Performance Evaluation of Scheduling in 5G-mmWave Networks under Human Blockage

by   Fadhil Firyaguna, et al.

The millimetre-wave spectrum provisions enormous enhancement to the achievable data rate of 5G networks. However, human blockages affecting the millimetre-wave signal can severely degrade the performance if proper resource allocation is not considered. In this paper, we assess how conventional schedulers, such as the Proportional Fair scheduler, react to the presence of blockage. Our results show that the resource allocation may disfavour users suffering from blockage, leading to low data rate for those users. To circumvent this problem, we show that the data rate of those users can be improved by using a scheduler adapted to react to upcoming blockage events. The adapted scheduler aims at proactively allocating the resources before a blockage happens, mitigating losses. Such adaptation is motivated by recent progress in blockage prediction for millimetre-wave signals in a dynamic human blockage scenario. Our extensive simulations indicate gains in the 1st percentile rate and fairness with respect to Proportional Fair scheduler when blockage conditions are severe.



There are no comments yet.


page 1

page 2

page 3

page 4


Recent Results on Proportional Fair Scheduling for mmWave-based Industrial Wireless Networks

Millimeter wave (mmWave) communication has recently attracted significan...

Scheduling for VoLTE: Resource Allocation Optimization and Low-Complexity Algorithms

We consider scheduling and resource allocation in long-term evolution (L...

Radio Resource and Beam Management in 5G mmWave Using Clustering and Deep Reinforcement Learning

To optimally cover users in millimeter-Wave (mmWave) networks, clusterin...

Proportionally Fair approach for Tor's Circuits Scheduling

The number of users adopting Tor to protect their online privacy is incr...

Block Error Rate and UE Throughput Performance Evaluation using LLS and SLS in 3GPP LTE Downlink

This paper provides a link and system level study of the downlink using ...

Distributed Learning for Proportional-Fair Resource Allocation in Coexisting WiFi Networks

In this paper, we revisit the widely known performance anomaly that resu...

Fair Resource Allocation in Optical Networks under Tidal Traffic

We propose a game-theoretic alpha-fair routing and spectrum allocation (...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

I Introduction

The mmWave (mmWave) spectrum is key to enhance the capacity of the fifth-generation (5G) of cellular networks and support multi-Gbps applications. However, transmission in mmWave frequencies may be severely affected by obstructions, such as human bodies. The blockage caused by a human body may add as much as 40 dB of attenuation to the mmWave signal [maccartney2017rapid], which can lead to radio link failures. Human blockages can be frequent in crowded places such as transportation hubs and sport arenas. This frequent, and potentially abrupt, disruption to the signal quality may impact the effectiveness of radio resource scheduling algorithms.

Blockage-aware scheduling algorithms have been adopted in the context of mmWave for relay-assisted networks, where the blockage problem is handled through allocating the transmissions over multi-hop relay paths and thus preventing possible disruptions and enabling seamless connectivity [lella2019blockage]. However, there is a lack of studies in the context of networks where relay nodes or other AP are not available for re-connection in the event of a blockage. In this context, a UE (UE) suffering from frequent blockage can have its transmission rate severely reduced, increasing the discrepancy to other UE operating in favourable channel conditions, as we will also show in our numerical evaluation.

In this paper, we evaluate the performance of resource allocation algorithms, based on the conventional PF (PF) scheduler, in the presence of mmWave signal blockages. We consider that the resource allocation is performed to multiple users, each of which may experience abrupt disruptions to signal quality caused by the human body blockage. Since the PF scheduler prioritises the allocation based on the achievable transmission rate, UE suffering from blockage may have the feasible rate degraded and thus be given progressively lower priority, disfavouring the assignment to those UE

. Hence, the challenge is to adapt the scheduler in such way that the allocation considers these signal disruptions and possible rate degradation in the following moments, and thus produces more fair assignment to

UE suffering from blockage.

With the development of blockage prediction mechanisms, a scheduler may effectively mitigate the blockage effects, as we will show in this work. Blockage prediction in mmWave communications can be performed by: (i) inferring the mmWave channel from the sub-6 GHz channel observation [alrabeiah2019deep, ali2019early], (ii) observing the past mmWave

beamforming vectors

[alkhateeb2018machine, liu2018interference] or (iii) using camera images to track the movement of potential obstructions [nishio2019proactive, alrabeiah2019millimeter]. Those methods can produce blockage indicators that will help the system predict the channel status. The scheduler can use these methods to anticipate transmission to a user that might be blocked and preemptively allocate resources, as illustrated in Fig. 1.

