In-network Angle Approximation for Supporting Adaptive

Beamforming
Hiba Mallouhi Jaspreet Kaur
ELTE Eötvös Loránd University University of Glasgow
Budapest, Hungary Glasgow, United Kingdom
[email protected] [email protected]

Hasan Tahir Abbas Sándor Laki

University of Glasgow ELTE Eötvös Loránd University
Glasgow, United Kingdom Budapest, Hungary
[email protected] [email protected]

ABSTRACT KEYWORDS
There is a great interest in utilizing P4 for in-network computing 5G, P4, In-Network Computing, Beamforming
along with programmable data planes. This use is emerging as a
new network paradigm that can not just reduce the complexity ACM Reference Format:
but the delay as well. Beamforming is now an integral feature of Hiba Mallouhi, Jaspreet Kaur, Hasan Tahir Abbas, and Sándor Laki. 2022.
modern wireless communication systems and its implementation In-network Angle Approximation for Supporting Adaptive Beamforming.
In P4 Workshop in Europe (EuroP4 ’22), December 9, 2022, Roma, Italy. ACM,
calls for an accurate beam alignment by estimating the direction of
New York, NY, USA, 6 pages. https://doi.org/10.1145/3565475.3569076
signal arrival. However, this estimation is computationally complex,
especially in a dynamic environment where a user is constantly
on the move. In this paper, we propose a user-assisted in-network
method to optimally approximate the angle of arrival by segmenting 1 INTRODUCTION
the cell area into an exponentially binned grid and make use of the The electromagnetic spectrum available to current-day consumers
advantages offered by programmable data planes and their match- in a wireless network is increasingly getting stretched owing to the
action table (MAT) logic. The method expects location messages massive demands of data. Exploring higher frequency in wireless
periodically reported by user equipment, processes them in the communication networks is one of the telecommunications sec-
network and reconfigures the base station antennas accordingly, tor’s most crucial tasks in order to address this problem [13]. As a
implementing user-assisted in-network beam control. The proposed result, millimetre wave (mmWave) frequencies, which range from
method is implemented in P4 and runs on a Tofino ASIC. Our 30-300 GHz, are being investigated for use in the fifth generation of
evaluation proves a theoretical bound on the absolute error of the wireless communication networks (5G). One of the most promising
proposed MAT-based angle approximation and shows that it is in technologies for future wireless networks, mmWave can accom-
accordance with the empirical error distributions. Moreover, there is modate high data rate transfers thanks to its abundant spectrum
no significant increase in errors attributed to the latency of various resources [7, 10, 13–15]. High propagation loss, which results in
control cycle times (less than 100ms) and the user’s movement short propagation lengths, and signal blockage brought on not only
at moderate speeds (of less than 90km/h.) We also show that the by building materials and greenery but also by the human body and
resource usage is only affected by the size of the TCAM table used to high oxygen absorption are some of the difficulties that mmWave
store the angle approximation values and that the proposed method must overcome [8].
has no significant per-stage resource usage on the pipeline. Beamforming [5], which has many advantages, such as increased
coverage at a cell’s edge, enhanced signal quality, tracking the user
CCS CONCEPTS equipment (UE), and enabling collaboration among base stations
(BS), is one practical solution to mitigate the reduced propagation
• Networks → Programmable networks; In-network process-
distance in mmWave communications. The high route loss that
ing; Mobile networks.
mmWave signals experience can be partially offset by directional
beamforming, but it creates a challenging beam alignment problem.
In order to establish the beamforming direction, a BS particularly
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed needs to know the direction of arrival (DOA) of its users. Search-
for profit or commercial advantage and that copies bear this notice and the full citation ing exhaustively for all potential matches to get the optimal beam
on the first page. Copyrights for components of this work owned by others than the alignment is a natural way to carry out beamforming training to in-
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission crease alignment accuracy [12]. However, this technique consumes
and/or a fee. Request permissions from [email protected]. so much computational power. As a result, we propose to realize a
EuroP4 ’22, December 9, 2022, Roma, Italy P4-enabled switch that can perform (or approximate) basic math-
© 2022 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-9935-7/22/12. . . $15.00 ematical operations and also hold the prior information in form
https://doi.org/10.1145/3565475.3569076 of tables. This is done to estimate the angle from the transmitter

