High Performance Switching and Routingtiny-tera.stanford.edu/~nickm/talks/ResearchOverview...Why we need faster switches/routers High Performance Switching and Routing 7 Trafﬁc Inversion

Nick McKeownAssistant Professor of Electrical Engineeringand Computer Science

[email protected]://www.stanford.edu/~nickm

High PerformanceSwitching and Routing

High Performance Switching and Routing

2

1. The Demand for Bandwidth

2. The Shortage of

3. The Architecture of Switches

Switching/Routing Capacity

and Routers

4. Some (of our) solutions


3

What’s the Problem?

Time

Mos

t thi

ngs


4

Source: http://www.mfsdatanet.com/MAE/west.stats.html

Feb Apr Jun Aug Nov

200

400

600

800

460

200150

123

Agg

rega

te b

andw

idth

(M

bps)

100

620

520

720

Feb Apr

The San Jose NAPThe demand


5

The supplyR

oute

r P

erfo

rman

ce (

pack

ets/

seco

nd)

103

19901986 1994 1997

104105

106


6

1986 1992 1997

pack

ets/

seco

nd

Demand

Supply

Why we need faster switches/routers


7

Traffic Inversion10 years ago


8

ISP

ISP

Traffic InversionToday


9

November 1st, 1996

Why is this a problem?

Pac

ket L

oss

(%)


10


2. The Shortage of



and Routers



11

The Architecture of

Forwarding

OutputScheduler

Signaling &

Generic Packet Switch:

Decision

Processor

Switches and Routers

Mgmt

Interconnect

(e.g. IP Router, ATM Switch, LAN Switch)

Data Hdr


12

Performance of IP Routers

Min back-to-back packet size Packet size

ArrivalTime

CopyTime

Forwarding

Time

Time

Decision


13

Performance of IP Routers

Min back-to-back packet size Packet size

Arrivalrate

Copyrate

Headerprocessingtime

Most routersdo this poorly!

Most routersdo this ~ ok

Time


14

The Evolution of RoutersThe first shared memory routers

RoutingCPU

LineCard

LineCard

LineCard

BufferMemory

DMA DMA DMA

MAC MAC MAC


15

The Evolution of RoutersThe first shared memory routers

RoutingCPU

LineCard

LineCard

LineCard

BufferMemory

DMA DMA DMA

MAC MAC MAC


16

The Evolution of RoutersReducing the number of bus copies

RoutingCPU

LineCard

BufferMemory

DMA

MAC

BufferMemory

RouteCache

DMA

MAC

BufferMemory

RouteCache

DMA

MAC

BufferMemory

RouteCache


17

The Evolution of RoutersReducing the number of bus copies

RoutingCPU

LineCard

BufferMemory

DMA

MAC

BufferMemory

RouteCache

DMA

MAC

BufferMemory

RouteCache

DMA

MAC

BufferMemory

RouteCache

updates


18

The Evolution of RoutersAvoiding bus contention

MA

C

Buffer

Mem

ory

Route

Cache

MAC

BufferMemory

RouteCache

MA

C

Buf

fer

Mem

ory

Rou

teC

ache

BufferMemory

CPUROUTE

Advantage:Non-blocking backplane—

Disadvantage:Difficult to provide QoS

high throughput


19


2. The Shortage of



and Routers



20

Some (of our)Solutions

1. Accelerating Forwardng Decisions:• Longest-matching prefixes

2. Interconnections: Switched Backplanes• Input Queueing

— Theory— Unicast— Multicast

• Fast Buffering

• Speedup

• TheTiny Tera Project


21

Routing Lookups

212.17.9.4

Class A

Class B

Class C212.17.9.0 Port 4

Class A Class B Class C D

ExactMatch

(hash, cache,pipeline...)

Routing Table:


22

212.0.0.0/8

212.17.0.0/16212.17.9.0/24

212.17.9.4

CIDR uses “longest matching prefix” routing:

Hashing, caching and pipelining are hard!