Fig. 1: Preemptive resource allocation for users that might be blocked.

I-a Related Works

Prediction-based schedulers have been studied in the context of previous generations of wireless systems [bang2008channel, hajipour2010proportional, schmidt2011prediction, margolies2016exploiting]

. These schedulers rely on predicted and measured information that includes fading channel gains, mobility pattern (predicted route of a user through a region) and radio coverage map (mapping between measured signal strengths and geographical locations). The prediction of the fading channel in the upcoming transmission intervals allowed these schedulers to estimate the achievable rates with increased precision, improving the network performance under the effects of rapid fluctuations of the received signal strength due to fast-fading at sub-6 GHz frequencies. However, as the accuracy of fading prediction can degrade rapidly with the prediction horizon (few milliseconds, depending on the mobility scenario)

[schmidt2011prediction], the schedulers proposed in [bang2008channel, hajipour2010proportional, schmidt2011prediction, shen2019proactive] have to update their prediction and re-optimise scheduling every millisecond. When considering mobility, the scheduler in [margolies2016exploiting] uses the prediction of the mobility pattern together with the radio coverage map to estimate the achievable rate and allocate resources accordingly. It was shown that this approach can significantly improve network performance for urban mobility scenarios and slow-fading channels. However, none of these works, e.g., [bang2008channel, hajipour2010proportional, schmidt2011prediction, margolies2016exploiting], account for the effect of the mmWave channel propagation characteristics, including the blockage effect.

For 5G wireless systems, most of the works have been focusing on developing schedulers to improve the fairness in heterogeneous networks [noliya2020performance] and the cell edge throughput [deniz2018performance]. Still, they do not consider the blockage effects on the channel and on the scheduler design. Therefore, it is not clear how those schedulers will react to blockages. In such conditions, the channel quality of a user can be more affected than, or as much as, the channel of cell edge users in 5G-mmWave networks.

On the other hand, blockage-aware schedulers have been adopted for relay-assisted networks (WPAN, WLAN, 5G wireless backhaul) [wang2019relay, lella2019blockage, niu2019relay, liu2019blockage, chen2018multipath, chang2018efficient, vu2018path], where the scheduling algorithm can allocate a transmission to a relay AP when the UE link to the serving AP becomes blocked. Such scheduling algorithms make their decision after a blockage event (reactive scheduling) and may have to re-calculate the decision every time a blockage occurs, potentially increasing overhead as blockage frequency increases.

I-B Contributions

In contrast to the reactive methods proposed in [wang2019relay, lella2019blockage, niu2019relay, liu2019blockage, chen2018multipath, chang2018efficient, vu2018path], our work considers that the scheduler could instead have a proactive approach with the aid of prediction mechanisms. Since human blockage represents a significant challenge for the 5G systems design, new prediction methods have been recently studied in [ali2019early], where one proposes the utilisation of sub-6 GHz signal to provide an early warning of blockage in the mmWave signal, and in [alkhateeb2018machine, alrabeiah2019deep, alrabeiah2019millimeter, zarifneshat2019learning, nishio2019proactive] where one proposes to use camera images to construct a prediction model from a dataset of sequential images involving the geometry and mobility of obstacles and labelled with the received power. Using such methods, blockages can be forecasted up to tens of milliseconds ahead, giving the system sufficient time to adapt. These positive blockage prediction results motivate us to revisit scheduler design for mmWave networks. In our work, we assume there exists a prediction mechanism, such as the one proposed in [nishio2019proactive] that allows the system to predict the received power in a given time window and to estimate future data rates based on that prediction. The early-warning blockage indicator allows the scheduler to have sufficient time to allocate resources in order to mitigate the losses in performance caused by the transition from LOS (LOS) to blocked state. With this system, we can implement a scheduler that allocates resources in future transmission intervals according to the estimated data rates and test its performance against the conventional schedulers.

