Tutorial on Network Layers 2 and 3

Radia Perlman Intel Labs ([email protected])

1

Why?
Demystify this portion of networking, so people dont drown in the alphabet soup Think about these things critically N-party protocols are the most interesting Lots of issues are common to other layers You cant design layer n without understanding layers n-1 and n+1
2

What can we do in 1 hours?

Understand the concepts Understand various approaches, and tradeoffs, and where to go to learn more A little of the history: without this, its hard to really grok why things are the way they are

Outline
layer 2 issues: addresses, multiplexing, bridges, spanning tree algorithm layer 3: addresses, neighbor discovery, connectionless vs connection-oriented
Routing protocols
Distance vector Link state Path vector
4

Why this whole layer 2/3 thing?

Myth: bridges/switches simpler devices, designed before routers OSI Layers
1: physical

Why this whole layer 2/3 thing?

Myth: bridges/switches simpler devices, designed before routers OSI Layers
1: physical 2: data link (nbr-nbr, e.g., Ethernet)

Why this whole layer 2/3 thing?

Myth: bridges/switches simpler devices, designed before routers OSI Layers
1: physical 2: data link (nbr-nbr, e.g., Ethernet) 3: network (create entire path, e.g., IP)

Why this whole layer 2/3 thing?

Myth: bridges/switches simpler devices, designed before routers OSI Layers
1: physical 2: data link (nbr-nbr, e.g., Ethernet) 3: network (create entire path, e.g., IP) 4 end-to-end (e.g., TCP, UDP)
8

Why this whole layer 2/3 thing?

Myth: bridges/switches simpler devices, designed before routers OSI Layers
1: physical 2: data link (nbr-nbr, e.g., Ethernet) 3: network (create entire path, e.g., IP) 4 end-to-end (e.g., TCP, UDP) 5 and above: boring
9

Definitions
Repeater: layer 1 relay

10

Definitions
Repeater: layer 1 relay Bridge: layer 2 relay

11

Definitions
Repeater: layer 1 relay Bridge: layer 2 relay Router: layer 3 relay

12

Definitions
Repeater: layer 1 relay Bridge: layer 2 relay Router: layer 3 relay OK: What is layer 2 vs layer 3?

13

Definitions
Repeater: layer 1 relay Bridge: layer 2 relay Router: layer 3 relay OK: What is layer 2 vs layer 3?
The right definition: layer 2 is neighborneighbor. Relays should only be in layer 3!

14

Definitions
Repeater: layer 1 relay Bridge: layer 2 relay Router: layer 3 relay OK: What is layer 2 vs layer 3? True definition of a layer n protocol: Anything designed by a committee whose charter is to design a layer n protocol
15

Layer 3 (e.g., IPv4, IPv6, DECnet, Appletalk, IPX, etc.)

Put source, destination, hop count on packet Then along came the EtherNET
rethink routing algorithm a bit, but its a link not a NET!

The world got confused. Built on layer 2 I tried to argue: But you might want to talk from one Ethernet to another! Which will win? Ethernet or DECnet?
16

Layer 3 packet

source

dest

hops

data

Layer 3 header

17

Ethernet packet

source

dest

data

Ethernet header

18

Ethernet (802) addresses

OUI

group/individual global/local admin

Assigned in blocks of 224 Given 23-bit constant (OUI) plus g/i bit all 1s intended to mean broadcast
19

Its easy to confuse Ethernet with network

Both are multiaccess clouds But Ethernet does not scale. It cant replace IP as the Internet Protocol
Flat addresses No hop count Missing additional protocols (such as neighbor discovery) Perhaps missing features (such as fragmentation, error messages, congestion feedback)
20

So where did bridges come from?

