLC) portion of Ethernet would be employed to
achieve inter-interface reliability of frames via
retransmission.  This would supplant the reliability
mechanism of TCP so that applications would
essentially "send and know that transmission was
successful."  He claims that the scalability and
security problems with Ethernet would be eliminated.
He claims that performance would become more
deterministic due to absence of queuing, and that
power consumption would be reduced as all these
devices would receive power over Ethernet (POE).

CRITIQUE:

"The UETS Communications Architecture and Reference
Model [1, 2] is the natural
evolution in Computers and Network convergence:"

Not true (yet). It is not a natural evolution until it
evolves naturally.

"The model has the ability to provide the same
services using a layer-2 hardware-based
operation device, breaking the limits of hosts
collapsed by very-high speed TCP/IP
connections [3]. The Logical Link Control (LLC) is
better than the TCP/UDP for offering
end-to-end services, because it is optimized to
hardware operation, and it also has a reduced
overhead and a tighter loop control."

This is not true.  When two applications are at either
end of a network, the end-to-end principle is not only
a nice-to-have, it is a necessity.  LLC cannot
guarantee successful end-to-end transmission of a
packet - think what would happen if a fiber cable were
cut.  Because the Ethernet hardware transmitting the
frames do not communicate with the end applications,
the applications would not know if the cable had been
cut, and so would still have to do end-based state
determinism.  It _is_ a good idea for each link to
make best effort transmission with minimized BER,
perhaps using FEC or even a retransmission after a
collision, but beyond that, end to end is a necessity.
 Note that there is no such thing as two applications
being "connected" anyway.  The notion of
"connectivity", unfortunately, is a misconception that
has been allowed to pervade the network architecture
community.  There is no such thing - it is essentially
the Three-Army Problem.

"The fixed and mobile terminals can be discriminated
by one bit, as described
in Figure 3. If the terminal is fixed, the address is
used directly to perform the switching, if
mobile, the address should be associated with a fixed
one by the Base Station."

Think about the mobility problem, for example, and it
will become apparent why this scheme will not scale,
or at very minimum, will be horribly inefficient.  If
I, and a friend of mine who lives in Austin, Texas,
USA, are far away from home, perhaps at a conference
in Barcelona, Spain; what addresses will be assigned
to our two computers?  If we use the "old" addresses
bound to the Base Station in the Austin, Texas, USA,
then even though were are in the same room in
Barcelona, the packets will be routed through a
machine in Austin.  If we get "new" addresses, each
time, there will be an identity problem.  A node will
never know if a packet was intended for it, or for the
device that acquired the same address 5 minutes
earlier.  Also, during acquisition of an address from
a Base Station, there will be, again, an identity
problem - the node requesting a new address will
attempt to communicate using and address that it
already has, which, as Dr. Barroso stated, would not
be valid, as it would not have been allocated by the
new Base Station.  Even with a scheme that I could
imagine which would be the only scheme that could
circumvent this problem, all Base Stations and mobile
nodes would have operate in promiscuous mode until
they are able to ascertain their link-local addresses.
This architecture will not only hinder a solution to
the mobility problem, it will render it unsolvable. 

"With an efficient structure, only one UETS domain
provide for very big number of
connections. "

Dr. Barraso allowed 5 bits to identify the country
within a world zone for his addressing scheme. 
2^5=32.  Africa is a world zone. Africa has more than
50 countries.

"
France - 547,030 x 59,765,983 = 32,693,785,680,490
Spain - 504,782 x 40,077,100 = 20,230,198,692,200
Italy - 301,230 x 57,715,625 = 17,385,677,718,750
U.K. - 244,820 x 59,778,002 = 14,634,850,449,640"

It is not clear to me why the land area of a country
is being multiplied by the number of address states
that would be allocated to that country.