Secondly, as the main objective of the allocation techniques proposed in [xu2019adaptive, ali2019early, alkhateeb2018machine, zarifneshat2019learning, liu2018interference, nishio2019proactive, alrabeiah2019millimeter, moltafet2018joint, gerasimenko2019capacity, polese2018distributed] is to select a different path to avoid blockage, in our work we study an alternative way to mitigate blockage effects when there is no other cell to hand-off to. Indeed, our work aims to evaluate how the conventional resource allocation methods behave under blockage when there is no spatial macro-diversity (single AP), and to study how they can be adapted to cope better with the blockage effects.

In summary, the main contributions of this work account for:

  • We show that users suffering from blockage are disfavoured in the resource allocation because conventional schedulers do not react quickly enough in case of blockage. This result shows that conventional schedulers, to be applied in 5G and beyond networks, should be adapted to cope with the impairments of the mmWave propagation.

  • We show that an adaptation of the PF scheduler can be made by changing the priority metric to consider the blockage prediction. We refer to this adapted scheduler as the BA-PF (BA-PF) scheduler. We also show that using BA-PF it is possible to increase resource allocation before the channel quality drops and reduce the allocation thereafter.

The rest of the paper is organised as follows: we describe our system model in Section II, we evaluate the user data rate performance, comparing our BA-PF scheduler to conventional schedulers in Section III and we summarise our findings in Section IV.

Ii System Model

Ii-a Frame Structure

We study the downlink transmission of a single ceiling-mounted indoor mmWave AP which transmits to UE using OFDMA (OFDMA) during  ms. Data transmission occurs in time-slots with duration and in a bandwidth divided in sub-carriers. At the MAC (MAC) layer, we assume a scheduler assigns all sub-carriers at a time-slot to a UE having index . This assignment is represented by a vector , with size elements, defined for each UE. When the time-slot is assigned to UE , the scheduler sets . Otherwise, the scheduler sets .

Ii-B Channel Model

The LOS path between AP and UE can be obstructed by blockage, such as a moving pedestrian or the user body itself, which can decrease the received signal power. We assume the blockage event is modelled according to the four-state model described in [maccartney2017rapid] and affects the channel quality of the whole link bandwidth for each UE independently. We denote as the states of the AP-UE link and we represent them in Fig. 2. The LOS state denotes the link having LOS propagation conditions, while the NLOS state denotes the link being obstructed by blockage. Moreover, the DECAY state denotes the transition state from LOS to NLOS and is characterised by a decay factor  dB/s. The RISE state denotes the transition state from NLOS to LOS and is characterised by a factor  dB/s.

Fig. 2: Four-state blockage model.

We assume that the blockers’ arrivals follow a Poisson process with average arrival rate

 blockers/s, and the blockage state has duration which is exponentially distributed with average value

 ms [jain2019impact]. The blockage attenuation power varies according to the channel state and can be expressed as


where and represent the time instants when the states DECAY and RISE start. We also assume that if multiple blockers arrive together, the channel remains in NLOS state for the duration of the longest blockage event.

The link budget for the channel between AP and UE in the time-slot can expressed as,


where is the transmit power, is the transmit antenna gain, is the receive antenna gain, is the path loss at one metre distance under free space propagation conditions, is the AP-UE distance and is the attenuation exponent.

We define the signal-to-noise ratio (SNR) as:


where represents the noise power over bandwidth . The instantaneous user data rate can be expressed as:


where is a binary function that indicates when the link goes to outage, i.e. , and the data rate becomes .

Ii-C Conventional Scheduling Algorithms

The conventional PF scheduler aims to provide a fair distribution of resources among the set of UEs that are served by the AP. Within the pool of UEs to schedule there can be UE experiencing favourable channel conditions and UE with less favourable channel conditions. To obtain a fair allocation, the PF scheduler gives priority to the UE that have the highest ratio between the feasible rate at the current time-slot and the average rate over former successive time-slots [kim2005proportional, liu2008proportional]. The PF scheduler policy to assign the user in the time-slot can be expressed as


where is the average rate of user in the time slot , and is calculated based on an exponential moving average:


where is a system parameter that weights the importance of the current feasible rate with respect to the average rate when computing the average rate metric.

