Congestion Control in Distributed Congestion Control in Distributed Media StreamingMedia Streaming

Lin MaWei Tsang Ooi

School of ComputingSchool of ComputingNational University of SingaporeNational University of Singapore

IEEE INFOCOM 2007IEEE INFOCOM 2007

outlineoutline

Introduction Distributed Media Streaming (DMS) Task-level TCP Friendliness Framework of DMSCC Throughput Control Congestion Control Simulation and Conclusions

IntroductionIntroduction

Distributed Media Streaming (DMS) coined by Nguyen and Zakhor

(IEEE transaction on multimedia 2004) A client receives a media stream from multiple

servers simultaneously multiple media flows may or may not pass

through the same bottleneck

Aggregate congestion control as task-level congestion control

DMS

Distributed Media Streaming client receives a media stream from multiple

servers simultaneous improves robustness allows aggregation of bandwidth among peers

Sender 1

receiver

Sender 2

Internet

0.8Mbps

0.4Mbps

1.2Mbps

ScenarioScenario

Receiver-driven

receiver Sender 1

Sender 2

Estimates round trip timeEstimates round trip time

Estimates round trip timeEstimates round trip time

Sender1’s RTT

Sender1’s RTT

Sender2’s RTT

Sender2’s RTT

Estimates Estimates sender’s loss sender’s loss ratesrates

Decide each sender’s Decide each sender’s sending rates (RAA)sending rates (RAA)

control packet

control packet

Determine the next Determine the next packet to be sent (PPA)packet to be sent (PPA)

Determine the next Determine the next packet to be sent (PPA)packet to be sent (PPA)

Send packet

Send packet

DMS protocol

Rate allocation algorithm (RAA) Run at receiver

minimize the probability of packet loss splitting the sending rates appropriately across

the senders

Packet partition algorithm (PPA) Run at individual senders

Ensure every packet is sent by only one sender minimize the probability of packet arriving late

Bandwidth Estimation

TCP Friendly rate control algorithm (TFRC)

B : current available TCP-friendly bandwidth between each sender and receiver

Trto : TCP timeout

R : estimated round-trip time in seconds

p : estimated loss rate

S : TCP segment size in bytes

232183

332

ppp

Tp

R

SB

rto

Rate Allocation Algorithm

Receiver computes the optimal sending rate for each sender based on loss rate and bandwidth

Subject to

During the interval (t, t + ∆)F(t) : total number of loss packet

L(i, t) : estimated loss rates S(i, t) : estimated sending rates

Sreq(t) : is the required bit rate for the encoded video

B(i, t) : TCP-friendly estimated bandwidth

tiStiLtFN

i

,,1

tiBtiS ,,0

tStiS req

N

i

1

,

Rate Allocation Algorithm

Sort the senders according to their estimated loss rates from lowest to highest

Assign each sender its available bandwidth, until the sum of their available bandwidths exceeds the bit rate of the encoded video

Rate Allocation Algorithm Suppose there exists a rate allocation with M' senders in which

F(t) is min, say F'(t) for some 1≤ i’< M‘ ,but S(i’,t)≠B(i‘,t) ,implies that S(i',t) < B(i',t) Proof ： Allocating some of the bit rate from sender M' to

i' we can achieve smaller loss rate than F'(t)

Ω = min[ S(M',t) , B(i',t)−S(i',t)]

F*(t) < F'(t)

Each sender receive control packet from the receiver through a reliable protocol in the following format

Di : Denote the estimate delay from each sender to receiver

Si : Denote the sending rate for each sender

Sync : Synchronization sequence number

Packet Partition Algorithm

D1 D2 S1 S2 Sync

Packet Partition Algorithm

Only the sender who maximizes Ak’(j, k) will be assigned to send kth packet Each sender computes Ak’(j, k) for each packet k for

itself and all other senders

Ak’(j, k) = Tk’(k) – [nj,kσ(j)+2D(j)]

D(j) : estimated delay

nj,k : Number of packets already sent since k’ to packet k

σ(j) = P/S(j) : Sending interval between packets for sender j

(P : packet size)

