Mario Gerla and Gianluca Reali

Computer Science Department

University of California, Los Angeles (UCLA)

gerla@cs.ucla.edu

www.cs.ucla.edu/NRL/

IP DiffServ Architecture for Providing Voice/Video QoS Guarantees:

CAC, QoS Routing and Forwarding

Multiple constraints QoS RoutingGiven:

- a (real time) connection request with specified QoS requirements (e.g., Bdw, Delay, Jitter, packet loss, path reliability etc)

Find:

- a min cost (typically min hop) path which satisfies such constraints

- if no feasible path found, reject the connection

Example of QoS Routing

A

B

D = 30, BW = 20D = 25, BW = 55

D = 5, BW = 90

D = 3, BW = 105

D =

5, B

W =

90

D = 1, BW = 90

D = 5, B

W = 90

D =

2, B

W =

90

D = 5, BW = 90D = 14, BW = 90

Constraints: Delay (D) <= 25, Available Bandwidth (BW) >= 30

2 Hop Path --------------> Fails (Total delay = 55 > 25 and Min. BW = 20 < 30)3 Hop Path ----------> Succeeds!! (Total delay = 24 < 25, and Min. BW = 90 > 30)5 Hop Path ----------> Do not consider, although (Total Delay = 16 < 25, Min. BW = 90 > 30)

A

B

D = 30, BW = 20D = 25, BW = 55

D = 5, BW = 90

D = 3, BW = 105

D =

5, B

W =

90

D = 1, BW = 90

D = 5, B

W = 90

D =

2, B

W =

90

D = 5, BW = 90D = 14, BW = 90

Constraints: Delay (D) <= 25, Available Bandwidth (BW) >= 30

We look for feasible path with least number of hops

Benefits of QoS Routing

• Without QoS routing: • must probe path & backtrack; non optimal path, control

traffic and processing OH, latency

With QoS routing:• optimal route; “focused congestion” avoidance• more efficient Call Admission Control (at the source)• more efficient bandwidth allocation (per traffic class)• resource renegotiation possible

The components of QoS Routing• Q-OSPF: link state based protocol; it disseminates

link state updates (including QoS parameters) to all nodes; it creates/maintains global topology map at each node

• Bellman-Ford constrained path computation algorithm: it computes constrained min hop paths to all destinations at each node based on topology map

• Call Acceptance Control • Packet Forwarding: source route or MPLS

OSPF Overview

5 Message Types

1) “Hello” - lets a node know who the neighbors are

2) Link State Update - describes sender’s cost to it’s neighbors

3) Link State Ack. - acknowledges Link State Update

4) Database description - lets nodes determine who has the most recent link state information

5) Link State Request - requests link state information

OSPF Overview

A

B

C

D

E

1

1

2

2

2

3

3

“Link State Update Flooding”

OSPF Overview

“Hello” message is sent every 10 seconds and only between neighboring routers

Link State Update is sent every 30 minutes or upon a change in a cost of a path

Link State Update is the only OSPF message which is acknowledged

Routers on the same LAN use “Designated Router” scheme

Implementation of OSPF in the QoS Simulator

Link State Update is sent every 2 seconds

No acknowledgement is generated for Link State Updates

Link State Update may include (for example):

- Queue size of each outgoing queue (averaged over 10s sliding window)

- Throughput on each outgoing link (averaged over 10s sliding window)

- Total bandwidth (capacity of the link)

Source router can use above information to calculate

- end-to-end delay

- available buffer size

- available bandwidth

Bellman-Ford Algorithm

110

1 2 4 2

3 2

1

2 4

3 5

Bellman Equation : jDih+1

Dh=min[ d(i ,j) + ]

1

3

1

2 4

3 5

D11=0

D21=1

D31=3

D41=00

D51=00

one hop

1

3

1

2 4

3 5

D12=0

D22=1

D32=2

D42=11

D52=5

10

1

2

two hops

iDh

h

D

h*three hops1

1

2 4

3 5

D13=0

D23=1

D33=2

D43=7

D53=4

1

2

2

3

B/F Algorithm properties

• B/F slightly less efficient than Dijkstra ( O(N) instead of O (NlgN) )

