Appendix A

Resource Allocation

A.1 Dynamic Resource Allocation

A.1.1 Utility Predictive Scheduler

The main novelty introduced with the Utility Predictive Scheduler (UPS) is the formulation of its rate-based utility function, shown in Figure A.1. The scheduling priority of each user depends on the rate increase due to the allocation of the considered Physical Resource Block (PRB) to user, weighted by average user's throughput, similarly to Proportional Fair (PF) scheduler, with the users with the highest weighted rate being selected. Although the utility function is rate dependent, the design of the function is based on three parameters: α, β, and γ, as shown in Figure A.1. These parameters can be used to change the slope of selected sections of employed utility function to prioritize User Equipment (UE)s affected by higher latency, thus providing the support for different Quality of Service (QoS) classes. Moreover, even when considering only Best Effort (BE) traffic, this design of the scheduler and its utility function leads to a slight gain in spectral efficiency, as shown in WINNER+ (2009a,d).

Figure A.1 Utility function of UPS and its parameterization – mapping data rate to utility value

A.1.2 Resource Allocation with Relays

A very interesting approach to resource allocation with QoS support in relay-enhanced network has been proposed in WINNER+ (2009C). The HurrY-Guided-Irrelevant-Eminent-NEeds (HYGIENE) scheduling algorithm brings urgency on top of relaying, which means that it gives the priority to urgent UEs and then to relayed UEs. A rushing entity classifier is introduced, which determines whether the UEs are of urgent class or not. Then, in the second processing step, the Base Station (BS) identifies UEs that require relaying. Based on the above mentioned classification scheduling, priority values are assigned to UEs according to Table A.1.

The resource allocation is jointly performed for a group of two consecutive time slots. First, relayed urgent UEs are scheduled, with the Real-Time (RT) packet being prioritized according to their remaining time-to-live, and the PRBs allocated in order to maximize the spectral efficiency. For allocation of PRBs to Non-Real-Time (NRT) packets a Proportional Fair (PF) scheduler is used. In the second step, the urgent non-relayed UEs are scheduled according to the same allocation policy as above. Finally, the non-urgent UEs are scheduled in the third step according to the Max C/I policy WINNER+ (2009C).

Figure A.2 presents the performance evaluation of the HYGIENE algorithm in comparison with several well-known schedulers: Maximum Carrier to Interference (MCI), PF, Modified-Largest Weighted Delay First (M-LWDF) and Exponential Delay Fairness (EDF). One can notice the performance improvement for the delay-sensitive RT services, such as the Voice over IP (VoIP).

Figure A.2 Maximum achievable cell capacity with PF, MCI, M-LWDF, EDF and HYGIENE schedulers – mixed real-time traffic scenario

A.2 Multiuser Resource Allocation

A.2.1 PHY/MAC Layer Model

Resources are divided into Resource Blocks (RB) occupying a given bandwidth and time, which can be allocated flexibly to the K UEs. A scenario where UEs travel with potentially high velocities is assumed. The high dynamics of the time varying channel prohibit the utilization of instantaneous Channel State Information at the Transmitter (CSIT). However, long-term CSIT that includes distance-dependent path loss and log-normal shadowing is assumed to be available.

Given long-term CSIT the data rate served to UE k is given by:

(A.1a)

where accounts for the rate if UE k is assigned all available slots exclusively. Additionally, the constraints

(A.1b)

need to be fulfilled with being the set of all UEs.

A.2.2 APP Layer Model

The generic application characteristic resembles a bounded logarithmic relation between perceived quality and data rate, as illustrated in Figure A.3, described by the MOS as a function of the data rate R_k of UE .

(A.2a)

with

(A.2b)

(A.2c)

and

(A.2d)

Figure A.3 Generic application characteristic for one example application class

The semilogarithmic plot of Figure A.3 visualizes the related parameters: the parameter determines the slope of while shifts the curve along the x-axis.

The application characteristic of UE k, denoted by in Figure 2.5, can be parameterized by only two parameters, , or alternatively .

A.2.3 Optimization Problem

The objective of application driven multiuser resource allocation is to assign the available resources over the shared wireless medium described by the Physical (PHY)/Medium Access Control (MAC) layer model (A.1) running different applications modeled by (A.2). The share of resources for each UE, α_k, and the associated application data rates, R_k, are determined by solving an optimization problem between link and APP layer, as illustrated in Figure 2.5.

