Content uploaded by John D Day
Author content
All content in this area was uploaded by John D Day on Jan 05, 2016
Content may be subject to copyright.
© John Day, 2011
Rights Reserved
How in the Heck
Do You Lose a Layer!?
Future Network Architectures Workshop
University of Kaiserslautern, Germany
John Day
2012
OSI only had a Network Layer, but
the Internet has an Internet Layer!
- Noel Chiappa, 1999
© John Day, 2011
Rights Reserved
Preamble
• A Couple of Remarks on the Nature of Layering
• The advent of packet switching required much more
complex software than heretofore, and so the concept of
layers was brought in from operating systems.
• In operating systems, layers are a convenience, one design
choice.
• In networks, they are a necessity.
© John Day, 2011
Rights Reserved
The (really) Important Thing about Layers
(From first lecture of my intro to networks course)
© John Day, All Rights Reserved, 2009
Physical
Data Link
Network
Transport
Application
Host or End System
Router
Less
Scope
More
Scope
•Networks have loci of distributed shared state with different scopes
•At the very least, differences of scope require different layers.
•It is this property that makes the earlier telephony or datacomm
“beads-on-a-string” model untenable.
–Or any other proposal that does not accommodate scope.
•If there are multiple layers of the same scope, their functionality must
be independent.
–This has always been understood.
© John Day, 2011
Rights Reserved
1972 Was an Interesting Year
• Tinker AFB joined the ‘Net exposing the multihoming problem.
• The ARPANET had its coming out at ICCC ‘72.
• As fallout from ICCC 72, the research networks decided it would be
good to form an International Network Working Group.
– ARPANET, NPL, CYCLADES, and other researchers
– Chartered as IFIP WG6.1 very soon after
• Major project was an International Transport Protocol.
– Also a virtual terminal protocol
– And work on formal description techniques
• But a major 3-sided war was just breaking out
© John Day, 2011
Rights Reserved
War on All Sides
• The Phone Companies don’t like the new model because an end-to-end
transport relegates them to a commodity business, thus not giving them
exclusive claim to “value-added services” or “services in the network.”
• IBM (with 80% of the computer market) doesn’t like the new model
because it has a hierarchical network architecture, SNA.
• The other computer companies (especially the minicomputer) love the
new model because it plays to their strength and because it nails down
the other two!
• This War (and it was) dominates everything for the next 2 decades and
really is still going on.
© John Day, 2011
Rights Reserved
Meanwhile Back at INWG
There Were Two Proposals
• INWG 37 based on the early TCP and
• INWG 61 based on CYCLADES TS.
• And a healthy debate, see Alex McKenzie, “INWG and the Conception of the
Internet: An Eyewitness Account” IEEE Annals of the History of Computing, 2011.
• Two sticking points
– How fragmentation should work
– Whether the data flow was an undifferentiated stream or maintained the integrity of
the units sent (letter).
• These were not major differences.
© John Day, 2011
Rights Reserved
After a Hot Debate
• A Synthesis was proposed: INWG 96
• There was a vote in 1976, which approved INWG 96.
• As Alex says, NPL and CYCLADES immediately said they would
convert to INWG 96; EIN said it would deploy only INWG 96.
• “But we were all shocked and amazed when Bob Kahn announced that DARPA researchers were too
close to completing implementation of the updated INWG 39 protocol to incur the expense of
switching to another design. As events proved, Kahn was wrong (or had other motives); the final
TCP/IP specification was written in 1980 after at least four revisions.”
– Neither was right. The real breakthrough came two years later.
• But the differences weren’t the most interesting thing about this work.
© John Day, 2011
Rights Reserved
The Similarity Among all 3
Is Much More Interesting
• This is before IP was separated from TCP.
• All 3 Transport Protocols carried addresses.
• This means that the Architecture that INWG was working to was:
• Three Layers of Different Scope each with Addresses.
• If this does not hit you like a ton of bricks, you haven’t been paying
attention.
• This is NOT the architecture we have.
Internetwork Transport Layer
Network Layer
Data Link
Layer
TCP
IP
MAC
LLC
SNDC
SNAC
© John Day, 2011
Rights Reserved
INWG’s Internet Model
• An Internet Layer addressed Hosts and Internet Gateways.
• Several Network Layers of different scope, possibly different
technology, addressing hosts on that network and that network’s routers
and its gateways.
– Inter-domain routing at the Internet Layer; Intra-Domain routing at the
Network Layer.
• Data Link Layer smallest scope with addresses for the devices (hosts or
routers) on segment it connects
Internet Gateways
Data Link
Network
Internet
Application
Data Link
Network
Internet
Application
Net 1 Net 2 Net 3
Host Host
© John Day, 2011
Rights Reserved
Separating IP from TCP
• Shouldn’t have changed things much. But it did.
• They thought they needed a flag day to transition from NCP to TCP/
IP and when they were done, it looked like this:
Internetwork Layer
Network Layer
Data Link Layer
Transport Layer
Network Layer
Data Link Layer
Transport Layer
© John Day, 2011
Rights Reserved
How Did They Lose A Layer?
