This section presents an initial performance evaluation of HARQ in our en-visioned 5G concept.
The results are extracted from our event driven based system level simu-lator. It implements the physical (PHY), medium access control (MAC), RLC, transmission control protocol (TCP), user datagram protocol (UDP) and in-ternet protocol (IP) layers. It also features a vertical Radio Resource Man-agement (RRM) layer which interacts with the PHY, the MAC and the RLC.
Figure A.4 shows the simulator structure.
At the transmitter side, the data generator is characterized by a traffic model, such as constant bit rate (CBR) or exponential on/off. It creates data units (DUs) according to the defined parameters (packet size, generation pe-riod, on and off time). Afterwards, the DUs are passed to the lower layers (IP and TCP/UDP) that add the headers (modelled as overhead). The DUs are then buffered in the RLC layer, until the RRM entity located in the AP de-cides that a new data transmission will be scheduled. A scheduling decision is taken according to the channel state, SINR conditions and CQI information from UE to AP of the previous TTIs. Then, a set of DUs that fits the transmis-sion bandwidth are aggregated and passed to the MAC, that adds its header and passes the resulting protocol data unit (PDU) to the PHY layer. After-wards, a coding block, whose size depends on the selected MCS and rank, is
4. Performance Evaluation
Fig. A.5:10x2 grid scenario
created once the corresponding PHY header has been added. Finally, a signal is transmitted over the wireless medium.
At the receiver side, both desired and interfering streams arrive to the antennas. The IRC receiver computes the effective SINR, which is passed to the decoding module that decides, according to a block error rate target of 10%, whether the packet is decodable. In case of failure, the HARQ manager takes care of notifying the RRM that a retransmission is required. If the packet is decodable, it is passed to the higher layers up to the sink, where the delay and throughput statistics are computed. The throughput is calculated as the total number of received bits divided by the total simulation time, and the delay is computed from the time that the DU reaches the transmission buffer to the instant that it arrives at the sink.
We define an SINR soft combining model used by the receiver to get the effective SINR upon retransmissions. Soft combining keeps memory of previous transmission of the same packet to achieve SINR gain and then improve the probability of correct detection [6]. The model can be expressed as follows:
SINRe f f ective=
n
∑
i=1
SINRi•ηn−1 (A.1)
wheren is the transmission number,SINRi is the SINR for theith trans-mission/retransmission of the same packet, andηis the combining efficiency, used to model the non-ideality of the combining process. It is set to 1.0 for simplicity.
The simulated scenario is a 10x2 grid, shown in figure A.5. One AP and one UE are randomly deployed per room. Results are collected over 100 simulation drops; at each drop, a UE is always affiliated to the AP which is in the same room (closed subscriber group). Each node is equipped with 4 antennas and an IRC receiver, and all of them use a fixed rank (1,2,3 or 4) during the whole simulation. Simulation parameters are summarized in Table A.1.
Paper A.
Table A.1:Simulation parameters
Parameter Value/State/Type
System parameters BW = 200MHz; fc= 3.5GHz Frequency reuse 1 (whole band)
Propagation model WINNER II A1 w/fast fading [16]
Antenna configuration 4x4
Scenario 10x2
Rank (fixed) 1, 2, 3 and 4
UL/DL decision maker HOL and random
Max HARQ processes 4
Max retransmission counter 4
RLC mode Acknowledged
Transport protocol UDP
Simulation time per drop 1 second
Number of drops 100
This first study shows the performance of an asynchronous and non-adaptive HARQ. Frequency reuse 1 is used, i.e., each transmission occupies the whole band, and the link adaptation decides the MCS to transmit with according to an average of the latest 5 SINR measurements. This is meant to mitigate the impact of measurement errors. The RLC layer operates in acknowledged mode, which includes a retransmission recovery procedure.
If an RLC PDU is not acknowledged within a given timeout, a retransmis-sion is triggered. Moreover, the RLC layer also aims at delivering RLC PDUs in-sequence. In case eventual missing PDUs are not recovered within a pre-defined time, they are considered dropped. Finally, neither QoS nor priority are here considered.
The used traffic model is CBR. Note that this represents a challenging situation for HARQ since there is always new data in the buffer to transmit, and this data has to be delayed if there are pending retransmissions. In addition, two UL/DL DM schemes are evaluated: one based on the buffer size and the Head-of-Line (HOL) delay [3], where the direction is chosen to avoid the buffer from overflowing and to bound the delay experienced by a packet, and a random one, meaning that the direction for each TTI is randomly chosen, with 50% probability for each link to be selected. Finally, we compare the performance of the system when the nodes transmit with fixed rank from 1 to 4.
The usage of rank 1 allows the IRC receiver to use the degrees of freedom of MIMO for suppressing the significant interfering streams [15], and then the gain from HARQ is expected to be limited since no many retransmissions are
4. Performance Evaluation
Fig. A.6:Throughput with rank 2
Fig. A.7:Number of failures per attempt
Fig. A.8:Throughput with rank 4
likely to occur. Nevertheless, the interference suppression capability of the IRC receiver diminishes for higher ranks.
Paper A.
Figure A.6 shows the cumulative distribution function (CDF) of the cell throughput when nodes operate with rank 2, for both DM schemes. We can observe that the gain of HARQ when operating with the HOL algorithm is smaller than in the case of random scheduling. This is due to the fact that, with the considered CBR traffic model, the HOL algorithm converges to a fixed downlink:uplink pattern because both buffers have the same amount of data. It is worth to notice the presence of some nodes in outage in the lower part of the CDF, i.e., nodes getting no throughput, a situation solved when HARQ is enabled.
Another interesting aspect is the lower throughput of the random DM compared to the HOL algorithm. This behaviour is due to the uncorrelated decisions on the link direction at each TTI, leading to larger SINR variance.
Therefore, a higher impact of HARQ is expected. To prove such behaviour, figure A.7 shows the total number of failures per transmission attempt. The first bar indicates the number of first retransmissions, the second bar denotes the number of second retransmissions, and so on. The last bar indicates the number of dropped packets, since the maximum attempts to retransmit a packet is set to 4. As we can see in the graphic, the number of first retrans-missions in the random case is significantly higher than the HOL algorithm case (x8 times). This is due to the fact that with the random DM we spend more time retransmitting, and therefore we reduce the chance of transmitting new data.
Figure A.8 shows the CDF of the throughput when the nodes operate with rank 4. In this case, there are 40% of the nodes in outage when the HOL algorithm is used and 22% in the random case. Such percentages are reduced to 5% and 0% when HARQ is enabled, respectively. It is interesting to notice that the random DM case is less affected than the HOL algorithm case by the high transmission rank, since the random DM leads to unpredictable interference in all rank configurations, and therefore increasing the number of streams has a lower impact on the system performance.
Table A.2 shows the throughput results in percentiles for all the possible ranks, in Mbps. We can see that in all cases HARQ gives a gain, as already visible from inspection from the plots in figure A.6 and figure A.8. The most significant gain is obtained for higher rank transmissions and in the 5%
percentile.
Finally, table A.3 shows the delay reduction per rank in percentiles when using HARQ. We can notice that significant delay reduction is obtained for the random DM case; this is due to the faster recovery mechanism provided by HARQ with respect to RLC. Nevertheless, such delay reduction is not vis-ible in the HOL DM case (negative delay detection), since HOL DM does not