Flow and Congestion Control for Reliable Multicast Communication

In Wide-Area Networks

Supratik Bhattacharyya

Department of Computer Science

University of Massachusetts Amherst

Talk Overview

General Problem

Single-rate source-based congestion control (CC) :

the Loss Path Multiplicity problem

a scalable and “fair” congestion control approach

a prototype implementation for active networks

Multi-rate flow-controlled bulk data transfer

Future Research Ideas

Flow/Congestion Control in Wide-Area Networks

Congestion Control short term : adapt

transmission rate to changing traffic conditions.

Flow Control : longer term : tailor rate

to available capacity

End-to-end approach suitable for today’s networks

Internet

Data

Data

Source

Receiver

Feedback

Feedback

Multicasting

My focus : one-to-many reliable multicasting

Network nodes replicate data packets

Network bandwidth used efficiently

Source

R1

R2R3

R4

Router

Multicast Flow/Congestion Control : a hard problem

Challenges - many rcvrs, many network paths :

Heterogeneity

– links, receiver capabilities

Scale– feedback implosion

Fairness – how to share bandwidth

with unicast: end-to-end feedback

Source

R1 R4R3R2

Talk Overview

General Problem

Single-rate source-based congestion control (CC) :

the Loss Path Multiplicity problem

a scalable and “fair” congestion control approach

a prototype implementation for active networks

Multi-rate flow-controlled bulk data transfer

Future Research Ideas

Feedback Aggregation

Challenge : How to aggregate feedback into single rate control decision

loss indications (LI)

filterfilter Rate controlRate control

algorithmalgorithm

congestion signal (CS)

rate change

Congestion signals (CS): filtered versions of loss indications (LI) : congestion signal probability filters can be distributed

Problem : Loss Path Multiplicity (LPM)

Copies of same packet lost on many network paths

Set of receivers treated as single aggregate receiver

Example :

N : no. of receivers

p : loss prob. on link to each rcvr.

: congestion signal probability

) 1 (1 p N

R2

?

R1 R3

LI LI

1 as N

How Severe is the LPM Problem?

Severe degradation in throughput with -

no. of receivers independent losses

0

2

4

6

8

10

12

0 50 100150200 250300350400450 500

No. of Receivers

f=0.1

f=0.5

f=0.9

p=0.05

Example :

p

f : fraction of end-to-end loss on independent link

. . .

fpend-to-end loss prob. =

Feedback Aggregation/Filtering :Related Work

Restrict response to one LI per time interval T Montgomery 1997

Restrict response to subset of receivers :

choose K receivers out of N as representatives

Delucia et al. 1997

Reduce response to each LI :

Golestani, Bhattacharyya 1998, Delucia et al. 1997

Q : How much bandwidth should a multicast session get?

Background : “Fair” Bandwidth Sharing

Challenge : How to achieve “fair” sharing among multicast and unicast sessions

Multicast allocation according to “worst” end-to-end path

Multicast session shares equally with a unicast session on its “worst” end-to-end path.

L1 - 1 Mbps, L2 - 2 Mbps

Ucast1

L2

L1

Mcast

Mbpsr 5.0 1Ucast

Mbpsr 5.0 Mcast

Mbpsr 5.1 Ucast2

Ucast2

L2

Background : End-to-end Rate Control Algorithms

: rate after i-th update

Additive increase, multiplicative decrease :

on congestion signal :

else, per T :

We derive average session throughput B

1 1 rr ii

)11( 1 Crr ii

ri

, ,TCCTB

Solution to LPM Problem : Our Approach

Identify (estimate) “worst” receiver

Respond to LIs from only “worst” receiver

prevents throttling of multicast transmission rate

allows fair bandwidth sharing

Bhattacharyya, Towsley, Kurose. Infocom ‘99

. . .

Modified Star

0

2

4

6

8

10

12

14

16

0 5 10 15 20 25 30 35 40 45 50

No. of Representatives (K)

representativeapproach

worst rcvr. approach

Simulation of LPM Solution

Simulation Settings: 5 multicasts over L1, L2, each

tracks L1 A : 5 unicasts over L1, 5 over L2 B : 5 more unicasts on L1 C : same as B, each multicast

tracks L2 instead

Example topology :

L1 L2

L1, L2 : 300 pkts/sec

Sources

Rcvrs

mcast ucastover L1

ucastover L2

Simulation Settings

A

B

C

29.8 30.2 30.3

Throughput (pkts/sec)

20.9

30.0

20.9 39.9

17.1 30.5

Rcvrs

Realizing the Worst Receiver Approach

Use end-to-end loss probability estimates :

N rcvrs - rcvr i reports Xi losses out of S pkts

choose rcvr with highest no. of losses

Worst Estimate-based Tracking (WET)

WET is sensitive to S : large S good estimate small S likely to choose wrong receiver as worst

Q : What can we do for small S ?

Challenge : How to identify the worst receiver?

Current Work : Robust Congestion Control

Our Idea : On LI from receiver i, reduce rate with probability

Linear Proportional Response (LPR) :

Observation : small S : LPR more robust

S : LPR allocates more than fair share to multicast session !