Tk’(k) : Time difference between arrival and playback time

of kth packet (not affect the value)

arrival time of the kth packet sent by sender j

Packet Partition Algorithm

Among all senders j=1,…N, the one that maximizes Ak’(j,k) is assigned to send kth packet

Each sender keeps track of all the values of Ak’(j,k) for all N senders, and updates every time a packet is sent

Task-level TCP Friendliness

Congestion Control in Single-flow streaming Flow f is TCP-friendly:

B = BTCP

Comparable network environment same loss rate, RTT and packet size

0 2

1

A B

3

f

TCP

TCPTCP RTTBRTTB

Task-level TCP Friendliness

Congestion Control on aggregate

O: set of flows in flow aggregatebi :throughput of flow fiRtti :round trip time of flow fi

0 2

1

A B

3

f1, f2, and f3

TCP

TCPTCP

Ofi

ii RTTBrttb

Task-level TCP Friendliness

Congestion Control in DMS

0

2

1

3

RA

B

C

TCP

TCP

Set of DMS flows to be controlled depends on where congestion appears

jj TCPTCP

ljfi

ii RTTBrttb

Framework of DMSCC

Congestion location

Throughput control

Receiver side

Sender side

Congestion Control

Algorithm

DMS Flows

Update the increasing factor of

AIMD

Congestion Location

An ideal solution to locate a congestion congestion causes a packet loss on a DMS flow

should be able to tell which link is congested

congestion subsides should sense it, so that the regulation on the DMS flows

previously imposed can be lifted

Such ideal solution is difficult simultaneous congestion on different links in the

tree same flow might experience congestion on

different links

Congestion Location

One link congestion (Rubenstein et al. [15].) compare the cross-correlation of two flows and

the auto-correlation of one of them CorrTest(i,j) to denote the correlation test of

Rubenstein applied on flow i and flow j the two flows share a bottleneck

Return1 No shared bottleneck is detected

Return 0

Congestion Location

Idea: Packets passing through same Point of Congestion(POC) close in time experience loss and delay correlations

Using either loss or delay statistics, compute two measures of correlation: Mc: cross-measure (c between flows) Ma: auto-measure (c within a flow)

such that if Mc < Ma then infer POCs are separate else Mc > Ma and infer POCs are shared

Mc = C(Delay(i), Delay(i-1))

Ma = C(Delay(i), Delay(prev(i)) time

i-4

i-2

i

i-1

i-3

i+1

Flow 1 pkts

Flow 2 pkts

Congestion Location

Algorithm 1: find the shared bottleneck

0

2

1 R

A

BOutput ={ A-B }

Congestion Location

Algorithm 2: On every packet loss

C = {A-B, 0-A}

H = {A-B, 0-A, 0-A, A-B, A-B, 0-A,A-B,A-B …}

0

21

AB R

Throughput Control

W : size of congestion windowα : increasing factorL : period (in RTT) between every two

packet lossesp : packet loss rate

83

22

3 2WL

WS

pS

1

pW

3

8

Total number of packets received during that period

Mathis Equation[14]loss rate is p, then for every 1/p

packets, one packet is lost.

Throughput Control

If we want a DMS flow to have β times the throughput of a TCP flow

For the throughput of a DMS flow to be β times of a conformant TCP flow, we need to set its increasing factor α to β^2

2

3

8

4

3

3

8

4

3

pp

WW TCP

Throughput Control

Update Alpha: according to dominant bottleneck Algorithm 3: Update Increasing Factors

0

2

1

A

B R

C = {A-B, 0-A}

C ’ = {A-B}A-B dominates the other link

Throughput Control

Bottleneck Recovery Congestion subsides and there is no more

packet loss, we need to reset αi to 1 DMS flow to fully utilize available bandwidth

A timer is refreshed when packet loss is detected If no packet loss is detected within t seconds

The increasing factors of all DMS flows are reset to 1

Simulation

Link Start End

0-A 0 100

A-B 150 200

B-C 50 150

C-R 250 350

0

21

3

RAB

C

Conclusions

Congestion control in DMS system Task-level TCP-friendliness Locate congestion in a reverse tree topology

Relies on Rubenstein’s method Control the throughput of a DMS flow using

AIMD loop combined throughput on the bottleneck is TCP-

friendly Based on Mathis equation