• However, B/F generates solutions by increasing hop distance; thus, the first found feasible solution is “hop” optimal (ie, min hop)

• polynomial performance for most common sets of multiple constraints (e.g., bandwidth and delay )

CAC and packet forwarding

• CAC: if feasible path not found, call is rejected; alternatively, source is notified of constraint violation, and can resubmit with relaxed constraint (call renegotiation)

• Packet forwarding: (a) source routing (per flow), and (b) MPLS (per class)

Application I: IP Telephony

• CAC at source; no bandwidth reservation along path• 36 node, highly connected network• Trunk capacity = 15Mbps• Voice calls generated at fixed intervals• Non uniform traffic requirement• Two routing strategies are compared:

Minhop routing (no CAC)

QoS routing • Simulation platform: PARSEC wired network simulation

QoS simulator: priority queues

PRIORITY 1

PRIORITY 2

PRIORITY 3

SIGNALLING (OSPF)

QoS Sources (Voice)

Non-QoS Sources (TCP)

RESERVED

QoS Simulator: buffer allocations

CBQ

1 MB

1 MB

1 MB

1 MB

1 MB

1 MB

QoS Simulator: Voice Source Modeling

TALK SILENCE

1/ = 352 ms

1/ = 650 ms

1 voice packet every 20ms during talk state

0

12

18

5421 3

6

30

24

14

87

13 1716

10 11

15

9

19

26 27

34

28

2220 21 23

33 35

2925

31 32

50 Km

15 Mbit/sec

Simulation Parameters

10 Minute Simulation Runs

Each voice connection lasts of 3 minutes

OSPF updates are generated every 2 seconds (30 minute OSPF update interval in Minhop scheme)

New voice connections generated with fixed interarrival

Measurements are in STEADY-STATE (after 3 minutes)

100 msec delay threshold

• 3Mbit/sec bandwidth margin on each trunk

NON-UNIFORM TRAFFIC GENERATION

Light Load Experiment

Voice Call interarrival T = 375ms; No TCP traffic

MC Bellman Ford Minhop

#calls generated: 1120 1120

#calls accepted: 1120 1120

% packet lost: 0% 0%

% packet above Delay Thresh: 0% 0%

% packet below Delay Thresh: 100% 100%

DELAY THRESHOLD = 100ms

MC Bellman Ford Minhop

#calls generated: 2800 2800

#calls accepted: 2800 2800

% packet lost: 0.0% 22.61%

% packet above Delay Thresh: 0.0% 48.83%

% packet below Delay Thresh: 100.0% 28.56%

DELAY THRESHOLD = 100ms

Heavy Load Experiment

Voice Call interarrival T = 150ms; No TCP traffic

0

12

18

5421 3

6

30

24

14

87

13 1716

10 11

15

9

19

26 27

34

28

2220 21 23

33 35

2925

31 32

50 Km

15 Mbit/sec

MINHOP ROUTING

0

12

18

5421 3

6

30

24

14

87

13 1716

10 11

15

9

19

26 27

34

28

2220 21 23

33 35

2925

31 32

50 Km

15 Mbit/sec

MINHOP ROUTING

0

12

18

5421 3

6

30

24

14

87

13 1716

10 11

15

9

19

26 27

34

28

2220 21 23

33 35

2925

31 32

50 Km

15 Mbit/sec

QoS ROUTING

0

12

18

5421 3

6

30

24

14

87

13 1716

10 11

15

9

19

26 27

34

28

2220 21 23

33 35

2925

31 32

50 Km

15 Mbit/sec

QoS ROUTING

(src, dst) Throughput

(3, 32) 471353bps = 471Kbps

(4, 32) 578806bps = 579Kbps

(5, 32) 729800bps = 730Kbps

(10, 32) 796863bps = 797Kbps

(16, 32) 404768bps = 405Kbps

Sum 2981590bps = 3.0Mbps

- TCP source/destination pairs same as for voice

- FTP infinite backlog

- TCP Reno

Heavy Load Experiment with TCP

Voice Call interarrival T = 150ms

TCP steady state performance with QoS Routing

Throughput of FTP traffic running concurrently with Minhop routed voice traffic

0

5000000

10000000

15000000

200000000.