• What Appears to Have Happened:
– By 1980 or so: A Large Central ARPANET surrounded by LANs
• ARPANET was a black box run by BBN with its own routing and congestion control.
– Routers start appearing with Ethernet on one side and ARPANET on the other
• Called Internet Gateways at the time.
– Some are directly connected, with Static Routing or Simple Routing schemes
– As the ‘Net expands and the ARPANET shrinks . . .
• More routers are directly connected but still all under DoD, so not seen as separate
networks with distinct routing domains. And not big enough that it is a problem.
– The Non-ARPANET network is routing on IP as a Network Layer protocol!
• Little overlap between the INWG and the Internet, by now IP is the Network Layer!
– No one notices that there needs to be a network layer and an internetwork layer.
And anyway functions don’t repeat in layers.
• But Two Layers of Different Scope is Not a Repetition
• Scope? What is that?
– The Same Protocol can used to Provide Functions of Different Scope.
• To refine Dykstra, Functions shouldn’t repeat in layers of the same scope
© John Day, 2011
Rights Reserved
Okay, What Did OSI Do?
• The other major effort at the time.
• The well-known 7-layer model was adopted at the first meeting in
March 1978 and frozen. After that, they had to work within that.
• One of the concerns in the Network Layer deliberations was
interworking over a less capable network:
• Would need to enhance the less capable network with an additional
protocol
high-quality
network
low-quality
network
high-quality
network
high-quality
network
low-quality
network
high-quality
network
© John Day, 2011
Rights Reserved
OSI Sub-Divided the Network Layer
• This concern and the recognition that there would be different
networks interworking lead OSI to posit three sublayers, which might
be optional depending on configuration:
Subnet Independent
Convergence (SNIC)
Subnet Dependent
Convergence (SNDC)
Subnet Access (SNAC)
© John Day, 2011
Rights Reserved
OSI Also Came to the INWG Model
• With a Transport Layer, this is the same as the INWG model.
• OSI stayed the course with a similar split to TCP/IP.
• So OSI had an Internet Architecture and the Internet has a Network
Architecture.
• And signs of a repeating structure.
Transport Layer
Subnet Independent
Subnet Dependent
Subnet Access
Data Link LLC
Data Link MAC
© John Day, 2011
Rights Reserved
Was Splitting TCP and IP Right?
• The Rules say if there are two layers of the same scope, the functions
must be independent for it to work well.
• Error and Flow Control separated from Relaying and Multiplexing are
independent. But IP also handles fragmentation across networks.
– Remember, Don’t repeat functions in different layers of the same scope?
• Problem: IP fragmentation doesn’t work.
– IP has to receive all of the fragments of the same packet to reassemble.
– Retransmissions by TCP are distinct and not recognized by IP.
• Must be held for MPL (5 secs).
• There can be considerable buffer space occupied.
• There is a fix:
– The equivalent of “doc, it hurts when I do this!” “Then don’t do it.”
MTU Discovery.
© John Day, 2011
Rights Reserved
But it is the Nature of the Problem
that is Interesting
• The problem arises because there is a dependency between IP and
TCP. The rule is broken.
– It tries to make it a beads on a string solution.
• A Careful Analysis of this Class of Protocols shows that the Functions
naturally cleave along lines of Control and Data.
• TCP was split in the Wrong Direction!
• It is one layer, not two.
– IP was a bad idea.
Relaying/ Muxing
Data
Transfer
Data Transfer
Control
Delimiting
Seq Frag/Reassemb
SDU Protection
Retransmission and
Flow Control
© John Day, 2011
Rights Reserved
A Chance to Get Things on Track
• We knew in 1972, that we needed Application Names and some kind
of Directory.
• Downloading the Host file from the NIC was clearly temporary.
– And eventually we would have many more applications than the basic 3.
• When the time came to automate it, it would be a good time to
introduce Application Names!
• Nope, Just Automate the Host file. Big step backwards with DNS.
• Now we have domain names
– Macros for IP addresses
• And URLs
– Macros for jump points in low memory
© John Day, 2011
Rights Reserved
Then in ‘86: Congestion Collapse
• Caught Flat-footed. Why? Everyone knew about this?
– Had been investigated for 15 years at that point
• With a Network Architecture they put it in Transport.
– Worst place.
• Most important property of any congestion control scheme is
minimizing time to notify. Internet maximizes it and its variance.
• And implicit detection makes it predatory.
– Virtually impossible to fix
• Whereas,
© John Day, 2011
Rights Reserved
Congestion Control in an Internet is
Clearly a Network Problem
• With an Internet Architecture, it clearly goes in the Network Layer
– Which was what everyone else had done.
• Time to Notify can be bounded and with less variance.
• Explicit Congestion Detection confines its effects to a specific
network and to a specific layer.
Internet Gateways
Data Link
Network
Internet
Application
Data Link
Network
Internet
Application
Net 1 Net 2 Net 3
Host Host
© John Day, 2011
Rights Reserved
Would be Nice to Manage the Network
• With a choice between a modern object-oriented protocol (HEMS) and
a traditional approach (SNMP), err sorry . . . “a simple approach.”
IETF goes with “simple.”