"POWER OVER ETHERNET FOR EMERGENCY CALLS"

If it were known how insignificant the power
consumption of an Ethernet interface is compared to
overall consumption (and waste) of the devices in
which they are embedded, I doubt that one would spend
much time thinking about this.

"
- ESCALABILITY
- HIGHER SPEEDS
LINK CONTROL (HDLC/LLC)
- FLOW
- CONGESTION"

Congestion cannot be managed without end-to-end
control.  Some might claim that this is not true -
that a network could simply "reject" inrush at the
edges, but the only way to know whether there is
inrush or not is to have end-to-end communication,
where the "ends" can be in the "middle" also.  

ON THE MATTER OF SECURITY:

The security problem goes far beyond link-layer
address spoofing.  There is privacy, authenticity,
replay, etc.  Certainly one would _not_ try to
promulgate a new system without having figured out all
these security issues simultaneously.

ON THE MATTER OF RETRANSMISSION:

As stated earlier, there will always have to be end to
end state determination as long as there does not
exist the facility of real-time, hardware-based,
detection, indication, and propagation of
link-failure.  It should be intuitively obvious that
this will never happen.  Wireless nodes alone attest
to this fact.

ON THE MATTER OF USING ETHERNET ADDRESSES: 

Some people reading the proposal might have become
intrigued by the use of Ethernet addresses.  After
all, aside from the paucity of state allocated to
Africa and other minor anomalies, it seems, at least
partially reasonable that the addressing scheme is not
*that* bad.  If so, may I suggest that, what you find
intriguing is not so much that Ethernet addresses are
being used, but that multi-level scoping is in the
addresses.  Note that if we were to declare that all
Ethernet hardware were to be redesigned with 128-bit
addresses, with scoping, our proposal would be just as
intriguing.  In other words, there is nothing special
about using Ethernet hardware.

However, using Dr. Barroso's scheme, all Ethernet
hardware _would_ have to be redesigned.  Currently, an
Ethernet interface does not examine the destination
address of a frame before transmitting that frame.  In
UETS, this would be necessary, whether done "in
software" or "in hardware" - the interface will have
to know that the address in the destination address
field identifies a machine that is perhaps 5,000
kilometers distant.  Also, the source address in the
frame would have to change as the frame traversed each
link due to the anti-spoof argument - if it did not,
the receiving Interface would think a node that does
not have the right to be attached to the medium is
trying to send a spoof frame.

If the author were successful in his proposal, he
would discover that the trillions of addresses
"allocated to Spain", would not be "enough".  The
enough here refers not to total enumeration, but to
space partitioning.  I offer no discussion to support
this assertion. :)

MY ANALYSIS:

The author of the proposal attempts to reintroduce the
notion that intelligence should lie in the core of the
network.  The end-to-end principle has not been
appreciated as a necessity.  No real attention as been
giving to the security, mobility, and multicasting
problems.  No real attention has been given to the
naming problem.  False reasoning was made about the
reliable transmission problem, namely that the core of
the network would be able to supplant end-to-end state
determination.  The author underestimates the cost of
having hardware designers redesign new Ethernet
hardware to accommodate the proposed protocol.  The
author has not considered well the address space
partitioning problem, where, even though 1,000,000,000
 bit patterns might be available for a given entity,
the utilization of those 1,000,000,000 addresses by
sub-entities within the entity will not only be
suboptimal, but "horrifically poor", by necessity of
recursive scoping.

The only thing virtuous about this proposal, in my
opinion, is that it hints at the notion that scoped
addresses is a good idea, something that we already
know.  Apart from that, tt mostly represents a
reversion to ideas that have long since been regarded
as lacking in virtue when compared to the end-to-end
principle.

Regards,

-J.C. Jones-
Electrical Engineer
Austin, Texas, USA


 
____________________________________________________________________________________
Be a PS3 game guru.
Get your game face on with the latest PS3 news and previews at Yahoo! Games.
http://videogames.yahoo.com/platform?platform=120121