forwarding decisions - univ-pau.frcpham.perso.univ-pau.fr/enseignement/iup/routerforwarding.pdf ·...

From a Nick McKeown's tutorial, 1999 andslides from Kalyanaraman (with figure fromKeshav) Some slides modified by C. Pham 1

Forwarding Decisions• ATM and MPLS switches

– Direct Lookup• Bridges and Ethernet switches

– Associative Lookup– Hashing– Trees and tries

• IP Routers– CIDR– Patricia trees/tries– Other methods– Caching

• Packet Classification


ATM and MPLS SwitchesDirect Lookup

VCI

Address

MemoryD

ata(Port, VCI)


Bridges and Ethernet SwitchesAssociative Lookups

NetworkAddress

AssociatedData

AssociativeMemory or CAM

Search Data

48

log2N

AssociatedData

Hit?

Address{

Advantages:• Simple

Disadvantages• Slow• High Power• Small• Expensive


Bridges and Ethernet SwitchesHashing

HashingFunction

Memory

Add

ress

Dat

a

Search Data

48

log2N

AssociatedData

Hit?

Address{16


Lookups Using HashingAn example

Hashing Function

CRC-1616

#1 #2 #3 #4

#1 #2

#1 #2 #3Linked lists

Memory

Search Data

48

log2N

AssociatedData

Hit?

Address{M entries

N lists


Lookups Using HashingPerformance of simple example

Most addresses in their own list

Most addresses in one list


Lookups Using Hashing

Advantages:• Simple

• Expected lookup time can be small

Disadvantages• Non-deterministic lookup time

• Inefficient use of memory


Trees and Tries

Binary Search Tree

< >

< > < >

log2 N

N entries

Binary Search Trie

0 1

0 1 0 1

111010


Tries

• An entry is:– a pointer to another array,– a special symbol indicating

no better match– a null pointer indicating that

the longst match is the parentnode

• Two ways to improveperformance– cache recently used addresses

in a CAM– move common entries up to a

higher level (match longerstrings)

128.32.1.2 ?


TriesMultiway tries

16-ary Search Trie

0000, ptr 1111, ptr

0000, 0 1111, ptr

000011110000

0000, 0 1111, ptr

111111111111


Trees and TriesMultiway tries

Table produced from 215 randomly generated 48-bit addresses


IP RoutersClass-based addresses

Class A Class B Class C D

212.17.9.4Class AClass BClass C 212.17.9.0 Port 4

Exact Match

Routing Table:

IP Address Space


IP RoutersCIDR

A B C D0 232-1

0 232-1

128.9/16

128.9.0.0

216

142.12/19

65/24

Classless:

Class-based:

128.9.16.14


IP RoutersCIDR

0 232-1

128.9/16

128.9.16.14

128.9.16/20 128.9.176/20

128.9.19/24

128.9.25/24

Most specific route = “longest matching prefix”


IP RoutersMetrics for Lookups

128.9/16128.9.16/20

128.9.176/20

128.9.19/24128.9.25/24

142.12/19

65/24

Prefix Port35271013

128.9.16.14• Lookup time• Storage space• Update time• Preprocessing time


IP RouterLookup

IPv4 unicast destination address based lookup

Dstn Addr Next Hop

--------

---- ----

--------

Destination Next HopForwarding Table

Next Hop Computation

Forwarding Engine

Incoming Packet

HEADER


Lookup Performance Required

Gigabit Ethernet (84B packets): 1.49 Mpps


Size of the Routing Table

Source: http://www.telstra.net/ops/bgptable.html

Exponentialgrowth before

CIDR

About10k newprefixes per year


Size of the Forwarding TableSource: http://www.telstra.net/ops/bgptable.html

95 96 97 98 99 00Year

Num

ber

of P

refi

xes

10,000/year

Renewed Exponential Growth

Renewed growth due to multi-homing of enterprise networks!


Routing Lookups in Hardware

Prefix length

Num

ber

Most prefixes are 24-bits or shorter


Routing Lookups in Hardware14

2.19

.6.1

4

Prefixes up to 24-bits14

2.19

.614

1 Next Hop

24

Next Hop

142.19.6

224 = 16M entries


Routing Lookups in Hardware12

8.3.

72.4

4

Prefixes up to 24-bits12

8.3.

7244

1 Next Hop

128.3.72

24 0 Pointer

8

Prefixes above 24-bits

Next Hop

Next Hop

Next Hopof

fset

base


Caching Addresses

CPU BufferMemory

LineCard

DMA

MAC

LocalBuffer

Memory

LineCard

DMA

MAC

LocalBuffer

Memory

LineCard

DMA

MAC

LocalBuffer

Memory

Fast Path

Slow Path

Advantages

Increased averagelookupperformance

Disadvantages

Decreased locality inbackbone traffic

Cache size

Cache managementoverhead

Hardwareimplementation difficult


Caching Addresses

LAN:Average flow < 40 packets

WAN: Huge Number of flows

Cache = 10% of Full Table

Cache Hit Rate


IP Router LookupsReferences

• A. Brodnik, S. Carlsson, M. Degermark, S. Pink. “Small ForwardingTables for Fast Routing Lookups”, Sigcomm 1997, pp 3-14.