61
EuroP4 ’22, December 9, 2022, Roma, Italy Hiba Mallouhi, Jaspreet Kaur, Hasan Tahir Abbas, and Sándor Laki

to the mobile user or vice versa by means of performing simple

trigonometric calculations.
The motivation behind this study is as follows:
(1) Complex tasks can be performed at the network switch level
(data link layer.)
(2) A truly distributed architecture.
(3) P4 is going to reduce the congestion caused at the network
layer by moving some of the operations and tasks to the link
layer.
(4) Network customization as a result of software-defined radio.
As a result, we get flexibility; ideal for 5G networks.
Beamforming requires training of the network that determines
the best beams within a wireless channel. There are various algo-
rithms such as binary search beam training over a beamforming
code-book. Typically this pre-configuration is done at the MAC
sub-layer (part of the link layer.) MAC protocols provide control to
physical devices such as antennas, etc. Although the state of the art
in beamforming undoubtedly increases the network performance
at the user end, work needs to be done to assess the overall increase Figure 1: Main concept of the proposed user-assisted in-
in performance when the backhaul network is also considered. network beam-control.
In this paper, we propose a user-assisted in-network approach
for accurately steering the beam towards the UE. Accordingly, we
assume that the UE periodically reports its location to an in-network to a relative coordinate system where the origin (0,0) represents
computing node (i.e., a P4 switch in the transport network or RAN) the base station’s location. We assume that the shape of the cell is
that uses its location and the base station information to compute similar in the four quadrant around the base station and thus we
the angle that is then used to reconfigure the beam between the base use the absolute value of the two coordinates to determine the base-
station and the UE. We also show that the complex task of angle angle. One can observe that the base angle of a UE with relative
computation can be approximated with match-action tables (MATs) location (𝑎,𝑏) can be computed as 𝛽 = arctan |𝑏 |/|𝑎| (resulting in an
in the data plane and the approximation errors can be bounded. angle in the first quadrant). However, computation of the arctangent
The accuracy of the proposed system is examined in more dynamic function is not supported by P4 and in general far too complex for
scenarios with a moving UE and different control latency, while a P4-programmable switch.
the resource usage of our Tofino-based implementation is also Our approximation method is shown in Figure 2b. We take
analyzed. We believe that P4 and P4-enabled devices can help to a grid with exponential binning and approximate the UE’s loca-
develop unique edge-cloud or fog network architectures, and enable tion with the centroid of the containing grid-cell. These centroid
the provision of network services through which a user can avail points are then used to determine the approximated base-angles
personalized services [2–4, 6]. (𝛽 1 and 𝛽 2 in the figure). The grid points on both axes are the same:
0, 𝐵 1, 𝐵 2, . . . , 𝐵 𝑁 = 𝐷, where 𝐵 is the base number, 𝑁 is the gran-
2 SYSTEM DESIGN ularity factor (the number of bins on the axes) while 𝐷 is the cell
diameter (the maximum range).
Figure 1 depicts the segmented network scenario that is taken into
After acquiring the approximated 𝛽 angle in the first quadrant,
consideration in this paper. Mobile users are attached to 5G network
we finally map it to its correct quadrant by checking the sign of 𝑎
and indirectly connected to switches in the access aggregation
and 𝑏 coordinates, giving us a direct indication of the UE’s position
network by means of core network protocols. We assume that
in the other three quadrants as depicted in Figure 2. After this
some of the switches in this network are P4-programmable. The
step, we obtain the approximated angle (𝛼 1′ and 𝛼 2′ ) of the beam for
mobile user periodically sends its GPS location to a P4-switch, the
the given UE that needs to be set. The angle information is then
P4-switch computes the angle of user equipment (UE) around the
encapsulated into a beam configuration message and sent to the
corresponding base station and send a configuration message to
corresponding base station.
the base station to direct beam to the UE. We believe that this
The implementation design of the Angle Approximation method
user-assisted in-network method can reduce the control latency
in the data plane is shown in Figure 3. When a location message
and increase the steering of beam much smoother and faster than
is received, it is parsed and the UE’s location with information
traditional approaches.
about the base station is stored in a new header instance (𝑙𝑜𝑐). The
2.1 Angle Approximation in the Data Plane base station coordinates (𝑏𝑠.𝑥, 𝑏𝑠.𝑦) are resolved by an exact match-
Figure 2 summarizes the key conceptual steps of our in-network action table (MAT), mapping some identifier of the base station
method for beam angle computation. Figure 2a shows the positions (e.g., GTP-TEID, IP, etc.) to its absolute location. Since P4 language
of two UEs inside the range of a base station and the desired Angles does not support high-level computations such as arctan and ab-
of Arrival (𝛼 1 , 𝛼 2 ) that needs to be determined and used to steer the solute value (|𝑥 |), we calculate the relative coordinates in the first
two beams. The absolute coordinates of UEs are first transformed quadrant (𝑚.𝑎 = |𝑎|,𝑚.𝑏 = |𝑏 |) by comparing the coordinates of the

