I. I NTRODUCTION
The real-time nature of video transmission requires maintaining stringent delay bounds to ensure a good user experience. The stringency of the delay bound is further dependent
on the specific use case. For instance, interactive applications
such as video conferencing can only tolerate an end-to-end
delay in the order of few hundred milliseconds for a smooth
experience whereas with live streaming, the delay constraint
can be relaxed to few seconds. Since the bottleneck in the endto-end network delay is queuing, users with more stringent
delay constraints should be allocated more physical resources
to boost their service rates, reduce the queuing delay, and
support their QoS requirement. Furthermore, there is an inherent tradeoff between queuing delay and video source rate.
Therefore, rate distortion optimization should be applied to
ensure that satisfying delay constraints does not come at the
expense of maintaining high perceptual quality.
In this paper, we derive a resource allocation and rate
adaptation policy that maximizes the sum video quality such
that all users in the network can achieve a target statistical
delay bound. Statistical delay bounds are considered since
deterministic delay bounds are prohibitively expensive to
guarantee over wireless networks in the presence of fading.
We show that the optimal operating point per user is such that
the rate-distortion slope is the inverse of the supported video
source rate per unit bandwidth. The maximum source rate per
unit bandwidth is a fundamental measure of the number of
video bits per channel use that can be delivered subject to
the QoS requirement and is termed source spectral efficiency.
Downloaded
from http://www.elearnica.ir
978-1-4799-1353-4/13/$31.00
2013 IEEE
For the general case where the set of users in the network
cannot all be served, we solve the problem of selecting a
maximal subset of users to schedule such that each user can
meet his target QoS requirement. We show that the optimal
scheduling policy can be obtained in polynomial time in the
number of users and it involves computing the minimum
resource allocation required by each user to support their QoS
requirement, using it as a sorting criteria, and scheduling the
first sorted users such that the sum of their minimum resource
requirement does not exceed the total available resources.
Previous work on multiuser video transmission focuses on
stored video use cases in which content can be buffered ahead
opportunistically and large delays can be tolerated. For example, for stored video streaming, in [1], the knowledge of the
variations in the dependence of perceived video quality to the
compression rate is utilized for resource allocation to improve
video QoE across multiple users. In [2], a scheduling scheme
for streaming pre-encoded video over wireless networks is
presented in which rates are dynamically adjusted based on
channel quality. For real-time video, however, such as live
streaming or video conferencing, whereby the content in not
pre-encoded, large buffers are not feasible and stringent delay
constraints need to be guaranteed. For that important use case,
no previous work addresses the problem of multiuser resource
allocation under statistical delay constraints.
The effective capacity link layer model [3] characterizes the
capacity of the wireless channels in the presence of queues
using QoS exponents that describe the decay rate of the
queue length tail probability and characterize a corresponding
statistical delay bound. The notion of statistical QoS is tied
back to the well developed theory of effective bandwidth
[4][6] and its dual concept of effective capacity [3], [7],
[8]. The main limitation of previous work is that effective
capacity has not been applied previously in the context of a
cross-layer design to optimize a video quality-based utility
function. In [9], effective capacity analysis is applied for
scalable video transmission to minimize resource utilization
over unicast and multicast links. Considering energy-efficiency
as a target objective in an ad-hoc network, [10] derives energyefficient transmission schemes such that the end-to-end delay
bounds are satisfied. Our proposed approach in this paper,
on the other hand, provides a comprehensive framework to
jointly capture the tradeoff between end-to-end delay, video
perceptual quality, and resource allocation.
1736
Rate-Distortion
Characteristics
Q1(R1)
Q2(R2)
User 1
B1
User 2
B2
R1
....
QK(RK)
....
R2
User K
RK
BK
Rate-distortion characteristics
Delay QoS requirements
Channel state information
(1)
1737
maxR,B
s.t.
K
X
k=1
K
X
Qk (Rk )
(6)
Bk = B; Bk 0 k
(7)
Rk (Bk , k )
|
{z
}
k=1
Pr Dk > Dkth Pkth k
(8)
k
where pk = Pkth
. Furthermore, the solution to the
problem requires that k = dk /Bk where the bandwidthQoS exponent product dk is a constant dependent only on
the channel distribution and the delay constraint.
minB,
K
X
k=1
s.t.
K
X
Bk T
1
kln(2)
Qk
ln E {(1 + k )
}
T k
Bk = B; Bk 0 k
(10)
k=1
n
k Bk T o
E (1 + k ) ln(2)
= pk k
(11)
Rk (Bk , k )
Bk
|
{z
}
= .
(12)
k B k T
where Rk (Bk , k ) = T1k ln E {(1 + k ) ln(2) } and
PK
is a constant chosen such that
k=1 Bk = B. Further,
Rk (Bk , k )/Bk is independent of Bk and is only a function
of the fading distribution fk (), the delay bound Dkth , and the
delay bound violation probability Pkth .