On the other hand, the MaxMin (MaxMin) scheduler aims to maximise the minimum rate of the UE. This is done by prioritising the UE with the lowest average rate. The MaxMin scheduler policy to assign the user in the time-slot can be expressed as


Under the effect of blockage, the policies described in Equations (5) and (7) may lead to a decrease of the UE average rate. This effect results from the detrimental impact of the blockage attenuation on the feasible data rate. In detail, the above scheduling policies’ specific behaviours can be described as follows:

  • The PF priority function in (5) is directly proportional to the current feasible rate. Thus, as the feasible rate decreases with blockage attenuation, the priority of UE suffering from blockage decreases, giving more priority to UE in favourable conditions. This decrease in the priority is not immediate since the priority is also inversely proportional to the average rate that decreases slower than the feasible rate, depending on the weight of the moving average.

  • The MaxMin priority function in (7) is inversely proportional to the average rate. It means that, even when the UE is in outage, the priority keeps increasing from slot to slot as the average rate decreases. Thus, more slots are allocated to the UE suffering from blockage and fewer slots to UE in favourable conditions.

Therefore, both the PF and MaxMin schedulers suffer in case of blockage. In the first case, only a few slots are allocated by the PF scheduler when it is most needed, while in the second case too many slots are allocated by the MaxMin scheduler when they are not needed.

Ii-D Blockage-Aware Proportional Fair Scheduling in mmWave Spectrum

To circumvent the scheduling inefficiencies caused by the blockage effect, we study the use of a predictive scheduler that leverages the estimation of the channel status in future slots similarly to the ones proposed in [bang2008channel, hajipour2010proportional, schmidt2011prediction, shen2019proactive]. Differently from [bang2008channel, hajipour2010proportional, schmidt2011prediction, shen2019proactive], we study the application to the mmWave network use case, where it is potentially possible to predict the received power degradation due to temporary link blockage [nishio2019proactive]. Therefore, we assume that the system can predict the received power slots in advance, where

is the prediction window. We model the potential error due to inaccurate predictions with a random variable

[hajipour2010proportional, duel-hallen2006long] which is added to the future ground-truth values of the received power obtained from simulations to model the predicted received power . We also assume that increases with the prediction window size given that the prediction accuracy decreases [nishio2019proactive].

Hence, the estimated SNR (SNR) at a future time-slot can be expressed as:


We assume that a blockage is detected by the predictor when the estimated received power shows significant variations within the prediction window. Conversely, we assume that there is no blockage when remains constant over time.111As the main focus of our work is to evaluate the scheduling performance under blockage effect, we assume that in this type of scenario the attenuation due to the blockers over time prevails on other time-dependent components such as fast fading or interference.

Once a blockage is predicted, and given the estimated channel quality, a predictive scheduler will allocate more resources to those UE which are to suffer from the blockage. In practice, the UE priority is modified expressing the denominator of the BA-PF scheduler policy as a sum of the predicted feasible rates at future time-slots. For the future time-slots at the end of the prediction window, the predicted rate for the blocked UE is expected to be low due to the blockage attenuation, making the denominator smaller than it would have been for if the UE had not been affected by the blockage. Consequently, the UE predicted as blocked are assigned a higher priority than the other UE and are likely to be granted more scheduling opportunities for the future time-slots at the beginning of the prediction window. For those time-slots, their channels still experience favourable propagation conditions. On the other hand, the UE which are predicted as non-blocked may have more scheduling opportunities in the last slots of the prediction window, when the channels of the blocked UE will be affected by a strong blockage attenuation and are assigned a low priority which allows to release the resources to those non-blocked UE.

Hence, we adapt the priority function in (5) to make the PF scheduler blockage-aware (BA-PF). The BA-PF scheduler policy to assign the user in the time-slot can be expressed as


where is the feasible instantaneous user rate at a future time-slot and can be computed as:


and is the estimated sum-rate of all time-slots from to given the allocation vector and can be expressed as:


Thus, the assignment of the first slots depends on how the following slots in the future are assigned. Because of that, the time-slot assignment starts from the end of the prediction window at time-slot and works backwards to time-slot , as shown in Algorithm 1.