1

48.7

97.3

146

195

243

292

340

389

438

486

535

583

Time (sec)

Thro

ughp

ut (b

its/s

ec) (8, 20)

(3, 32)

(4, 32)

(5, 32)

(10, 32)

(16, 32)

(0, 34)

Throughput of FTP traffic running concurrently with MC Bellman-Ford routed voice traffic

0

5000000

10000000

15000000

200000000.

1

48.7

97.3

146

195

243

292

340

389

438

486

535

583

Time (sec)

Thro

ughp

ut (b

its/s

ec) (8,20)

(3, 32)

(4, 32)

(5, 32)

(10, 32)

(16, 32)

(0, 34)

Scalability of OSPF with QoS enhancements

0

20

40

60

80

100

120

140

0 50 100 150

number of nodes

Kbi

t/sec Maximum OSPF control

traffic on a link

• OSPF packet size was 350 bytes

• OSPF (LSA) updates were generated every 2 seconds

• Measurements were performed on a “perfect square grid” topology

75ms voice call generation rate

Effect of OSPF update interval on call acceptance control of IP Telephony traffic

0

1000

2000

3000

4000

5000

6000

0.1 0.5 2 10 60

OSPF (LSA) update interval (sec)

# of

voi

ce c

alls

Accepted

Rejected

Offered Load

Application II: MPEG video

• Measurement based CAC (which worked well for IP telephony) is now replaced by reservation based CAC

• RSVP type signaling required• Effective bandwidth & buffer reservations• 36 node grid-type topology• Trunk capacity = 5.5 Mbps• Inputs = Measured MPEG traces• QoS guarantees: no-loss; delay < Tmax• Simulation platform: PARSEC wired network simulation

DLB #1

DLB #2

DLB #n

C

B

S1

S2

Sn

EDGE ROUTER

MPEGSOURCES

NETWORK ACCESS SECTION

DLB1

DUAL LEAKY-BUCKET REGULATOR

rs Ps

DLB2

BTS

bBrP

cPTS

ss

s

maxDc

b

bBrP

cPTS

ss

s

BTS

b

cPsrs c0

b0

LOSSLESS ALLOCATION

maxDc

b

SCHEDULING

Token buffer

Sustainable rate Peak rate

Input rate Output rate

NETWORK CORE SECTION

• Each router keeps a resource bookkeeping table:

OUGOING LINK; AB;AC ;NS1 ,NS2 ,...NSn ;

AB: Available Buffers; AC: Available Capacity

NSk , k=1,…,n: the number additional flows belonging to

class Sk that the router can accept within QoS constraints

• Each router loads the AB, AC and NS information in the

OSPF Link State Update packet

. Link State Parameters updated at each reservation

APPLICATION ORDER

SOURCE NODE DESTINATION NODE

PATH CACY/N

REJECTION

NODE

1 0 35 1 18 19 20 21 22 28 29 35 Y  

2 24 13 25 19 20 14 13 Y  

3 19 20 25 31 32 26 20 Y  

4 1 18 18 Y  

5 34 35 28 29 35 Y  

6 

18 21 

19 20 21 N 19

1 7 8 14 20 21 Y  

7 29 35 NO PATH FOUND    

8 9 18 8 7 1 18 N 1

8 14 20 19 18 Y  

9 27 21 26 20 21 N 20

26 20 14 8 9 15 21 Y  

10 22 28 28 Y  

11 22 28 21 20 19 25 24 34 28 Y  

SIMULATION RESULTS Bandwidth/link: 5.5 Mbps unidirectionalTmax: 0.1 s Effective bandwidth: 2.6 MbpsEffective buffer: 260 KB (no buffer saturation)