One commonly used utility function is to maximize the sum of the Mean Opinion Score (MOS) of all UEs, thereby maximizing the average perceived service quality. The corresponding optimization problem is formulated as in A.3 (Saul and Auer 2009):

(A.3)

where .

Alternatively, the max-min approach distributes resources such that all UEs experience the same utility degradation, which can be cast in the following optimization problem (Radunovic and Le Boudec 2007):

(A.4)

Provision of max-min fairness in terms of UE perceived quality implies that one single UE that cannot achieve a good perceived quality, for example, due to a poor wireless channel and/or a demanding application, forces all other UEs to share this poor experience. Furthermore, from an operator's point of view it might be desirable to provide premium services with a higher quality only to some UEs, which contradicts the idea of “equal loss”.

To mitigate these shortcomings, a resource allocation scheme is presented in (Saul 2008), which is described by the following optimization problem:

(A.5)

where is the subset of UEs that satisfy the specified MOS constrains, and is a strictly monotonic function that allows the operator to differentiate UEs, for example, to ensure that the MOS of a premium service is superior to the ordinary service quality. For instance, the condition that the MOS of UEs k₁ exceeds the MOS of UE k₂ by , is obtained by setting the inverse priority function to .

In order to provide at least minimum service quality an access control is established, in the way that K_act out of K UEs are served, while the remaining UEs are denied access. By incorporating admission control to the resource allocation problem, service guarantees of for admitted UEs are established, according to the desired UE priorities, that is , .

The optimization problem (A.5) can be solved by finding the roots of an equation system. The following algorithm determines the admitted UEs and the share of resources α:

Step 0:	Initialize,
Step 1:	Drop UEs until equation system is solvable (see step 1a–c)
Step 1a:	Solve ,
Step 1b:	If , continue with step 2
Step 1c:	Drop UE , that is, , and continue with step 1a
Step 2:	Solve equation system , , and

It may happen that the available resources are insufficient to serve all UEs with the desired quality . In such cases some UEs cannot be served, which is realized in the algorithm step 1c.

Implications for the System Architecture

For Cross-Layer Optimization (CLO) between link and Application (APP) layer, different timescales are involved. As timescales on the link layer are several orders of magnitude smaller than those of the APP layer, in order to limit the signaling overhead between optimizer, link and APP layer, the optimizer should be placed close to the BS. For tuning of the APP layer data rates on a shorter timescale, it is desirable that transcoding is supported at the BS, and/or scalable video codecs ITU-T 2007; Schwarz and Marpe 2007) are deployed.

A.2.4 Simulation Results

In order to evaluate the performance of the considered resource allocation schemes that aim to maximize the perceived quality as described in Section A.2.3, computer simulations are conducted. We consider a Long Term Evolution (LTE) downlink in the 10 MHz bandwidth Frequency Division Duplex (FDD) mode. By employing Orthogonal Frequency Division Multiple Access (OFDMA), UEs can be assigned to RBs in time and frequency. The Wireless World Initiative New Radio (WINNER) typical urban macrocell channel model (model C2, WINNERII 2007) is used, comprising path loss, shadowing and time-variant frequency-selective fading. The average Signal-to-Interference-plus-Noise Ratio (SINR) is constrained such that transmission at least with the lowest supported Modulation and Coding Scheme (MCS) is feasible. Taking into account velocities of UEs of 50 km/h, only long-term Channel State Information (CSI) at the transmitter is assumed to be available. While the same modulation and coding scheme is chosen for all RBs assigned to one UE, PRB's assigned to different UEs will typically use a different MCS. Furthermore, taking into account fast Automatic Repeat-reQuest (ARQ) at the BS, it is assumed that packets are always received error free. The applications are modelled by a bounded logarithmic utility (A.2), with a minimum required data rate of R_1.0 = 300 kbps and a desired data rate of R_4.5 = 3 Mbps.

Figure A.4 show the CDF of the MOS and the data rate, respectively, for the CLO techniques. There are K = 16 UEs with a guaranteed service quality of MOS, which corresponds to a data rate of 1.1 Mbps. While the conventional max-min MOS technique (A.4) can serve only 81% of the UEs with MOS, a significantly higher number of UEs achieves at least MOS for the max-min MOS utility (A.5) that incorporates admission control. This gain is achieved at the expense of not serving 1.3% of the UEs, who suffer a bad channel, for example due to strong shadowing at the cell edge. The max-sum MOS technique (A.3) serves more than 50% of the users with the best possible quality of MOS = 4.5. These are the users with good channels, for which high data rates can be achieved with comparably modest resource utilization. On the other hand, more than 12% of the users are not served with at least MOS, and 6.5% of the users are not served at all.

