A Monotonic-Decreasing Rate Scheduler for Variable-Bit-Rate Video Streaming Hin-lun Lai IEEE Transactions on Circuits and System for Video Technology, February 2005 Yiu-bun Lee Lian-kuan Chen
Jan 20, 2016
A Monotonic-Decreasing Rate Scheduler for
Variable-Bit-Rate Video Streaming
Hin-lun Lai
IEEE Transactions on Circuits and System for Video Technology, February 2005
Yiu-bun Lee Lian-kuan Chen
Outline
Introduction MDR (Monotonic-Decreasing Rate)
Scheduler– Admission Complexity– Peak Transmission Rate– Client Buffer Requirement
Aggregated MDR Scheduler Conclusion
VBR Video
VBR Video Smoothing
B(t) = A(t) + b
B(t) ≧ S(t) ≧ A(t)
Upward Adjustment
Downward Adjustment
Optimal Smoothing Method
MVBA : minimal bit-rate variance MCBA : minimal change rate
Problem
Bandwidth reservation fail : additional bandwidth is not available at upward adjustment
Bandwidth reservation processing delay– Network topology– Reservation protocol – Loss of control message
MDR Scheduler
Principle : – eliminate upward adjustment and only use
downward adjustment.
ri > rj , for j > I
– Provide guaranteed video delivery Advantage : solve previous two problems
– Have sufficient system bandwidth– Adjustment processing time will not be critical
Disadvantage : need more client buffer
MDR Scheduler
{ri , Ti | i = 1, 2, …n}
Admission Complexity
Notation
– U : System capacity
– ui : System utilization at time i
– w : video length (seconds)
– vj : video transmission bit-rate ( j = 0, 1, …,w-1 )
– A : video startup time
•General Optimal Smoothing
•ui = ui + vi-A , i = A, A+1, .. A+w-1 (addition computation)
•ui + vi-A U , i≦ = A, A+1, .. A+w-1 ( comparison computation )
Need w additions and w comparisons for a successful admission
Admission Complexity
•MDR Scheduler
• uA + v0 ≦ U (only one comparison)
•ui = ui + vi-A , i = A, A+1, .. A+w-1 (addition computation)
ui + vi-A , i = A, A+1, …, A+w-1 ≦ uA + vi-A, ∵ ui is nonincreasing ≦ uA + v0 , ∵ vj ( j = 0, 1, …, w-1) is nonincreasing ≦ U
For a unsuccessful admission, client will have to wait until the next round to repeat the admission test
Peak Transmission Rate
MDR Scheduler has the minimum peak rate among all feasible schedules with zero startup delay.
(1) Y(t) ≧ A(t)
(2) Y’(t) ≦ S’(t) , for 0 ≦t ≦T1
(3) Y’(t) < S’(t) , for 0 ≦t ≦T1
(4)
(5) Y(T1) < S(T1) = A(T1)
S(t)
A(t)
Y(t)
T1
(Contradiction)
Client Buffer Requirement
MDR scheduler generates schedules with the minimum buffer requirement among all feasible monotonic decreasing rate schedules
(1) X(t) ≧ A(t)
(2) Exist t0 such that S(t0) > X(t0) ≧A(t0)
(3) But S(Ti) = A(Ti), for i=1, 2, …,n
(4) so t0 ≠ Ti, for i=1, 2, …, n
(5) t0 not in (Ti-1, Ti), for i=2, 3,..n
(6) t0 not exist
S(t)
X(t)
A(t)
T1
t0
Buffer(Contradiction)
MDR Performance Evaluation Environment
– 274 different DVD video– Full length( average 5781 s , and 4348 MB)– 1-Gb/s backbone network
Admission Complexity
MDR Performance Evaluation
Client Buffer Requirement
Worst case buffer requirement : 394.5MB
Problem in MDR
Client buffer utilization will be low most of time.
The worst-case buffer requirement is unbounded.
Aggregated MDR Scheduler
Principle : apply the MDR principle to aggregated network flows.
Method : – Give a fixed client buffer B
– If B video buffer requirement ≧ use MDR schedulerelse use optimal smoothing algorithm with B buffer-constrainted
– Because server serves many videos simultaneously, so aggregate traffic conforms to the monotonicity property
Aggregated MDR Scheduler
Admission Complexity
Computation complexity– MDR case : need one comparison– Optimal smoothing case
transmission schedule {vi} rate-increasing round : vi > vi-1
increasing round point : hi , i=1, 2, .., g
shi+A + vhi ≦ U, for i=0, 1, …g
si+A + vi ≦ U, => si+A+1 + vi+1 ≦ U need g+1 comparisons
Performance Evaluation
Fix buffer size : 32MB
Performance Evaluation
Performance Evaluation
Conclusion
MDR is able to guarantee video delivery with tradeoff in client buffer requirement
The result of AMDR scheduler show that performance is nearly identical to optimal smoothing even for a buffer size as small as 32 MB