An Efficient Flow Control Plan for End-To-End Delivery of Pre-stored Compressed Videos

Sheau-Ru Tong and Sho-Chi LeeDepartment of Management Information SystemsNational PingTung University of Science and TechnologyPingTung, Taiwan, R.O.C.srtong @mail.npust.edu.tw

End-to-End Video Flow

Disk-based video Server

buffer

Video blocks(hundreds KB)

ClientStation

buffer

Video frames(tens KB)

Cells(tens B)

Network channel

• Disk-based video server

• Network channel• Client Station

• Objective: minimize buffer space, network bandwidth, initial delay

• Peer buffer: for reshaping traffic from one style to another

Feasible belt

Accumulative Traffic Model

• Let the time of data first being available be the reference time t=0.

• Accum. data amount of client: fc(t).

• Client buffer: Bc.

• Initial delay d.

d

Bc

fx(t).

•Accum. data amount of server: fs(t).•Server buffer: Bs.

fs(t)+ Bs

fs(t).

Bs

fx(t)+ Bc

fl(t)fn(t) fu(t), where fl(t)=max{fc(t),fs(t)-Bs} and fu(t)=max{fs(t), fc(t)+Bc}

Time

Dat

a am

ount

Accumulative Traffic Model (Contd.)

• How to keep the feasible belt contiguous? ( for existence of any fn)

• One possible way: keep a vertical spacing of between fu and fl, where =max transient increment of fl.

1. Non-zero startup delay d,

2. Bs, Bc ,

3. f s (t) – f c (t) Bs + Bc – , 0 t and

4. f s (t) – f c (t) , 0 t tq-1,

where max{umax, vmax}.

Necessary condition

Proposed Two-Phase Flow Control

Server fs

Network fn

Client fc

Phase 1: Peer-level Compute fs and fc

Phase 2: Network-level Compute fn

Phase1: Peer-Level

• Server: supplies data at specific CBR.

• Client: consumes data at VBR

• Average rate: rserver> rclient

• Threshold- test:

– Server pauses flow when the amount of backlog data is higher than a threshold =Bs+Bc--vmax. What is the min ?

• Server usually stores video by striping video blocks among regions (n).

• Stalled resumption:“Each server stop takes one or more disk cycles ().”

Disk cycle

regions

Disk head

Phase1: Peer-Level (contd.)

: supplies sufficient inventory for

– client to consume during the server stalling period (fc

ue(/n+)).

Bs+Bcfcue(/n+)+++vmax,

Bs, Bcd time for building up at last backlog.

Proposition

Stalling period

- keeping a spacing between fs and fc as of the stalling period ().

fcue(/n+))

fs(t)

fc(t)

FinalPreliminary

Phase 2: Network-Level

• Network delivers data based on a best-effort discipline: (simple & implementable)

– A m-rate channel is given.

– Channel transmits data when (1) server buffer has data and (2) client buffer has free space. Otherwise, it turns off.

• What is the min m?

• Preliminary fn(t): seeking min slop in every possible situation.

• Final fn(t): using the peak slop found in Preliminary fn(t).

Peak m

Performance Evaluation

• Trace four video tracks:

– one movie- Pbride (rc=1.15Mbps)

– one news- CNN (rc =1.15Mbps)

– one advertisement-Advert (rc =0.48 Mbps)

– one lecture- Lec (rc =0.34Mbps)

• Two disk cycle length:=2 and 4 seconds.

• Four server rates: rs=0.75, 1.5, 3.0 and 4.5 Mbps

Performance Evaluation (contd.)rs (Mbps)

Video track0.75 1.50 3.00 4.50

Name rc

(Mb/s)d

(33ms)Bt

d Bt

d Bt d Bt

Advert 0.483 105 0.717 60 0.764 60 1.027 60 1.345

CNN 1.147 5340 17.05 60 0.711 60 1.014 60 1.332

Lec 0.337 60 0.319 60 0.478 60 0.796 60 1.114

Pbrid 1.147 5340 17.05 60 0.703 60 1.009 60 1.327

rs (Mbps)Video track

0.75 1.50 3.00 4.50

Name rc

(Mb/s)d

(33ms)Bt

d Bt

d Bt d Bt

Advert 0.483 150 1.076 120 1.289 120 1.852 120 2.488

CNN 1.147 5400 17.57 120 1.390 120 2.008 120 2.644

Lec 0.337 120 0.601 120 0.919 120 1.555 120 2.190

Pbrid 1.147 5400 17.57 120 1.378 120 2.000 120 2.636

=2

=4

Bt=Bs+Bc

Performance Evaluation (contd.)

Lecture

89500000

90500000

91500000

92500000

93500000

94500000

95500000

96500000

97500000

98500000

1 14 27 40 53 66 79 92 105

118

131

144

157

170

183

196

209

222

235

248

261

274

287

300

313

326

339

352

365

378

391

404

417

430

443

456

469

482

495

frame#

accu

mul

ated

data

size

Fc

Fs

Fc+B c

Fs-Bs

Bandwid t h

• Example of Lec video track.

Performance Evaluation (contd.)

• Given Bt, there is a range of Bc and Bs, which results in the same min network bandwidth m.• Let Bs=+* and Bc=+(1-)*, where=max{0, Bt-2}.

Performance Evaluation (contd.)

Video Tracks =2 =4

Name rc

(Mbps)rs

(Mbps)m

(Mbps)Stable m

(Mbps)Stable

Advert 0.483 0.75 0.678 15%~75% 0.583 25%~70%

CNN 1.147 1.5 1.143 35%~99% 1.141 42%~96%

Lecture 0.339 0.75 0.386 30%~95% 0.382 40%~90%

Pbride 1.147 1.5 1.144 24%~99% 1.141 24%~96%

• Min network bandwidth m for all cases. (0.38~1.14 Mbps)

Conclusion

• The end-to-end flow control is performed by a two-level approach:

– Peer level: compute fs and fc

– Network level: compute fn.

– Minimizing buffer space (<1.4 MB) and network bandwidth ( average video rate).

• Implementation:

– Use three counters, server_data, net_data and client_data at the server site to keep tracks of all plans.

– Enforce threshold- (peer-level) and best-effort (network-level) principles at the run time.

• Open issue: How to perform flow control for interactive application?