Figure A.4 CDF of UE perceived quality (Saul 2008). Reproduced by permission of © 2008 IEEE

A.3 Busy Burst Extended to MIMO

For performance evaluation the downlink of a hexagonal cell deployment is considered. The BS transmitter selects one beam from a set of fixed beamforming vectors given by a Discrete Fourier Transform (DFT) codebook. Channels models are taken from WINNER scenario C1 (Döttling et al. 2009, Chapter 2). A full buffer traffic model is considered. Perfect time and frequency synchronisation of the network is assumed. The simulation parameters shown in Table A.2, are taken from the WINNER Time Division Duplex (TDD) mode (WINNERII 2006). Link adaptation is assumed, where the modulation scheme is adaptively controlled based on the achieved SINR.

A.4 Efficient MBMS Transmission

A.4.1 Service Operation

Concerning streaming services, the system must ensure maximum coverage for its transmission. Therefore, more robust modulation and coding schemes are selected at the expenses of a reduced bit rate capability. The file delivery case is much more challenging and thus we focus on this service. For the rest of the section, file delivery case is the transmission of a 2 MB file.

The file download service in the Multimedia Broadcast Multicast Service (MBMS) consists of three phases:

In case of selecting the p-t-m mode, the transmission duration has to be configured, so to ensure that the file is successfully received by, in this case, 95% of the users. In order to achieve this, there are two possible options:

• Transmit the file using only MBMS until the 95% acquisition probability is reached (referred to as Conventional File Delivery).
• Transmit data with MBMS during the time required to achieve a certain acquisition probability and then use a repair phase that will serve the remaining users (referred to as Hybrid Delivery).

Figure A.5 Dynamic switching criteria (left) and static switching criteria (right)

The decision must be made with the aim of optimizing the acquisition probability of the initial MBMS transmission phase while minimizing the delivery time.

To avoid congestion, error reporting messages from UEs can be distributed over time within a backoff window and across multiple repair servers. The backoff window should be large enough to prevent congestion, but should not increase the duration of the repair phase.

As mentioned before, UEs start the repair phase using dedicated p-t-p connections, but if the number of active users in this phase is high enough it is possible to employ a p-t-m connection with MBMS. However, during the initial file transmission there is no communication between the UEs and the server. Therefore, once the initial MBMS transmission is finished, the server does not have any information about the number of users that have not received the file and the amount of repair data needed by each of them. This information can be estimated in the beginning of the postdelivery repair session using application layer feedback.

With the aim of serving users with bad channel conditions, the new MBMS transmission should use a more robust MCS. The duration of this transmission would depend on the amount of repair data requested by the users at the time of the decision.

A.4.2 Frequency Division Multiplexing (FDM) Performance

For the FDM case, from the 10 MHz available, half will be dedicated to unicast, that is, 5 MHz, and the same for multicast. In terms of MBMS performance, this FDM distribution is equivalent to allocating five subframes from the ten available to MBMS. As shown in Figure A.6, FDM is never a good option for file delivery due to lower frequency diversity.

Figure A.6 Average user throughput versus the number of users per cell for unicast users with FDM

Table A.1 Priority rules of HYGIENE scheduler.

Table A.2 System level simulation parameters.

Parameter	Value
Carrier centre frequency	3.95 GHz
System bandwidth B	89.84 MHz
Number of subcarriers Nc	1840
Frame duration	0.6912 ms
OFDM symbols/frame	30
Total OFDM symbol length	22.48 μs
RB size	15 (time) × 8 (frequency) = 120
Number of RB/frame	2 (time) × 230 (frequency)
Number of sectors/cell	3
Number of antenna elements/sector	4
Average number of UEs/cell U	10
Transmit power per RB P	16.4 dBm
Elevation antenna gain Ae	14 dBi
Azimuth antenna element gain	[dB]where, and
Noise level N	−117.8 dBm/RB
Number of snapshots	50
Simulation duration per snapshot	50 ms