21

Problem Statement
Need something that will sit between two Ethernets, and let a station on one Ethernet talk to another

22

Basic idea
Listen promiscuously Learn location of source address based on source address in packet and port from which packet received Forward based on learned location of destination

23

Whats different between this and a repeater?

no collisions with learning, can use more aggregate bandwidth than on any one link no artifacts of LAN technology (# of stations in ring, distance of CSMA/CD)

24

But loops are a disaster

No hop count Exponential proliferation
S

B1

B2

B3

25

But loops are a disaster

No hop count Exponential proliferation
S

B1

B2

B3

26

But loops are a disaster

No hop count Exponential proliferation
S

B1

B2

B3

27

But loops are a disaster

No hop count Exponential proliferation
S

B1

B2

B3

28

But loops are a disaster

No hop count Exponential proliferation
S

B1

B2

B3

29

What to do about loops?

Just say dont do that Or, spanning tree algorithm
Bridges gossip amongst themselves Compute loop-free subset Forward data on the spanning tree Other links are backups

30

Algorhyme
I think that I shall never see A graph more lovely than a tree. A tree whose crucial property Is loop-free connectivity. A tree which must be sure to span So packets can reach every LAN. First the Root must be selected By ID it is elected. Least cost paths from Root are traced In the tree these paths are placed. A mesh is made by folks like me. Then bridges find a spanning tree. Radia Perlman
31

11 9 10 4

7 2

14

32

11 9 10 4

7 2

14

33

11 9 10 4

7 2

14

34

Bother with spanning tree?

Maybe just tell customers dont do loops First bridge sold...

35

First Bridge Sold

36

So Bridges were a kludge, digging out of a bad decision

Why are they so popular?
plug and play simplicity high performance

Will they go away?

because of idiosyncracy of IP, need it for lower layer.
37

Note some things about bridges

Certainly dont get optimal source/ destination paths Temporary loops are a disaster
No hop count Exponential proliferation

But they are wonderfully plug-and-play

38

So what is Ethernet?
CSMA/CD, right? Not any more, really... source, destination (and no hop count) limited distance, scalability (not any more, really)

39

Switches
Ethernet used to be bus Easier to wire, more robust if star (one huge multiport repeater with pt-to-pt links If store and forward rather than repeater, and with learning, more aggregate bandwidth Can cascade devicesdo spanning tree Were reinvented the bridge!
40

Basic idea of a packet

Destination address Source address

data

41

When I started
Layer 3 had source, destination addresses Layer 2 was just point-to-point links (mostly) If layer 2 is multiaccess, then need two headers:
Layer 3 has ultimate source, destination Layer 2 has next hop source, destination
42

Hdrs inside hdrs

R1 R2 R3 D

As transmitted by S? (L2 hdr, L3 hdr) As transmitted by R1? As received by D?

43

Hdrs inside hdrs

R1 R2 Dest= Source= Layer 2 hdr Dest=D Source=S Layer 3 hdr
44

R3

S S:

Hdrs inside hdrs

R1 R2 Dest= Source= Layer 2 hdr Dest=D Source=S Layer 3 hdr
45

R3

S R1:

Hdrs inside hdrs

R1 R2 Dest=D Source=S Layer 2 hdr Layer 3 hdr
46

R3

S R2:

Hdrs inside hdrs

R1 R2 Dest= Source= Layer 2 hdr Dest=D Source=S Layer 3 hdr
47

R3

S R3:

What designing layer 3 meant

Layer 3 addresses Layer 3 packet format (IP, DECnet)
Source, destination, hop count,

A routing algorithm
Exchange information with your neighbors Collectively compute routes with all rtrs Compute a forwarding table
48

Network Layer
connectionless fans designed IPv4, IPv6, CLNP, IPX, AppleTalk, DECnet Connection-oriented reliable fans designed X.25 Connection-oriented datagram fans designed ATM, MPLS

49

Pieces of network layer

interface to network: addressing, packet formats, fragmentation and reassembly, error reports routing protocols autoconfiguring addresses/nbr discovery/ finding routers

50

Connection-oriented Nets
S
8

3 4

(3,51)=(7,21) (4,8)=(7,92) (4,17)=(7,12) 7

92

R1

R3
2 3 4

R2
8

(2,12)=(3,15) (2,92)=(4,8)

R4
2 1

R5
3

VC=8, 92, 8, 6

(1,8)=(3,6) (2,15)=(1,7)

51

Lots of connection-oriented networks

X.25: also have sequence number and ack number in packets (like TCP), and layer 3 guarantees delivery ATM: datagram, but fixed size packets (48 bytes data, 5 bytes header)

52

MPLS (multiprotocol label switching)

Connectionless, like MPLS, but arbitrary sized packets Add 32-bit hdr on top of IP pkt
20 bit label Hop count (hooray!)