62
In-network Angle Approximation for Supporting Adaptive Beamforming EuroP4 ’22, December 9, 2022, Roma, Italy

Alpha2' =
180°- Beta2
Alpha2

Alpha1 Beta2 Alpha1' = Beta1

Beta1

(0,0)
(a) Direct Angle of Arrival (b) Base-angle Approximation on a Grid (c) Final Angle Transformation

Figure 2: Key steps of beam angle approximation with an exponentially binned grid.

base station and the UE and subtracting the higher value from the and for different cell diameters ranging from 200m to 1km. During
lower using simple if-statements, we also define variables (𝑚.𝑠𝑖𝑔𝑛𝑎, this experiment, we have taken 100.000 UE locations uniformly at
𝑚.𝑠𝑖𝑔𝑛𝑏) to hold the signs of 𝑎 and 𝑏, respectively. Note that in random within a specific diameter around a base station and com-
our implementation we involved extra MATs for these computa- puted the angle approximation error of our MAT-based method.
tions instead of a simple if-statement and comparison. The angle We also assume that the user cannot be closer to the base station
approximation then relies on two match-action tables. Table Angle than five meters which hold in most outdoor scenarios.
Approximation does range matching on keys 𝑚.𝑎 and 𝑚.𝑏. The One can observe that the applied grid granularity has a higher
table is filled with constant range entries representing the grid with impact on the approximation error than the cell diameter. The
exponential binning as shown in Figure 2b. Each entry represents maximum errors for cases 𝑁 =20,40,60 and 80 are approx. 8, 5, 4 and
a grid cell and the related action sets the cell centroid-based ap- 2.5 degrees respectively. From a practical point of view, larger errors
proximation of base-angle 𝛽. Note that this table requires the most require wider beams. In static cases, the beam needs to be two times
resources and in most P4-programmable hardware it is mapped wider than the maximum approximation error. For example, if the
to Ternary Content Addressable Memory (TCAM). The required beam is 5 degrees wide and the error is at most 2.5 degrees, the
TCAM space is mainly affected by the desired angle-approximation system can ensure that the UE still remains within the beam. In
accuracy as we will show in Section 3. As a last step, Table Final addition to the empirical analysis, we can also show a theoretical
Angle Transformation is applied that matches on keys 𝑚.𝑠𝑖𝑔𝑛𝑎, bound on the error of the applied grid-based angle approximation.
𝑚.𝑠𝑖𝑔𝑛𝑏 to obtain the final angle approximation 𝛼 ′ . For example,
Lemma 1. The absolute error of the proposed
√ MAT-based angle
the exact matched value (1,0) of 𝑚.𝑠𝑖𝑔𝑛𝑎 and 𝑚.𝑠𝑖𝑔𝑛𝑏 respectively,
approximation is bounded by 2 × arctan( 𝐵) − 𝜋2 (in rads), where 𝐵
indicates that the location of UE is in the second quadrant and
therefore the action to obtain final 𝛼 ′ results in 180 − 𝛽 (in degrees). is the base number in 𝐵 𝑁 = 𝐷, for a predefined maximum distance
Finally, the location header is removed and the control header is 𝐷 and granularity factor 𝑁 .
added, filled and the packet is readdressed before deparsing and Proof. (Proof sketch.) If the user’s relative coordinates (𝑎, 𝑏)
forwarding to the designated base station. lay in a cell grid |𝑎| ∈ [𝐵𝑖 , 𝐵 (𝑖+1) ) and |𝑏 | ∈ [𝐵𝑘 , 𝐵 (𝑘+1) ) for any
𝑖, 𝑘 positive integers. The angle is approximated according to the
3 PROOF OF CONCEPT central point of the cell. The deviation from the real angle is the
We have implemented the proposed method in P4 using the TNA largest in one of the corner points of the cell. We can define the error
(Tofino Native Architecture) architecture model and evaluated it functions for the corner points in this way and take the derivative
on a Tofino ASIC. During the evaluation we focus on three aspects: in 𝑖 and 𝑘 to find the location of maximum errors. After that, we
1) The approximation error of MAT-based angle estimation with can get the error bound trivially. □
different granularity factors (number of bins (𝑁 =20,40,60 and 80), 2) One can observe that this theoretical bound is in accordance
the effect of moving UEs and non-zero control cycle delays on the with the above empirical error distributions. For example, in cell
method’s accuracy, and 3) the relationship between the utilization of with diameter of 800 meters, the theoretical error bounds for gran-
Tofino resources (stages, SRAM, TCAM memories) and the applied ularity factors 𝑁 =20,40,60 and 80 are 9.5, 4.8, 3.2 and 2.4 degrees
granularity factor 𝑁 . respectively.

