Dual-Plan Bandwidth Smoothing for Layered-Encoded Video

Tong Gan, Kai-Kuang Ma, and Liren Zhang

IEEE Trans. Multimedia, Apr. 2005

Outline Bandwidth smoothing

Dual-plan bandwidth smoothing (DBS) scheme Trunk plan Branch plan

Simulation results

System architecture

Example of BS

Dual-plan bandwidth smoothing scheme Based on the renegotiated constant bit rate (RCBR) service.

The reserved bandwidth limits the peak rate of the schedule. Renegotiation may fail (the reservation request for a higher ban

dwidth may be rejected by the network) Upon a renegotiation failure, the server has no choice but to pa

use the transmission. Use layered video coding to guarantee continuous video playba

ck.

L layers: one base layer and (L-1) enhancement layers. Trunk plan A L(t): deliver all L layers Branch plan Am(t): the branch plan for the mth renegotiation i

nstance rm,j: the renegotiation instances (the jth renegotiation after the

mth rate change)

Example: DBS scheme

b1

b2

b3

b4

b5Trunk plan A L(t)

r1,0r0,0 r1,1 r1,2 r2,0 r2,1 r3,0 r3,1 r4,0 r4,1 r5,0 Time, t

Cum

ula

tive d

ata

si

ze, s

Branch plan A1(t)

Branch plan A3(t)

Branch plan A4(t)

bi = ci

Example: switch plan

b1

b2

b3

Trunk plan A L(t)

r1,0r0,0 r1,1 r1,2 r2,0 r2,1 r3,0 Time, t

Cum

ula

tive d

ata

si

ze, s

b2

b3

Branch plan A1(t)

Procedure Compute the trunk plan

Adjust renegotiation instances of the trunk plan Match frame boundaries

Compute the branch plans Frame boundaries Number of layers being transmitted

Trunk plans Minimum Polyline Smoothing (MPS) a

lgorithm Renegotiation instances

Switching procedure only occurs at frame boundaries.

If the server intends to switch to a branch plan, it must completely send out all L layers of the kth frame and then follows the branch plan to transmit the (k+1)th frame.

Adjust renegotiation instance (1/2)

A L(t)

cm

Cm+1

vm kmΔ

V L(t)

mkm tr 0,

renegotiate before the rate increases

Adjust renegotiation instance (2/2)

A L(t)

cm

Cm+1

vm kmΔ

V L(t)

mkm tr 0, renegotiate after the rate decreases

Possible cases of AL(t)

Branch plans For each bandwidth increasing instance, an associ

ated branch plan should be established. Renegotiation instances

The server should switch back to the trunk plan as early as possible.

However, more renegotiations imply more overhead. Use Tr to control renegotiation frequency.

Renegotiations should only occur at frame boundary. The branch plan continues until

the end of the video, renegotiate successfully, or (the bandwidth requirement of the video) < (the current r

eserved bandwidth).

Example: computing the branch plan Am(t)

rm,0 rm,1 rm,2

V L(t)

A L(t)Vm(t)

U L(t)

Um(t)Am(t)

Scheme (branch plan) The server transmits an equal number of l (l

≦L) layers of each frame. The goal: maximize l Procedure:

For a given l, the corresponding constraint region is constructed.

Apply MPS algorithm to compute a temporary transmission plan.

If the plan satisfies the bandwidth constraint, it is saved as the branch plan;

otherwise, l will be decremented and the procedure is repeated.

Exploit the MPS algorithm to compute A l(t)

Vm(t)

Am(t)V l(t)

Um(t)

U l(t)

r_right

r_left (p-1)Δ (q-1)Δ

Transmit the pth frame to the (q-1)th frame during r_left ≦ t < r_right. A l(t)

Finish line

Starting point

G

Simulation C: link capacity bmin, bavg, bmax: the minimum, average, and maximu

m transmission rate of the trunk plan. Kmin = C / bmax: the number of clients admitted when

the peak-rate reservation is employed. Kavg = C / bavg: the number of clients accepted when

each one can transmit at the average rate. Kmax = C / bmin: the maximum number of concurrent

connections admitted. (base layer only)

Configurations Playback buffer B = 1 MB Minimum time gap of renegotiation interval Tr = 10 sec.

Measurements

Network utilization Enhancement layer discarding

ratio

Video quality degradation interval

Different Tr (1/3)

In heavy load, renegotiation failures occur frequently, limiting the number of enhancement layers to be delivered.

L = 2

Different Tr (2/3)

L = 2

Different Tr (3/3)

L = 2

Different enhancement-layer granularities (1/3)

Tr = 10

Different enhancement-layer granularities (2/3)

Tr = 10

Different enhancement-layer granularities (3/3)

Two cases are the same!

Tr = 10

Playback buffer B = 5 MB (1/3)

?

better than the case B = 1MB

(L=2, 65%)

Playback buffer B = 5 MB (2/3)

better than the case B = 1MB

(L=2, 60%)

Playback buffer B = 5 MB (3/3)

better than the case B = 1MB

(Kmax, 2000 s)

Summary

The performance of DBS can be improved by reducing the minimum time gap of reneg

otiation interval Tr; employing multilayer video encoding wit

h fine granularity; and/or increasing the playback buffer size B.