53

Hierarchical connections (stacks of MPLS labels)

D8 S1 S2 S8 S4 S3 S6 S9 R2 S5 R1 D2 D3 D5 D4 D9 D2 D1

Routers in backbone only need to know about one flow: R1-R2

54

MPLS
Originally for faster forwarding than parsing IP header later traffic engineering classify pkts based on more than destination address

55

Connectionless Network Layers

Destination, source, hop count Maybe other stuff
fragmentation options (e.g., source routing) error reports special service requests (priority, custom routes) congestion indication

Real diff: size of addresses

56

Addresses
802 address flat, though assigned with OUI/rest. No topological significance layer 3 addresses: locator/node : topologically hierarchical address interesting difference:
IPv4, IPv6, IPX, AppleTalk: locator specific to a link CLNP, DECnet: locator area, whole campus
57

Hierarchy within Locator

Assume addresses assigned so that within a circle everything shares a prefix Can summarize lots of circles with a shorter prefix
2428* 27*
279* 272*

2* 23*

58

New topic: Routing Algorithms

59

Distributed Routing Protocols

Rtrs exchange control info Use it to calculate forwarding table Two basic types
distance vector link state

60

Distance Vector
Know
your own ID how many cables hanging off your box cost, for each cable, of getting to nbr
cost 3 cost 2 j k I am 4 n m cost 2 cost 7

61

cost 3 cost 2

j k I am 4

m n

cost 2 cost 7

distance vector rcvd from cable j cost 3 cost 2 12 3 15 3 12 5 3 18 0 7 15 distance vector rcvd from cable k

5 8 3 2 10 7 4 20 5 0 15 distance vector rcvd from cable m cost 2 0 5 3 2 19 9 5 22 2 4 7 distance vector rcvd from cable n cost 7 6 2 0 7 8 5 8 12 11 3 2 your own calculated distance vector 2 6 5 0 12 8 6 19 3 your own calculated forwarding table m j m 0 k j k/j n j ? ? ? ?
62

cost 3 cost 2

j k I am 4

m n

cost 2 cost 7

distance vector rcvd from cable j cost 3 cost 2 12 3 15 3 12 5 3 18 0 7 15 distance vector rcvd from cable k

5 8 3 2 10 7 4 20 5 0 15 distance vector rcvd from cable m cost 2 0 5 3 2 19 9 5 22 2 4 7 distance vector rcvd from cable n cost 7 6 2 0 7 8 5 8 12 11 3 2 your own calculated distance vector 2 6 5 0 12 8 6 19 3 your own calculated forwarding table m j m 0 k j k/j n j ? ? ? ?
63

cost 3 cost 2

j k I am 4

m n

cost 2 cost 7

distance vector rcvd from cable j cost 3 cost 2 12 3 15 3 12 5 3 18 0 7 15 distance vector rcvd from cable k

5 8 3 2 10 7 4 20 5 0 15 distance vector rcvd from cable m cost 2 0 5 3 2 19 9 5 22 2 4 7 distance vector rcvd from cable n cost 7 6 2 0 7 8 5 8 12 11 3 2 your own calculated distance vector 2 6 5 0 12 8 6 19 3 your own calculated forwarding table m j m 0 k j k/j n j ? ? ? ?
64

cost 3 cost 2

j k I am 4

m n

cost 2 cost 7

distance vector rcvd from cable j cost 3 cost 2 12 3 15 3 12 5 3 18 0 7 15 distance vector rcvd from cable k