3.1 Approximation Error 3.2 Moving UEs and Control Latency

Figure. 4 depicts the absolute error distribution of angle approxi- To demonstrate our concept in a more dynamic scenario, we
mation for grids with different granularity (𝑁 =20, 40, 60 and 80) consider a small cell in a semi-urban environment located at the

63
EuroP4 ’22, December 9, 2022, Roma, Italy Hiba Mallouhi, Jaspreet Kaur, Hasan Tahir Abbas, and Sándor Laki

Location Configuration
Message Message

Table Angle Approximation

Table Final Angle
m.a:range, m.b:range Action Transformation
Resolve BS Info Preprocessing Computation
(0 .. 136, 0 .. 136) m.beta = 45
m.signa, Action
loc.bsid Action (0 .. 136, 136 .. 186) m.beta = 67 m.signb
m.a = | loc.x-bs.x |
bs.x = 120 m.b = |loc.y-bs.y| … … (0, 0) m.alpha = m.beta
4432 bs.y = 320 m.signa = sign(loc.x-bs.x) (C*Bj-1 .. C*Bj, C*Bk-1 .. C*Bk) m.beta = (1, 0) = 180 – m.beta
bs.ip = …
m.signb = sign(loc.y-bs.y) 180 𝐵𝑘 +𝐵𝑘−1
arctan( 𝑖 𝑖−1 ) (1,1) = 180 + m.beta
… … 𝜋 𝐵 +𝐵

… … (0,1) = 360 – m.beta

Stage 1 Stage 2 Stage 3 Stage 4 … … Last Stage

Figure 3: Data Plane Design of the Angle Approximation

1.0 1.0 1.0 1.0

0.8 0.8 0.8 0.8

0.6 0.6 0.6 0.6

CDF

CDF

CDF

CDF
200 m 200 m 200 m 200 m
0.4 400 m 0.4 400 m 0.4 400 m 0.4 400 m
600 m 600 m 600 m 600 m
0.2 800 m 0.2 800 m 0.2 800 m 0.2 800 m
1000 m 1000 m 1000 m 1000 m
0.0 0.0 0.0 0.0
0 5 10 0 5 10 0 5 10 0 5 10
Absolute Error [degrees] Absolute Error [degrees] Absolute Error [degrees] Absolute Error [degrees]

(a) 𝑁 =20 (b) 𝑁 =40 (c) 𝑁 =60 (d) 𝑁 =80

Figure 4: Error Distribution of MAT-based Angle Approximation.

of angle-approximation and beam reconfiguration requires more

than zero time.
We assume that the UE periodically sends a location message to
the P4-based angle approximation node that - as a result - generates
a beam configuration message and forwards it to the correspond-
ing base station. The round trip delay between the departure of
the location information message from the UE and the arrival of
control message to the base station is varied in a range of 5, 20 and
100ms. The angle approximation errors depicted in Figure. 6-7 were
calculated at the time of arrival of the configuration message to the
base station, thus it also takes the user’s movement into account.
Figure. 6 and Figure. 7 prove that larger granularity factor (𝑁 )
Figure 5: University campus showing the transmitter and the results in better angle approximation even when the control cycle
mobile user. is longer and the UE is moving at a moderate speed. Comparing the
case of 𝑁 = 20 to 𝑁 = 80, one can see that the increase in the error
caused by the latency and movement is less important for 𝑁 = 20
when the base error is also high (around 9 degrees). The error
campus of the University of Glasgow, UK. Figure 5 shows the base
distributions are similar to our static scenarios depicted in Figure. 4.
station, located on James Watt South (JWS) building in the campus.
The large deviance can only be seen when the control cycle time is
A user is assumed to be moving on a route at a constant speed
100ms and the UE moves fast (>90kmph), as depicted in Figure. 7c. It
through the scene shown as a red arrow. The route denoted covers
also indicates that at high speed wider beams are needed otherwise
the Byres Road, James McCune Smith (JMS), Library and Glasgow
the UE can easily get out of the coverage as the beam cannot keep
University Union (GUU). The user’s constant speed is varied from
track of the UE. Note that a real-world environment is more complex
5 to 200kmph (pedestrian to high-speed trains). Note that we only
where the radio propagation is influenced by many environmental
emulate the movement of the UE in this scenario and evaluate the
factors, also affecting the UE’s quality of experience.
effect of displacement in the UE’s location when the feedback loop