– Must be simple has the Largest implementation of the 3:
• SNMP, HEMS, CMIP.
– So simple too complex to use
• IEEE had tried the SNMP approach several years earlier so the shortcomings
were well-known.
• Everything connectionless making it impossible to snapshot tables
• No attempt at commonality across MIBs.
• Router vendors played them for suckers and they fell for it.
– Not secure, can’t use for configuration.
© John Day, 2011
Rights Reserved
IPv6 Still Names the Interface?
Why on Earth?
• Known about this problem since 1972
– No Multihoming, kludged mobility
– Notice in an Internet Architecture this is straightforward.
– Signs the Internet crowd didn’t understand the Tinker AFB lesson.
– DARPA never made them fix it.
• By the Time of IPng, Tradition trumps Everything.
• IPv7 would have fixed it.
– But that fight was too intense. This is no longer science, let alone
engineering.
• When can’t ignore, and given post-IPng trauma they look for a
workaround.
Violá!
Loc/Id Split!
© John Day, 2011
Rights Reserved
Loc/ID Split
(these are people who
lost a layer to begin with, right?)
• You’ve got to be kidding?! Right!
• First off:
– Saltzer [1977] defines “resolve” as in “resolving a name” as “to locate an
object in a particular context, given its name.”
– All names in computing locate something.
• So either nothing can be identified without locating it, nor located without
identifying it, OR
• anything that doesn’t locate something is being used outside its context
• Hence it is either a false distinction or it is meaningless.
• Second, one must route to the end of the path.
– The locator is on the path to the end, it isn’t the end.
– The “identifier” locates the end of the path but they aren’t routing on it.
– No wonder it doesn’t scale
• There is no workaround. IP is fundamentally flawed.
© John Day, 2011
Rights Reserved
Never Get a Busy Signal on the Internet!
• Golly Gee Whiz! What a Surprise!!
• With Plenty of Memory in NICs, Getting huge amounts of buffer
space backing up behind flow control.
• If peer flow control in the protocol, pretty obvious one needs interface
flow control as well.
• Well, Duh! What did you think was going to happen?
– If you push back, it has to go somewhere!
– Now you can have local congestion collapse!
Flow Control
2010 They Discovered Buffer Bloat!
No Interface
Flow Control
© John Day, 2011
Rights Reserved
Taking Stock
• The Internet has:
– Botched the protocol design
– Botched the architecture
– Botched the naming and addressing
– When they had an opportunity move in the right direction with application
names, they didn’t. They did DNS.
– When they had an opportunity to move in the right direction with node
addresses, they didn’t. They did IPv6.
– More than Botched Network Management
– Botched the Congestion Control twice
– Once so bad it probably can not be fixed.
• By my count this makes them 0 for 8!
• It defies reason! Do these guys have an anti-Midas touch or wha!?
© John Day, 2011
Rights Reserved
But It is a Triumph!
• But It Works!
• So did DOS. Still does.
• With enough Thrust even Pigs Can Fly!
• As long as fiber and Moore’s Law stayed ahead of Internet
Growth, there was no need to confront the mistakes.
• Now it is catching up to us and can’t be fixed.
– Fundamentally flawed, a dead end.
– Any further effort based on IP is a waste of time and effort.
• Throwing good money after bad
– Every patch (and that is all we see) is taking us further from where we
need to be.
(By that argument, so was DOS)
© John Day, 2011
Rights Reserved
INWG Was on The Right Track!!
• They were Close to Seeing it was a Repeating Structure
– A Structure we arrived at independently by a similar approach.
– Is RINA the answer? Who knows? We are doing the science and
letting the chips fall where they may.
– Can you propose something that is simpler and answers more
questions, makes predictions about things we haven’t seen?
– We can and have. . . .
Internetwork Transport Layer
Network Layer
Data Link
Layer
TCP
IP
MAC
LLC
SNDC
SNAC
© John Day, 2011
Rights Reserved
How Lucky Can You Get!?
• Now that We are Back on Track, There is so much to explore!
– It is Interesting How Different the Fundamentals Are
• By Maximizing Invariances and Minimizing Discontinuities:
– To scale, resource allocation problems require a repeating (recursive)
structure. (confirmed by Herb Simon’s Science of the Artificial)
– A Layer is a Distributed Application for Managing Interprocess
Communication.
– Watson’s results are fundamental: Bounding 3 Timers are the Necessary
and Sufficient conditions for synchronization.
– Besides a simpler protocol
• Implies decoupling port allocation and synchronization
• Implies that for it to be networking (IPC) MPL must be bounded.
• Has important implications for security
© John Day, 2011
Rights Reserved
Really Lucky!
• So Much to Explore!
– Multiple layers of the same rank implies a means to choose which one to use.
• Neither a global address space nor a global name space are absolute requirements
– The nature of security is much clearer, simpler, and more robust.
– Many new avenues to explore in congestion control and quality of service.
– And the adoption can be seamless.
– And there are implications beyond IPC to distributed systems in general that
we are just beginning to understand:
• Hint: Distributed Applications are Local Computations.
• Is there more to Discover?
• Undoubtedly!