• B. Lampson, V. Srinivasan, G. Varghese. “ IP lookups using multiwayand multicolumn search”, Infocom 1998, pp 1248-56, vol. 3.

• M. Waldvogel, G. Varghese, J. Turner, B. Plattner. “Scalable highspeed IP routing lookups”, Sigcomm 1997, pp 25-36.

• P. Gupta, S. Lin, N.McKeown. “Routing lookups in hardware atmemory access speeds”, Infocom 1998, pp 1241-1248, vol. 3.

• S. Nilsson, G. Karlsson. “Fast address lookup for Internet routers”,IFIP Intl Conf on Broadband Communications, Stuttgart, Germany,April 1-3, 1998.

• V. Srinivasan, G.Varghese. “Fast IP lookups using controlled prefixexpansion”, Sigmetrics, June 1998.


Course Outline

• Packet Lookup and Classification:Where does a packet go next?

• Switching Fabrics:How does the packet get there?


Switching Fabrics• Overview• Output and Input Queueing• Output Queueing• Input Queueing

– Scheduling algorithms– Combining input and output queues– Multicast traffic– Other non-blocking fabrics

• Multistage Switches


Basic Architectural ComponentsDatapath: per-packet processing

ForwardingDecision

ForwardingDecision

ForwardingDecision

ForwardingTable

ForwardingTable

ForwardingTable

Interconnect

OutputScheduling

1.2.

3.

Transfers data from aninput to an output

many ports (density),high speeds


Background: Circuit switch

• A switch that can handle N calls has N logical inputs andN logical outputs– N up to 200,000

• Moves 8-bit samples from an input to an output port– Recall that samples have no headers– Destination of sample depends on time at which it arrives at the

switch• In practice, input trunks are multiplexed

– Multiplexed trunks carry frames = set of samples• Goal: extract samples from frame, and depending on

position in frame, switch to output– each incoming sample has to get to the right output line and the

right slot in the output frame


Call blocking

• Can’t find a path from input to output• Internal blocking

– slot in output frame exists, but no path• Output blocking

– no slot in output frame is available• Output blocking is reduced in transit

switches– need to put a sample in one of several slots

going to the desired next hop


Multiplexors and demultiplexors

• Most trunks time division multiplex voicesamples

• At a central office, trunk is demultiplexedand distributed to active circuits

• Synchronous multiplexor– N input lines– Output runs N times as fast as input

…

123

N

MUX…

123

N

De-MUX1 2 3 … N


Time division switching

• Key idea: when de-multiplexing, position in framedetermines output trunk

• Time division switching interchanges sample positionwithin a frame: time slot interchange (TSI)


Time Division SwitchingLimitations

• To build a 120,000 circuit switch– read and write samples 120,000 every 125us, a

R&W operation in 0.5 ns!– Today DRAM has access time from 80 to 40 ns– If we use 40 ns DRAM, it's 80 times more than

what we need– Maximum #circuit= 120,000/80=1500!– Too small!!


Space division switching

• Each sample takes a different path throughthe switch, depending on its destination


Crossbar

• Simplest possible space-division switch

• Crosspoints can be turnedon or off, long enough totransfer a packet from aninput to an output

• Expensive• Internally nonblocking

– but need N2 crosspoints– time to set each crosspoint

grows quadratically

configuration

Dat

a In

Data Out


Multistage crossbar (1)

• In a crossbar during eachswitching time only onecross-point per row orcolumn is active

• Can save crosspoints if across-point can attach tomore than one input line

• This is done in a multistagecrossbar

N/narraysn x k

karraysN/n x N/n

N/narraywk x n


Multistage crossbar (2)

• Can suffer internal blocking– unless sufficient number of second-level

stages, k ≥ n• Number of crosspoints < N2

• Finding a path from input to outputrequires a depth-first-search

• Scales better than crossbar, but still not toowell– 120,000 call switch needs ~250 million

crosspoints


The true cost of telephoneswitching

• In a central switching system, the high costis the line card.

• Now the true cost is the copper wire to thecustomer premises!!

• In long-distance, the high cost is in layinglines, acquiring rights of way and switch-control software!

• So, saving a few thousand crosspoints isnot going to make phone call cheaper!