0

12

18

5421 3

6

30

24

14

87

13 1716

10 11

15

9

19

26 27

34

28

2220 21 23

33 35

2925

31 32

1

2

3

4

5

6

6

7

8

8

9

9

10

11

0

12

18

5421 3

6

30

24

14

87

13 1716

10 11

15

9

19

26 27

34

28

2220 21 23

33 35

2925

31 32

1

2

3

4

5

6

6

7

8

8

9

9

10

11

SIMULATION RESULTSno reservations; measurement based CAC

Bandwidth/link: 5.5 Mbps unidirectionalTmax: 0.1 s

CONNECTIONORDER

SOURCE NODE

DESTINATIONNODE

PATH CACY/N

REJECTION

1 0 35 1 18 19 20 21 22 28 29 35 Y

2 24 13 25 19 20 14 13 Y

3 19 20 20 Y

4 1 18 18 Y

5 34 35 28 29 35 Y

6 18 21 19 20 21 Y

7 29 35 35 Y

8 9 18 8 7 1 18 Y

9 27 21 26 20 21 Y

10 22 28 28 Y

11 22 28 28 Y

Video Only Result Comparisons Class based QoS routing with reservation

vs. Measurement based QoS routing without reservation

Bandwidth/link: 5.5 Mbps unidirectionalTmax: 0.1 s, Duration 10 min

Class Based QoS routing withreservation

Measurement based QoS routingw/o reservation

Number of packets sent 551820 607002

Percentage of packets lost 0% 0%

Percentage of packets received: 100% 100%

Max delay for video packets: 0.0806 s 0.3725 s

Percentage of Packets exceedingdelay threshold

0% 0.8%

Number of connection requests 11 11

Number of rejections 1 0

Number of routing retries 3 N/A

Effects of Reservation Algorithms on TCPScenario: 5 ftp (TCP) applications run between

Node 0 (server) and Node 35(client) over 1 min period. Case 1: TCP only

Case 2: TCP & video using class based algorithms with reservationCase 3: TCP & video using measurement based algorithms w/o reservation

TCP

SESSION

SOURCE

NODE

DEST.

NODE

TIME

(s)

RECEIV.

DATA

case 1 (bytes)

RECEIV.

DATA

case 2 (bytes)

RECEIV.

DATA

case 3 (bytes)

1 0 35 60 7639208 5870767 4251468

2 0 35 60 7604068 5853718 4035787

3 0 35 60 7604068 5808244 3968244

4 0 35 60 7604068 6220530 4360450

5 0 35 60 7604068 5817280 4053859

Effects of Reservation Algorithms on TCPContinued

Number ofbottleneck

videoconnections

Availablethroughput

for TCPGood put Utilization

TCP only 0 330Mb 304.44Mb 92.2%

TCP &Video no

reservation4 185.32Mb 165.35Mb 89.2%

TCP &Video withreservation

2 257.66Mb 236.56Mb 91.8%

Conclusions• QoS routing beneficial for CAC, enhanced routing,

resource allocation and resource renegotiation• Can efficiently handle flow aggregation (Diff Serv)• Q-OSPF traffic overhead manageable up to

hundreds of nodes• Can be scaled to thousands of nodes using

hierarchical OSPF• Major improvements observed in handling of IP

telephony and MPEG video• MPEG video best served via reservations

Future Work

• Use of MPLS instead of source routing• extension to hierarchical OSPF• extension to Interdomain Routing• extension to multiple classes of traffic • Statistical allocation of MPEG sources

Throughput of FTP traffic running concurrently with Minhop router voice traffic

(detail)

01000000200000030000004000000500000060000007000000

0.1

47.6

95.1 143

190

238

285

333

380

428

475

523

570

Time (sec)

Thro

ughp

ut

(bits

/sec

)

(16, 32)

(3, 32)

(4, 32)

(5, 32)

(10, 32)

Throughput of FTP traffic running concurrently with QoS routed voice traffic

( detail )

0

1000000

2000000

3000000

4000000

5000000

6000000

70000000.1

47.4

94.7

142

189

237

284

331

379

426

473

520

568

Time (sec)

Th

rou

gh

pu

t (b

its/s

ec)

(16, 32)

(3, 32)

(4, 32)

(5, 32)

(10, 32)

75ms voice call generation rate

Effect of OSPF update interval on voice packet delivery rate

0

20

40

60

80

100

120

0.1 0.5 2 10 60

OSPF (LSA) update interval (sec)

deliv

ery

rate

(%)

% packets below100ms delaythreshhold