5 8 3 2 10 7 4 20 5 0 15 distance vector rcvd from cable m cost 2 0 5 3 2 19 9 5 22 2 4 7 distance vector rcvd from cable n cost 7 6 2 0 7 8 5 8 12 11 3 2 your own calculated distance vector 2 6 5 0 12 8 6 19 3 your own calculated forwarding table m j m 0 k j k/j n j ? ? ? ?
65

cost 3 cost 2

j k I am 4

m n

cost 2 cost 7

distance vector rcvd from cable j cost 3 cost 2 12 3 15 3 12 5 3 18 0 7 15 distance vector rcvd from cable k

5 8 3 2 10 7 4 20 5 0 15 distance vector rcvd from cable m cost 2 0 5 3 2 19 9 5 22 2 4 7 distance vector rcvd from cable n cost 7 6 2 0 7 8 5 8 12 11 3 2 your own calculated distance vector 2 6 5 0 12 8 6 19 3 your own calculated forwarding table m j m 0 k j k/j n j ? ? ? ?
66

cost 3 cost 2

j k I am 4

m n

cost 2 cost 7

distance vector rcvd from cable j cost 3 cost 2 12 3 15 3 12 5 3 18 0 7 15 distance vector rcvd from cable k

5 8 3 2 10 7 4 20 5 0 15 distance vector rcvd from cable m cost 2 0 5 3 2 19 9 5 22 2 4 7 distance vector rcvd from cable n cost 7 6 2 0 7 8 5 8 12 11 3 2 your own calculated distance vector 2 6 5 0 12 8 6 19 3 your own calculated forwarding table m j m 0 k j k/j n j ? ? ? ?
67

Looping Problem
A B C

68

Looping Problem
A 2 B 1 C 0 Cost to C

69

Looping Problem
direction towards C A 2 B 1 direction towards C C 0 Cost to C

70

Looping Problem
A 2 B 1 What is Bs cost to C now? C 0 Cost to C

71

Looping Problem
A 2 B 1 3 C 0 Cost to C

72

Looping Problem
direction towards C A 2 direction towards C B 1 3 C 0 Cost to C

73

Looping Problem
direction towards C A 2 4 direction towards C B 1 3 C 0 Cost to C

74

Looping Problem
direction towards C A 2 4 direction towards C B 1 3 5 C 0 Cost to C

75

Looping Problem worse with high connectivity

C N M V H

76

Split Horizon: one of several optimizations

Dont tell neighbor N you can reach D if youd forward to D through N A B C

A C D

77

Link State Routing

meet nbrs Construct Link State Packet (LSP)
who you are list of (nbr, cost) pairs

Broadcast LSPs to all rtrs (a miracle occurs) Store latest LSP from each rtr Compute Routes (breadth first, i.e., shortest path firstwell known and efficient algorithm)
78

A 2 D A B/6 D/2 B A/6 C/2 E/1

B 1 2 E C B/2 F/2 G/5

C 2 F D A/2 E/2

5 G 1 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

79

Computing Routes
Edsgar Dijkstras algorithm:
calculate tree of shortest paths from self to each also calculate cost from self to each Algorithm: step 0: put (SELF, 0) on tree step 1: look at LSP of node (N,c) just put on tree. If for any nbr K, this is best path so far to K, put (K, c +dist(N,K)) on tree, child of N, with dotted line step 2: make dotted line with smallest cost solid, go to step 1
80

Look at LSP of new tree node

A B/6 D/2 B A/6 C/2 E/1 C B/2 F/2 G/5 D A/2 E/2 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

C(0) B(2) G(5)

F(2)

81

Make shortest TENT solid

A B/6 D/2 B A/6 C/2 E/1 C B/2 F/2 G/5 D A/2 E/2 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

C(0) B(2) G(5)

F(2)

82

Look at LSP of newest tree node

A B/6 D/2 B A/6 C/2 E/1 C B/2 F/2 G/5 D A/2 E/2 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

C(0) B(2) E(4) G(5) G(3)

83

F(2)

Make shortest TENT solid

A B/6 D/2 B A/6 C/2 E/1 C B/2 F/2 G/5 D A/2 E/2 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

C(0) B(2) E(4)