Packet switches

• In a circuit switch, path of a sample is determinedat time of connection establishment

• No need for a sample header--position in frameused

• In a packet switch, packets carry a destinationfield or label– Need to look up destination port on-the-fly

• Datagram switches– lookup based on entire destination address (longest-

prefix match)• Cell or Label-switches

– lookup based on VCI or Labels


Blocking in packet switches

• Can have both internal and output blocking• Internal

– no path to output• Output

– trunk unavailable• Unlike a circuit switch, cannot predict if

packets will block (why?)• If packet is blocked, must either buffer or

drop


Dealing with blocking in packetswitches

• Over-provisioning– internal links much faster than inputs

• Buffers– at input or output

• Backpressure– if switch fabric doesn’t have buffers, prevent

packet from entering until path is available• Parallel switch fabrics

– increases effective switching capacity


Switch Fabrics: Bufferedcrossbar

• What happens ifpackets at two inputsboth want to go tosame output?

• Can defer one at aninput buffer

• Or, buffer cross-points: complex arbiter


Switch fabric element

• Goal: towards building “self-routing” fabrics• Can build complicated fabrics from a simple

element

• Routing rule: if 0, send packet to upper output,else to lower output– If both packets to same output, buffer or drop

0

1

data 10

data 00


Banyan• Simplest self-routing recursive fabric, 2n output

need n stages with 2n-1 components in each stage

• What if two packets both want to go to the sameoutput→output blocking

000001

010011

100101

110111

000001

010011

100101

110111


Blocking in Banyan S/wsSorting

• Can avoid blocking by choosing order in which packetsappear at input ports

• If we can– present packets at inputs sorted by output– remove duplicates– remove gaps– precede banyan with a perfect shuffle stage– then no internal blocking

• For example: [X, 011, 010, X, 011, X, X, X]:• Sort => [010, 011, 011, X, X, X, X, X]• Remove dups => [010, 011, X, X, X, X, X, X]• Shuffle => [010, X, 011, X, X, X, X, X]• Need sort, trap and shuffle networks.


Sorting using Merging

• Build sorters from merge networks• Assume we can merge two sorted lists• Sort pairwise, merge, recurse

Sort {5,7,2,3,6,2,4,5}

1/ sort 2 by 22/ merge adjacent lists

to get two 4-el lists3/ merge de two lists

with a merge network

2357

2

54

6


Putting it together- BatcherBanyan

• What about trapped duplicates?– recirculate to beginning– or run output of trap to multiple banyans (dilation)


Non-Blocking Batcher-Banyan

3

7

5

2

6

0

1

4

7

2

3

5

6

1

0

4

7

5

2

3

1

0

6

4

7

0

5

1

3

4

2

6

7

4

5

6

0

3

1

2

7

6

4

5

3

2

0

2

7

6

5

4

3

2

1

0

000001010011100101110111

Batcher Sorter Self-Routing Network

• Fabric can be used as scheduler. •Batcher-Banyan network is blocking for multicast.

a dans le sens de la flèche si a > b,a dans le sens opposé si a est tout seul


InterconnectsTwo basic queueing techniques

Input Queueing Output Queueing

Usually a non-blockingswitch fabric (e.g. crossbar)

Usually a fast bus


InterconnectsOutput Queueing

Individual Output Queues Centralized Shared Memory

Memory b/w = (N+1).R

1

2

N

Memory b/w = 2N.R

1

2

N


Output QueueingThe “ideal”

1

1

1

1

1

1

1

1

1

11

1

2

2

2

2

2

2


Output QueueingHow fast can we make centralized shared memory?

SharedMemory

200 byte bus

5ns SRAM

1

2

N

• 5ns per memory operation• Two memory operations per packet• Therefore, up to 160Gb/s• In practice, closer to 80Gb/s


Switching Fabrics

• Output and Input Queueing• Output Queueing• Input Queueing

– Scheduling algorithms– Combining input and output queues– Multicast traffic– Other non-blocking fabrics

• Multistage Switches


InterconnectsInput Queueing with Crossbar

configuration

Dat

a In

Data Out

Scheduler

Memory b/w = 2R


Input QueueingHead of Line Blocking

Del

ay

Load58.6% 100%


Head of Line Blocking


Input QueueingVirtual output queues


Input QueuesVirtual Output Queues

Del

ay

Load100%


Input Queueing Virtual Output Queues

Scheduler

Memory b/w = 2R

Can be quitecomplex!


Input QueueingWhy is serving long/old queues better than

serving maximum number of queues?

• When traffic is uniformly distributed, servicing themaximum number of queues leads to 100% throughput.• When traffic is non-uniform, some queues become longer than others.• A good algorithm keeps the queue lengths matched, and services a large number of queues.

VOQ #

Avg

Occ

upan

cy Uniform traffic

VOQ #

Avg

Occ

upan

cy

Non-uniform traffic


Summary: progression

Shared Memory

InputQueued

Combined Input and

Output QueuedParallelPacket

Switches37526014

72356104

75231064

70513426

74560312

76453202

76543210

000001010011100101110111

Batcher Sorter Self-Routing Network

Multistage

forwarding decisions - univ-pau.frcpham.perso.univ-pau.fr/enseignement/iup/routerforwarding.pdf ·...

Documents