Lecture 2: The Internetworking Problem

gazecummingΔίκτυα και Επικοινωνίες

26 Οκτ 2013 (πριν από 3 χρόνια και 9 μήνες)

89 εμφανίσεις

Lecture 2: The Internetworking Problem.


Problem. How to interconnect
different networks
together? Look into IP, TCP, end
-
to
-
end issues.


The Internetworking Problem. Many different
networks.


Link heterogeneity issue. Single
-
link LANs do not
scale well o
ver hundreds of miles distances. LAN
switching for unknown destination implies
broadcasting, a poor scheme

in large nets.


Link technology: LANs, point
-
to
-
point, Packet radio,
wireless, satellites. An extremely heterogeneous
community.


Challenge: Design a

scalable network infrastructure
that connects different types of smaller networks
enabling packet flows between hosts across different
nets.


Issues to reckon:


■ Addressing: Each net may have its own addressing
approach. Ethernet uses 6
-
byte ID, Telephon
e uses
10
-
digit numbers.


■ Bandwidth and Latency: Link bandwi
dth would
vary from net to net from bits/sec to Gb/sec.
Latencies can be from μs to several secs.


■ Packet size: In general varies between networks

■ Loss rates & QoS: All networks vary conside
rably
in these. These depend on traffic volume, congestion,
service demands, …

■ Packet routing: Routing of packets may be handled
differently by different nets.


The original TCP/IP (Cerf & Kahn).


■ Key idea. Gateways. Reject translation in favor of
gat
eways.


■ IP over everything.

■ Standard packet header format

■ Gateways perform fragmentation, reassembly at
end
-
hosts

■ TCP: Process
-
level communication via TCP, a
reliable, in
-
order, byte
-
stream protocol
. Original TCP
did reassembly; now it’s done at t
he receiver. TCP
and IP were tightly interwined originally, now they
are separate.

■ How to launch/sustain/deliver multiple TCP
between two processes? Original design didn’t
consider this.

1.

Use same IP packet for multiple TCP
segments?

2.

Use different IP pa
ckets for different
segments?


■ 2 is simpler than 1. It’s easier to demux and out
-
of
-
order packets are difficult to handle in 1. Demux
key is the port number of the receiving process.

■ Sliding window, cumulative ACK
-
based ARQ
protocol.

■ Window
-
based fl
ow control

■ Heavily dependent on connection setup and release
scheme.

■Way off the mark on accounting issues

■One line summary




Gateways:



Primarily a protocol converter.
It interfaces distinct
networks with different chara
cteristics.


ISDN

LAN

GW

Two types: Translation GW, and Unified Network
Layer.


Translation
-
based Gateways: translates packet
headers and proto
col from one network to another
seamlessly.

Problems:


a.

Feature deficiency
: Not all features are
“preserved” as one moves fr
om one net to
another. Not all features are supported by the
translation.

b.

Poor scalability
: Translation =

where
is
the number of networks connected. Untenable at
large networks


Internet Design principles
:


Id
entify key parameters (a small number) which must
be preserved through translation. All hosts and
networks must implement them.

Check the seminal
paper: “End
-
to
-
end arguments”
http://www.reed.com/Pa
pers/EndtoEnd.html



Universality principles
:


U1.
IP
-
over
-
everything
. A standard network
protocol IP would be used as Internet Protocol. Must
have a standard well
-
defined addressing convention.
Then scaling is not an issue.


U2.
Best
-
effort service model
. All packets are
treated alike, if any gets lost address this at the end,
not inside the network to recover it. (Increases
scalbility and throughput)

U3.
End
-
to
-
end arguments
. Keeps the interior of the
network as simple as possible. Move all controls for
reliability, congestion control etc. at the end
-
hosts.
TCP functionality removed from Gateways.


Robustness principle:


R1.
Soft
-
state inside the network
. Refresh soft
-
states such as Router tables as often as possible.
Thus, even if it gets lost once in a

while, it’s no big
deal.


R2.
Fate sharing
.
Even if a GW crashes, the end
-
to
-
end communication must continue. The GW state is a
soft
-
state, and it shouldn’t share its fate with end
-
hosts.


R3.
Conservative
-
transmission/Liberal reception
.
Be conservative
in what you send, be liberal in what
you receive. What does a TCP sender do when it sees
an ACK for a message it didn’t expect?




Weaknesses


Internet architecture
has its own weaknesses.


■ Architecture relies heavily on the trustworthiness
of end
-
syste
ms. How does one build a trusted
networks when hosts (or a subset of them) couldn’t
be trusted?


Greedy sources aren’t handled well. TCP does a
good job in sharing its bandwidth, but more and more
traffic isn’t TCP

■ Security issues have been neglected


Administration and management tools are not all
that matured

■ Incremental deployment. Not a fundamental
problem, but ought to be recognized by every
protocol designer.












http://www.csee.umbc.edu/~phatak/publications/trans
port
-
striping+mobility.pdf