F(2) G(3)
84

Look at LSP of newest tree node

A B/6 D/2 B A/6 C/2 E/1 C B/2 F/2 G/5 D A/2 E/2 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

C(0) B(2) A(8) E(3)

F(2) G(3)
85

Make shortest TENT solid

A B/6 D/2 B A/6 C/2 E/1 C B/2 F/2 G/5 D A/2 E/2 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

C(0) B(2) A(8) E(3)

F(2) G(3)
86

Look at LSP of newest tree node

A B/6 D/2 B A/6 C/2 E/1 C B/2 F/2 G/5 D A/2 E/2 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

C(0) B(2) A(8) E(3) D(5)

87

F(2) G(3)

Make shortest TENT solid

A B/6 D/2 B A/6 C/2 E/1 C B/2 F/2 G/5 D A/2 E/2 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

C(0) B(2) A(8) E(3) D(5)

88

F(2) G(3)

Look at newest tree nodes LSP

A B/6 D/2 B A/6 C/2 E/1 C B/2 F/2 G/5 D A/2 E/2 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

C(0) B(2) A(8) E(3) D(5)

89

F(2) G(3)

Make shortest TENT solid

A B/6 D/2 B A/6 C/2 E/1 C B/2 F/2 G/5 D A/2 E/2 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

C(0) B(2) A(8) E(3) D(5)

90

F(2) G(3)

Look at newest nodes LSP

A B/6 D/2 B A/6 C/2 E/1 C B/2 F/2 G/5 D A/2 E/2 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

C(0)

B(2) A(8) E(3) D(5)

F(2) G(3)

A(7)
91

Make shortest TENT solid

A B/6 D/2 B A/6 C/2 E/1 C B/2 F/2 G/5 D A/2 E/2 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

C(0)

B(2) E(3) D(5)

F(2) G(3)

A(7)
92

Were done!
A B/6 D/2 B A/6 C/2 E/1 C B/2 F/2 G/5 D A/2 E/2 E B/1 D/2 F/4 F C/2 E/4 G/1 G C/5 F/1

C(0)

B(2) E(3) D(5)

F(2) G(3)

A(7)
93

A miracle occurs
First link state protocol: ARPANET I wanted to do something similar for DECnet My manager said Only if you can prove its stable Given a choice between a proof and a counterexample
94

Routing Robustness
This failure mode actually occurred in the ARPANET I designed self-stabilizing link state protocol but only after sick/evil node gone IS-IS and OSPF do it self-stabilizing My thesis: robust even if some of the routers are evil. More than securing the routing protocol: it deals with packet delivery
95

Distance vector vs link state

Memory: distance vector wins (but memory is cheap) Computation: debatable Simplicity of coding: simple distance vector wins. Complex new-fangled distance vector, no Convergence speed: link state Functionality: link state; custom routes, mapping the net, troubleshooting, sabotage-proof routing
96

Specific Routing Protocols

Interdomain vs Intradomain Intradomain:
link state (OSPF, IS-IS) distance vector (RIP)

Interdomain
BGP

97

BGP (Border Gateway Protocol)

Policies, not just minimize path Path vector: given reported paths to D from each nbr, and configured preferences, choose your path to D
dont ever route through domain X, or not to D, or only as last resort

Other policies: dont tell nbr about D, or lie to nbr about D making path look worse
98

Path vector/Distance vector

Distance vector
Each router reports to its neighbors {(D,cost)} Each router chooses best path based on min (reported cost to D+link cost to nbr)

Path vector
Each rtr R reports {(D,list of ASs in Rs chosen path to D)} Each rtr chooses best path based on configured policies
99

BGP Configuration
path preference rules which nbr to tell about which destinations how to edit the path when telling nbr N about prefix P (add fake hops to discourage N from using you to get to P)

100

TRILL working group in IETF

TRILL= TRansparent Interconnection of Lots of Links Use layer 3 routing, and encapsulate with a civilized header But still look like a bridge from the outside

101

Goal
Design so that change can be incremental With TRILL, replace any subset of bridges with RBridges
still looks to IP like one giant Ethernet the more bridges you replace with RBridges, better bandwidth utilization, more stability

102

Run link state protocol

So all the RBridges know how to reach all the other RBridges But dont know anything about endnodes

103

Why link state?