Result: Allocation assignment
if blockage is predicted to happen during any of the next slots then
       for  down to t do
             for  to  do
             end for
       end for
      Execute PF algorithm.
end if
Algorithm 1 BA-PF scheduling algorithm

For instance, consider a UE that suffers from blockage and has the estimated rate () degraded in the future. Because of the backward iteration, other UE were allocated in the last slots as is lower than the estimated rates of the other UE. Hence, as the iteration reaches the first slots, is high and is low since very few time-slots have been allocated to UE , leading to a increased priority function according to (9). Therefore UE has the transmission anticipated as it gets more priority in the first slots and releases the last slots for UE with better channel conditions.

Iii Numerical Results

Iii-a Simulation Setup

In this section, we evaluate and compare the PF, MaxMin and BA-PF schedulers performance. As depicted in Fig. 3, we consider a single ceiling-mounted AP that is positioned in the centre of the network area at a height of 2 m above the UE antennas and is equipped with a directive antenna facing downward having beamwidth 170  and gain of 3.16 dBi [firyaguna2020performance]. The UE

are uniformly distributed in the

AP coverage area (of radius 15 m), are equipped with an omnidirectional receiver antenna and are all served from the same AP. The channel model parameters are in line with empirically-tested model described in [yoo2017measurements] having LOS channel state with parameters 63.4 dB and 1.72 . We assume maximum blockage attenuation  dB, decay rate  dB/ms, and rise rate  dB/ms, in line with the measurements reported in [ju2019millimeter].

Fig. 3: Ceiling-mounted AP setup with downward facing beam illuminating the UE.

We simulate a sequence of 48000 time-slots of an OFDMA frame transmission with carrier frequency of 60 GHz according to the 5G NR (NR) standard [3gpp_tr_38211]. We gain insights on the various schedulers by changing the characteristics of the blockage scenario, including different blockers’ arrival rate and blockage duration

. To analyse the impact of the blockage effects on scheduling, we evaluate the empirical cumulative distribution function (ECDF) of the average user data rate. The 1st percentile rate, mean rate and fairness metrics are selected for performance evaluation. The 1st percentile rate can be interpreted as the average data rate of the

UE that are most affected by blockage. The mean rate is defined as the sum of the average user rates divided by the number of UE in the cell, and fairness is defined as the Jain’s fairness index [jain1984quantitative]. Settings for the system parameters are provided in Table I.

Number of UE 8
Slot duration 62.5 s
Frame bandwidth 2 GHz
Moving average weight 0.5
Transmit power 100 mW
Noise figure 9 dB
Noise power -71.99 dBm
SNR threshold 0 dB
AP beamwidth 170 
AP antenna gain 3.16 dBi
UE antenna gain 0 dBi
Prediction window {50, 200, 500} ms
Prediction error { ,,}
TABLE I: System parameters

Iii-B User Performance

(a)  blockers/s,  ms
(b)  blockers/s,  ms
(c)  blockers/s,  ms
Fig. 4: Empirical cumulative distribution function of the user rate for different blockage scenarios with eight UE in the cell.

In Fig. 4 we show the ECDFs of the average user rate for the PF, MaxMin and BA-PF schedulers in three different blockage scenarios: (a) low blockers’ arrival rate, short blockage duration (Fig. 3(a)); (b) mid-range blockers’ arrival rate, short blockage duration (Fig. 3(b)); and (c) high blockers’ arrival rate, long blockage duration (Fig. 3(c)).

In the blockage scenario (a) depicted in Fig. 3(a), the MaxMin scheduler (red curve) has the worst performance with respect to PF and BA-PF schedulers. This is because the UE that suffered from blockage were mostly prioritised even though they had poor channel quality, leading to low user rate. On the contrary, the UE with better channel quality were less prioritised, also leading to low user rate. Moreover, we observe a smaller discrepancy in performance between PF (blue curve) and BA-PF (yellow, purple and green curves) schedulers. In fact, when there are fewer blockage events, this is expected since the BA-PF priority function in (9) is executed only a few times. In the vast majority of cases, the conventional PF is executed during the prediction window because no blockage is detected.