Routing Lookups withCIDR (“supernetting”)


23

Perform Lookups Faster

Size ofRouting Tables

Cost of Memory(per byte)

Observation #1:

Time


24

Performing Lookups Faster

Prefix length

Numberin routingtable

24

212.17.9.0/24

212.17.9.40 232-1

256

Observation #2:


25

212.17.9.1 101

0

1

1

Port 4

Port 4Port 3

look further

look furtherPort 3

256Port 4Port 5

Port 4Port 5

16Mbytes of 50ns DRAM

<1Mbyte of 50ns DRAM

20 million lookups per second


26




• Fast Buffering

• Speedup



27

Shared Memory:

Input Queueing:

Should we use shared memory

Advantages:SimplicityHigh Bandwidth

Disadvantages:HOL BlockingLess efficient

Advantages:Highest Throughput.

Disadvantages:N-fold internal speedup

N p

orts

Difficult to control packet delay.

or input-queueing?

Possible to control packet delay.

Shared Memory


28

TimeM

emor

y S

ize

Time

Mem

ory

Spe

ed

SRAM

DRAM

Memory Bandwidth


29

An aside: How fast can sharedmemory operate?

Route Lookup

SharedMemory

200byte packet

5ns SRAM

How fast can a 16 port switch run with this architecture?

5ns per packet 2 memory operations per cell time×aggregate bandwidth is 160Gb/s⇒


30

Because of ashortage of memory bandwidth, mostmultigigabit and terabit switches and routers useeither:

1. Input Queueing, or2. Combined Input and Output Queueing.

Should we use shared memoryor input-queueing?


31

1 24 2

1 2

Inputs Outputs

ρmax 2 2– 58%= =

Input Cell Buffer

Cells CellsN

The Problem A Solution....

Head of Line Blocking

“Virtual Output Queueing”

ρmax 100%=


32

Input 1

Q(1,1)

Q(1,n)

A1(t)

Input m

Q(m,1)

Q(m,n)

Am(t)

D1(t)

Dn(t)

Output 1

Output n

Matching, MA1,1(t)

?

...but requires scheduling...


33

RequestGraph

BipartiteMatching

1234

1234

1234

1234

(Weight = 18)

25

242

7

....which is equivalent to graph matching


34

Practical Algorithms

1. iSLIP — Weight = 1

2. iLQF — Weight = Occupancy

— Simple to implement

— Iterative round-robin

3. iOCF — Weight = Cell Age

4. LPF — Weight = Backlog

Simple, fast,efficient

Good fornon-uniformtraffic.Complex!Good fornon-uniformtraffic.Simple.


35

Achieving 100% ThroughputLongestQueueFirst & OldestCell First

1234

1234

1234

1234

10 1

1

1

1 10

Maximum weight

WeightWaiting Time 100%Queue Length{ }=


36

Theorems

DefE Li j, n( )i j,∑ ∞< n∀,

100% throughput

Lyapunov Stability Criterion:

E V L n 1+( )( ) V L n( )( ) L n( )–[ ] 0 L n( ) k>∀,≤

Theorem:Both LQF and OCF can achieve 100% throughput forindependent traffic both uniform and non-uniform.

Proof:

http://tiny-tera.stanford.edu/~adisak/research.html


37

Approximating LQF and OCF

Iteration steps

Step 1. Request

Step 2. Grant to the largest request

Step 3. Accept grant to the largest request

iLQF&iOCF

10 1

1

1

1 10

10 1

1 10

10

10

1 2 3


38

Problem is in ComparatorsiLQF and iOCF

>input i

Grant Arbiter , 1

1

N

>

Accept Arbiter N

>input i

Grant Arbiter N

1

i

N

i>

Accept Arbiter, 1

Decoder

Decoder

b

b

b

b

b

b

b

b

b

b

b

b

logN

logN

Req

uest

s

Mat

ches

Clear Requests

L1,1

L2,1

LN,1

L1,N

L2,N

LN,N

L1,1

L1,2

L1,N

L1,N

L1,N

L1,N


39

Solution toComplexity Problem

☛ Longest Port First (LPF)

☛ Oldest Port First (OPF)

Advantages

— SIMPLER.