Example : 2 receivers, loss prob. 0.05 and 0.10

13

14

15

16

17

18

19

20

21

0 50 100 150 200 250 300 350 400 450 500

No. of observations (S)

LPR

WETFair Share

Ongoing Work

Related : Random Listening Algorithm (RLA) [Wang98]

Result : Our approach (LPR) provides tighter upper bound on r

LPR :

RLA : Nr

4 Nr 0

1

2

3

4

5

6

0 5 10 15 20 25 30

No. of receivers (N)

RLA

LPR

A Prototype of Worst Receiver Approach for Active Networks

“Worst” receiver has largest value of

Active Servers : aggregate feedback

help in identifying “worst” receiver

p : loss prob. estimateRTT : round trip time estimate

Source

R1 R2 R3 R4

AS1 AS2

Our Rate Control Algorithm

pRTTv

v1 v2v3 v4

v1 v4

Worst : R1

Talk Overview

General Problem

Single-rate source-based congestion control (CC) :

the Loss Path Multiplicity problem

a scalable and “fair” congestion control approach

a prototype implementation for active networks

Multi-rate flow-controlled bulk data transfer

Future Research Ideas

Flow-controlled Bulk Data Transfer : Overview

Challenge : reliable delivery of finite volume

of data diverse receive-rates

Goal : minimize average completion

time

Approach : multiple IP multicast groups

(channels)

R1=1 R2=2 R3=3

Bhattacharyya, Kurose, Towsley, Nagarajan. Infocom ‘98

R4=4

Flow-controlled Bulk Data Transfer

2 pkts/sec4 pkts/sec

1 pkt/sec

a

b

c

d

b dr1 = 1

r2 = 1

r3 = 2 c d

R1R2

R4

a

a

a

b

b

cd

R1,R2,R4

R2,R4

R4

Q : How to : assign channel rates? assign receivers to channels? partition data among

channels?

Assumptions : error-free channels known, static receive-rate

constraints

Solution with unlimited channels :

minimizes average completion time

minimizes bandwidth

Flow-controlled Bulk Data Transfer

2 pkts/sec4 pkts/sec

1 pkt/sec

a

b

c

d

b dr1 = 1

r2 = 1

r3 = 2 c d

R1

R2R4

a

a

a

b

b

cd

R1,R2,R4

R2,R4

R4

Q : How to : assign channel rates? assign receivers to channels? partition data among

channels?

Assumptions : error-free channels known, static receive-rate

constraints

Solution with unlimited channels :

minimizes average completion time

minimizes bandwidth

c

cd

Flow-controlled Bulk Data Transfer

2 pkts/sec4 pkts/sec

1 pkt/sec

a

b

c

d

b dr1 = 1

r2 = 1

r3 = 2 c d

R1

R2R4

a

a

a

b

b

cd

R1,R2,R4

R2,R4

R4

Q : How to : assign channel rates? assign receivers to channels? partition data among

channels?

Assumptions : error-free channels known, static receive-rate

constraints

Solution with unlimited channels :

minimizes average completion time

minimizes bandwidth

c

cd

d

b

Limited Number of Channels

Static rate assignment :

Q : Given K channels and N (>K) receive rates, which K rates to match?

Approach : minimize average completion time

dynamic programming solution - O(N3 K)

Dynamic rate assignment : reassign rates when faster receivers finish optimization problem too hard Our approach : Simple heuristics

Heuristics for Channel Rate Assignment

Fastest Receivers First (FRF)

Slowest Receivers First (SRF)

Equal Partitions (EQ) distribute rates “smoothly” over entire

range of receive rates

Maximize Utilized Capacity (MUC) :

allocate channel rate to maximize sum of rates at which unfinished receivers receive

dynamic programming solution

no. of receivers

receive rates

Example :

Choose rates for 3 channels

EQ:

MUC:

G1G2

G3

G4

Summary of Results

Average Completion time scales well :

200

1000

1500

0 100 200 300 400

No. of Receivers (X)

SRF

STATIC

MUC

IDEAL

Small no. of channels reqd :

200

1000

2600

0 2 4 6 8 10 12 14 16

Number of Channels (K)

SRF

STATICMUC

IDEAL

Summary of Contributions

Single-rate source-oriented multicast CC : identified and studied Loss Path Multiplicity

problem proposed a scalable and “fair” congestion control

approach current work : robust congestion control schemes developing a prototype implementation for active

networks

Developed efficient algorithms for flow-controlled multicast of bulk data 1

1 : U.S. patent pending

Other Interesting Projects

RMTP : A Reliable Multicast Transport Protocol 1

A Class of End-to-end Congestion Control Algorithm for the Internet 2

Design and Implementation an Adaptive Data Link Layer Protocol for an ATM Wireless LAN

2 : Golestani and Bhattacharyya. ICNP ‘98

1 : Paul, Sabnani, Lin, Bhattacharyya. JSAC 97

Future Research Ideas

Immediate : prototype CC protocol for

active networks robust multicast CC

schemes

Short Term : multicast CC for continuous

media CC with enhanced network

support

Looking ahead :

network measurements support for adaptive

applications active services differentiated services

Open to new ideas and collaborations !