64
In-network Angle Approximation for Supporting Adaptive Beamforming EuroP4 ’22, December 9, 2022, Roma, Italy

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

CDF

CDF

CDF
5 kmph 5 kmph 5 kmph
0.4 50 kmph 0.4 50 kmph 0.4 50 kmph
90 kmph 90 kmph 90 kmph
0.2 130 kmph 0.2 130 kmph 0.2 130 kmph
200 kmph 200 kmph 200 kmph
0.0 0.0 0.0
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
Absolute Error [degrees] Absolute Error [degrees] Absolute Error [degrees]

(a) Control Cycle=5ms (b) Control Cycle=20ms (c) Control Cycle=100ms

Figure 6: Angle approximation error when the user moves at a constant speed and the control cycle is not zero. 𝑁 is set to 20.

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

CDF

CDF

CDF
5 kmph 5 kmph 5 kmph
0.4 50 kmph 0.4 50 kmph 0.4 50 kmph
90 kmph 90 kmph 90 kmph
0.2 130 kmph 0.2 130 kmph 0.2 130 kmph
200 kmph 200 kmph 200 kmph
0.0 0.0 0.0
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
Absolute Error [degrees] Absolute Error [degrees] Absolute Error [degrees]

(a) Control Cycle=5ms (b) Control Cycle=20ms (c) Control Cycle=100ms

Figure 7: Angle approximation error when the user moves at a constant speed (5kmph to 200kmph) and the control cycle is
5ms, 20ms or 100ms. 𝑁 is fixed to 80.

3.3 Resource Usage SRAM usage is almost constant and independent of the granularity
factor. The grid-based angle approximation is implemented by a
50 TCAM table and thus the granularity factor has a strong correlation
SRAM/TCAM Usage [%]

8 TCAM #Stages
with the required TCAM space. For 𝑁 = 20 and 40, the TCAM table
SRAM 40
6 only occupies a single stage, for 𝑁 =50 and 60 the TCAM table is
#Stages

30 distributed among two stages while in the case of 𝑁 =80 three stages
4 are required to store this table. Note that the per-stage TCAM space
20
is limited. The number of stages is only affected by the size of the
2 10 TCAM table. Though the proposed method occupies the TCAM
space on one or more stages, its per-stage resource usage is not
0 0 significant in general thus other network functions/pipelines (e.g.,
20 40 50 60 80
Granularity factor (N) related to mobile transport or RAN) can also be co-located with the
proposed pipeline. The TCAM space of stages implementing the
MAT-based angle approximation table is in essence fully occupied,
Figure 8: The trade-off between granularity factor (𝑁 ) and
but other resources like SRAM and ALUs can potentially be used
resource usage (stages, SRAM and TCAM) on a Tofino ASIC.
in these stages by other pipelines.

We have shown in the previous sections that the grid resolu- 3.4 Customization
tion defined by the granularity factor 𝑁 has the highest impact on The surrounding around a base station may be different and thus
the achievable angle approximation accuracy. A larger 𝑁 results radio propagation would not occur along a straight line because of,
in better accuracy, but it requires a much larger TCAM space to e.g., reflections from buildings and other objects. To handle these
implement the grid searching logic in MAT. Figure. 8 depicts the re- effects, the customization of the proposed method may be needed.
lationship between the applied granularity factor 𝑁 and the number Instead of applying the global model implemented by our Angle Ap-
of required stages, SRAM and TCAM usage. One can observe that proximation table, we could 1) introduce customized tables for each