Since all switches know the complete topology, easy to compute lots of trees deterministically (well get to that later) Easy to piggyback nickname allocation protocol (well get to that later)

104

Routing inside campus

First RB encapsulates to last RB
So header is safe (has hop count) Inner RBridges only need to know how to reach destination RBridge

Still need tree for unknown/multicast

But dont need spanning tree protocol compute tree(s) deterministically from the link state database
105

Rbridging
R4 R7 R3 R6 a R1 R2 c R5

106

Details
What the encapsulated packet looks like How R1 knows that R2 is the correct last RBridge

107

Encapsulated Frame
(Ethernet) outer header dest (nexthop) srce (Xmitter) Ethertype=TRILL TRILL header first RBridge last RBridge TTL original frame

TRILL header specifies RBridges with 2-byte nicknames

108

2-byte Nicknames
Saves hdr room, faster fwding Dynamically acquired Choose unused #, announce in LSP If collision, IDs and priorities break tie Loser chooses another nickname Configured nicknames higher priority
109

How does R1 know that R2 is the correct last RBridge?

If R1 doesnt, R1 sends packet through a tree When R2 decapsulates, it remembers (ingress RBridge, source MAC)

110

Use of first and last RBridge in TRILL header

For Unicast, obvious
Route towards last RBridge Learn location of source from first RBridge

For Multicast/unknown destination

Use of first
to learn location of source endnode to do RPF check on multicast

Use of last
To allow first RB to specify a tree Campus calculates some number of trees
111

Algorhyme v2
I hope that we shall one day see A graph more lovely than a tree. A graph to boost efficiency While still configuration-free. A network where RBridges can Route packets to their target LAN. The paths they find, to our elation, Are least cost paths to destination. With packet hop counts we now see, The network need not be loop-free. RBridges work transparently. Without a common spanning tree.
Ray Perlner
112

Wrap-up
folklore of protocol design things too obvious to say, but everyone gets them wrong

113

Forward Compatibility
Reserved fields
spare bits ignore them on receipt, set them to zero. Can maybe be used for something in the future

TLV encoding
type, length, value so can skip new TLVs maybe have range of Ts to ignore if unknown, others to drop packet
114

Forward Compability
Make fields large enough
IP address, packet identifier, TCP sequence #

Version number
what is new version vs new protocol?
same lower layer multiplex info

therefore, must always be in same place! drop if version # bigger

115

Fancy version # variants

Might be security threat to trick two Vn nodes into talk V(n-1) So maybe have highest version I support in addition to version of this packet Or just a bit I can support higher (we did this for IKEv2) Maybe have minor version #, for compatible changes. Old node ignores it
116

Version #
Nobody seems to do this right IP, IKEv1, SSL unspecified what to do if version # different. Most implementations ignore it. SSL v3 moved version field!
v2 sets it to 0.2. v3 sets (different field) to 3.0. v2 node will ignore version number field, and happily parse the rest of the packet
117

Avoid flag days

Want to be able to migrate a running network ARPANET routing: ran both routing algorithms (but they had to compute the same forwarding table)
initially forward based on old, compute both one by one: forward based on new one-by-one: delete old
118

Parameters
Minimize these:
someone has to document it customer has to read documentation and understand it

How to avoid
architectural constants if possible automatically configure if possible
119

Settable Parameters
Make sure they cant be set incompatibly across nodes, across layers, etc. (e.g., hello time and dead timer) Make sure they can be set at nodes one at a time and the net can stay running

120

Parameter tricks
IS-IS pairwise parameters reported in hellos area-wide parameters reported in LSPs Bridges Use Roots values, sent in spanning tree msgs

121

Summary
If things arent simple, they wont work Good engineering requires understanding tradeoffs and previous approaches. Its never a waste of time to answer why is something that way Dont believe everything you hear Know the problem youre solving before you try to solve it!
122