• Can use maximum size matching — O(N 2.5).

— FASTER.• Move magnitude comparator out of the critical path.

• Lends itself well to pipelining.


40

1

2

1

2

Using Port Occupancy

i.e. w1,1 = L1,1 + L1,2 + L1,1 + L2,1

w1,1

Input occupancy Output occupancy

L1,1

L1,2

L2,1

L2,2

LPF Algorithm

w i j, L i j,j∑ L i j,

i∑+=


41

On The Theorems

Theorem:LPF can achieve 100% throughput for independenttraffic both uniform and non-uniform.

Theorem:An LPF match is of both maximum weight andmaximum size.

Proof:

E V L n 1+( )( ) V L n( )( ) L n( )–[ ] 0 L n( ) k>∀,≤

V L n( )( ) LT

n( )TL n( )=


42

Presorting Inputs & Outputs

1920 9

4142 31

4445 34

2021 10

22

25

1

0

1

0

1

1

0

1

0

1

2 20 0

17 80

0 0 1

19 20 9

41 42 31

44 3445

20 21 10

22

25

1

1

1

0

1

1

0

0

0

1

19 20 9

41 42 31

44 3445

20 21 10

22

25

1

1

1

0

1

1

0

0

0

1

Weight request

Permute


43

11 0

10 1

00 1

1920 9

4142 31

4445 34

2021 10

22

25

1

0

1

0

1

1

0

1

0

1

Remove Weights

11 0

10 1

00 1

Matching01 0

10 0

00 1


44

Implementation

Input Occup Output Occup

Input permutation Output permutation

Maxsize

Raw Requests

Sorter Sorter

X Bar X Bar

11 011 0

00 1

Matching

Permuted Requests

Match

10 001 0

00 1

{10, 20, 30} {20, 25, 15}

{3, 2, 1} {2, 1, 3}


45

Multicast TrafficQueue Architecture

2. Why treat multicast differently?1. Making use of the crossbar

3. Why maintain a single FIFO queue?4. Fanout-splitting

134


46

0.1

1

10

100

1000

10000

0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24

Ave

rage

Cel

l Lat

ency

Offered Load

Fanout-splittingNo Fanout-splitting

Fanout-Splitting


47

Multicast Traffic

1. Residue Concentration

2. Tetris-based schedulers


48




• Fast Buffering

• Speedup

• TheTiny TeraProject


49

Fast BufferingPing-pong Memory

BufferMemory

BufferMemory

BufferMemory

M

M/2

M/2


50

M/2

M/2

iii

iii

iv

v

X1 X2=

X1

X2

M/2

M/2

X1

X2 X1 X2=

memory 1

memory 2


51


M

t

Occupancy

M/2

M/2

Maximum “cost” = M/2


52


Buffer size, M

ping-pong: (M/2,M/2)

Log(

Ove

rflow

Rat

e)

In practise, cost <5%

single memory: M single memory: M/2


53

Wastage Factor,ω R( )M R( ) M̃ R( )–

M R( )-------------------------------------≡

• decreases with M

• decreases with burstiness

• decreases with load

• decreases with number of ports

ω R( )

ω R( )

ω R( )

ω R( )

Some ResultsInput Queued Switch


54




• Fast Buffering

• Speedup



55

Combined Input- and Output-Queueing:

Matching Output Queueingwith Input- and Output- Queueing

How much speedup is enough?

k readsand writes


56

Conventional wisdom suggests:


How much speedup is enough?

A speedupk 2 4–= leads to high throughput


57

Fact To match output queueing, with FIFO input queues:


=?Traffic

Output QueuedSwitch

Combined Input

Switchand Output Queued

Fact To match output queueing, with virtual output queues:

k 21N----–

is necessary and sufficient.=

k N is necessary and sufficient.=


58




• Fast Buffering

• Speedup


Documents

High Performance Switching and Routingtiny-tera.stanford.edu/~nickm/talks/ResearchOverview...Why we need faster switches/routers High Performance Switching and Routing 7 Trafﬁc Inversion