Furthermore, in the blockage scenario (b) depicted Fig. 3(b), where the blockage effect is more severe, the MaxMin scheduler (red curve) still has the worst performance, while the difference between PF and BA-PF performances increases when looking at the 1st percentile of the ECDF, where we observe a performance gain of 48 % with respect to the PF for all the prediction window values. This is because the BA-PF priority metric in (9) is designed to allocate resources at the beginning of the prediction window and to release them at the end considering that there is only one transition from LOS to NLOS states. Hence, when the blockage duration is short, or there are not as many arrivals as in the blockage scenario (c), multiple state transitions may occur within a large prediction window, leading to a non-ideal scenario for the BA-PF scheduling metric as the resource anticipation is not effective.

Finally, the best performance can be observed for the blockage scenario (c), as depicted in Fig. 3(c), when it is more likely that the blockage state is long enough to occur only once within the prediction window. When using the PF scheduler, the UE that are affected by the blockage are less prioritised as more blockage events occur, decreasing their user rate. Thus, the performance of the PF scheduler (blue curve) approaches the performance of the performance of the MaxMin scheduler (red curve) for the lower values of the ECDF. In contrast, when adopting the BA-PF scheduler, the UE that suffer from blockage are more prioritised than in the PF scheduler, and we observe gains up to 153 % in the 1st percentile rate with respect to the PF performance for the prediction window value  ms (yellow curves). Moreover, this gain is improved to 267 % when increasing the prediction window to  ms (green curve in Fig. 3(c)), as the pool of resources to allocate increase and there is more time to anticipate a transmission for a blocked UE.

Iii-C Network Performance

Fig. 5: Average fairness performance as a function of , with  ms.
Fig. 6: Mean rate performance as a function of , with  ms.

In Figs. 5 and 6, we show the performance of the network in terms of mean rate and fairness index as a function of the blockage effect. From Fig. 5 we can observe that the MaxMin fairness performance is high (0.95) and is not affected by blockage. The BA-PF performance degrades with , as the blockage scenario gets more severe. This is because the preemptive allocation for UE suffering from blockage drains a significant amount of resources taken from the pool of all the available resources. Hence, as the blockage severity increases, BA-PF assigns more resources to UE transitioning to blockage state, disfavouring other UE, and consequently decreasing fairness. However, the BA-PF fairness is still higher than PF when the blockage arrival rate is high ( blockers/s), as PF prioritises the allocation of resources to a decreasing number of UE that are not blocked.

Moreover, from Fig. 6, we can observe that the MaxMin mean rate is significantly lower compared to other schedulers, as the high fairness is achieved by taking away resources from UE with good channel quality and giving them to UE with worse quality, levelling down the user rates. We observe that BA-PF and PF mean rates significantly degrade with , as the number of blocked UE increases. We see that the BA-PF mean rate is up to 22 % lower than the PF mean rate, as the gains achieved by improving the data rate of UE suffering from blockage comes at the cost of reducing the data rate of non-blocked UE.

Iv Conclusion

In this work, we assessed the behaviour of schedulers in the event of a blockage affecting the LOS signal path. We showed that the conventional MaxMin scheduler, designed to improve the cell-edge user rate, cannot improve the rate of the users suffering from blockage and has its performance significantly degraded. We showed that the PF scheduler, although it is less affected by blockage compared to MaxMin, can have a user performance as low as MaxMin when blockage scenario is severe. This is because conventional schedulers react to blockage by allocating more resources to compensate the loss in the data rate (as the MaxMin scheduler does), or by allocating fewer resources to prioritise other UE with better channel conditions (as the PF scheduler does). However, when the UE is entirely in outage, a successful transmission might be impossible to achieve, and any amount of allocated resources may be unused, reducing the efficiency of the allocation.

This reduction in efficiency suggests that conventional schedulers should be adapted to mitigate the blockage effects in the context of 5G-mmWave networks. We therefore considered an adapted scheduling metric BA-PF which relies on blockage prediction to compute the estimated user rates, many slots ahead of a blockage event. The effectiveness of this adaptation is demonstrated by increasing the blockage arrival rate and the blockage duration. We showed that the gain in performance between the BA-PF scheduler and the conventional ones is more significant when the arrival rate and duration are high, i.e., when the channel stays in blockage state long enough to allow for an effective anticipation of resources. This gain is due to an improvement in the rate of users affected by blockage, which is reflected in the gains in the 1st percentile rate and fairness with respect to the PF performance. However, this improvement comes at the expense of performance experienced by other UE, as reflected on the losses in mean rate performance.