Proof: See [13], proof of Theorem 1.
We refer to the term Rk (Bk , k )/Bk as the source spectral
efficiency as it represents the number of video bits that can be
delivered per channel use subject to the QoS constraint. The
interpretation of the above result is that the optimal resource
allocation is such that the slope of the rate distortion curve
multiplied by the source spectral efficiency is the same for all
users. Furthermore, that constant determined by is computed such that the sum bandwidth constraint is satisfied with
equality. Note that as decreases, Bk increases for all k and
vice versa, thus reducing the problem intoP
a one-dimensional
root-finding problem involving solving for k Bk ()B = 0
to find .
Note from Theorem 1 that the source spectral efficiency
jointly characterizes the channel and the QoS requirement
since it is independent of the resource allocation Bk . It is only
a function of the channel statistics, the delay bound, and the
delay bound
probability. Specifically,Rk (Bk )/Bk =
violation
th
th
ln 1/Pk /Dk T dk . Furthermore, ln 1/Pkth /Dkth T dk is a
fundamental measure of the maximum number of video bits
per channel use that can be delivered subject to the QoS
requirement.
V. M AXIMAL U SER S UBSET S CHEDULING
In this section, we consider the problem of selecting video
users to serve given their QoS constraint. The objective is to
select the largest subset of users such that each user in the
subset can meet their delay constraint.
Theorem 2 presents the optimal scheduling policy that
selects a maximal subset of users such that all user meet their
statistical delay constraint. To summarize, the optimal scheduling policy can be obtained in polynomial time in the number
of users and it involves computing the minimum bandwidth
required by each user to support their QoS requirement, using
it as a sorting criteria, and scheduling the first sorted users
such that the sum of their minimum bandwidth requirement
does not exceed the total bandwidth.
1738
Theorem 2. Maximal User Subset Scheduling: Under maximum sum quality resource allocation, the scheduling policy
to maximize the number of users that can meet their QoS
requirement is as follows. Define
4.5
(13)
4
3
3.5
10
2.5
0.5
0.1
0.01
1.5
10
0.5
Practically
Infeasible
Region
0.1
1
0.0
10
Dkth dk T min
R
=
ln(1/Pkth ) k
Bkmin
5
Source Spectral Efficiency R /B (bps/Hz)
for = 20 dB
0.5
10
0.2
N = argmaxN
N
X
min
s.t.
Bl(j)
j=1
N
X
min
B
Bl(j)
(14)
j=1
Bk k k by solving E (1 + k )
= pk .
3)
Compute maximum
source spectral efficiency Rk (Bk )/Bk =
ln 1/Pkth /Dkth T dk .
end for
Step 2. Select Maximal User Subset
min is the j th sorted
1) Sort Bkmin with operator l(j) such that Bl(j)
element j = 1, , K
P
min
2) Compute maximum number of users N = argmaxN N
j=1 Bl(j)
PN
min B
s.t. j=1 Bl(j)
3) Schedule users l(1), , l(N ).
Step 3. Allocate Resources
1) Set = initial
Qk (Rk
(Bk
,k ))
T d
k ( )
2) Solve R
= ln(p
for Rk k = l(1), , l(N )
(B , )
k)
k
k k
3) Solve k = ln(pk )/(T Rk ) k = l(1), , l(N )
4) Solve
P Bk = dk /k k = l(1), , l(N )
5) If
Bk < B, decrease
P and repeat 2,3,4, otherwise increase and
repeat 2,3,4 until | Bk B| < .
4)
VI. R ESULTS
In this section, we present results and analysis to demonstrate the performance of the proposed scheduling and resource
allocation algorithms for real-time video transmission. First,
we present analysis of the source spectral efficiency under
different QoS requirements and channel conditions. Next, we
present results for the case of multiple users to demonstrate
the gains of QoS-aware scheduling and resource allocation.
0.5
1.5
2
2.5
Delay bound Dth (sec)
3.5
Fig. 2. Contour plot of source spectral efficiency for a Rayleigh channel with
1739
300
5 Km
184 users
3.4x
260
5 Km
5 Km
Proposed
280
240
Scheduling: Max SNR
Resource Allocation: Equal BW
220
200
160
2.1 Km
35 users
29 users
1.5 Km
14 users
180
Baselines
140
120
100
4.1x
80
60
40
20
0
1
10
10
10
5 Km
5 Km
5 Km
5 Km
Fig. 3. (a) Number of users supported by the proposed scheduling and resource allocation algorithm in comparison to the non-QoS-aware baseline vs. base
station transmit power; (b) Users covered by the proposed Maximal user subset scheduler and maximum sum quality resource allocation for Pt = 25 W; (c)
Users covered by the baseline Max SNR scheduler baseline and equal bandwidth allocation. Red markers correspond to video conferencing sessions and blue
dots correspond to live streaming users.
1740