65
EuroP4 ’22, December 9, 2022, Roma, Italy Hiba Mallouhi, Jaspreet Kaur, Hasan Tahir Abbas, and Sándor Laki

base station and 2) extend the angle approximation grid to all the seen by the UE. We also will investigate how our generalized model
possible quadrants around the base station, handling asymmetric can efficiently be customized so that we could take into account
signal propagation in the cell. In the latter case, the approximation the differences in the base stations’ surroundings (e.g., buildings
table requires four times more TCAM entries for covering all the and other obstacles influencing the signal propagation) without
four quadrants of the cell. However, in the first case we need to significantly increasing the required TCAM space.
create separate tables for each base station, multiplying the needed
TCAM space. Our future work includes the more detailed analysis ACKNOWLEDGMENT
of customization scenarios and their efficient realization. The research was supported by the SICSA Saltire Emerging Re-
searcher Scheme and the National Research, Development and
4 RELATED WORK Innovation Office – NKFIH, FK 138949.
New ways have been explored like channel sounding [11] in which
Synchronization signal block (SSB) and channel state information REFERENCES
[1] 3GPP. 2018. NR; Physical channels and modulation. 3rd Generation Partnership
reference signal (CSI-RS) for 5G downlink transmissions in accor- Project (3GPP), Technical Specification (TS) 38.211 9 (2018).
dance with the New Radio (NR) technical specifications suggested [2] Frederik Hauser, Marco Häberle, Daniel Merling, Steffen Lindner, Vladimir Gure-
by 3GPP has been extracted and SAGE algorithm has been used vich, Florian Zeiger, Reinhard Frank, and Michael Menth. 2021. A survey on
data plane programming with p4: Fundamentals, advances, and applied research.
to estimate the channel parameters. Their work gives the hope arXiv preprint arXiv:2101.10632 (2021).
to introduce location-aided beamforming if we are able to extract [3] Enio Kaljic, Almir Maric, Pamela Njemcevic, and Mesud Hadzialic. 2019. A survey
on data plane flexibility and programmability in software-defined networking.
the channel parameters. Apart from this, recently [9] proposed
IEEE Access 7 (2019), 47804–47840.
P4 based In band telemetry (INT) fully integrated with UEs. De- [4] Sukhveer Kaur, Krishan Kumar, and Naveen Aggarwal. 2021. A review on P4-
centralized steering rules, such as those involving UE and specific Programmable data planes: Architecture, research efforts, and future directions.
Computer Communications 170 (2021), 109–129.
switches, are made possible by the INT mechanism, which is built [5] Mehdi M Molu, Pei Xiao, Mohsen Khalily, Kanapathippillai Cumanan, Lei Zhang,
to deliver synchronized and precise end-to-end latency and geo- and Rahim Tafazolli. 2017. Low-complexity and robust hybrid beamforming
location information without the need for software-defined radio design for multi-antenna communication systems. IEEE Transactions on Wireless
Communications 17, 3 (2017), 1445–1459.
(SDN) controller intervention. They also incorporate a forecast sys- [6] Francesco Paolucci, Filippo Cugini, Piero Castoldi, and Tomasz Osiński. 2021.
tem helped by artificial intelligence that can anticipate latency and Enhancing 5G SDN/NFV edge with P4 data plane programmability. IEEE Network
localization in advance and initiate faster edge steering. 35, 3 (2021), 154–160.
[7] Theodore S Rappaport, Shu Sun, Rimma Mayzus, Hang Zhao, Yaniv Azar, Kevin
Both these works are giving motivation that in future for more Wang, George N Wong, Jocelyn K Schulz, Mathew Samimi, and Felix Gutierrez.
5G communications operations, such as mmWave communications, 2013. Millimeter wave mobile communications for 5G cellular: It will work! IEEE
access 1 (2013), 335–349.
which is being developed, and vehicle to everything (V2X), which [8] Wonil Roh, Ji-Yun Seol, Jeongho Park, Byunghwan Lee, Jaekon Lee, Yungsoo
was recently added in the 3GPP technical specification [1], will be Kim, Jaeweon Cho, Kyungwhoon Cheun, and Farshid Aryanfar. 2014. Millimeter-
commercially accessible. Then we can use the proposed frameworks wave beamforming as an enabling technology for 5G cellular communications:
Theoretical feasibility and prototype results. IEEE communications magazine 52,
enabling location-assisted beamforming. 2 (2014), 106–113.
Our proposed solution is to establish a direct link between UE [9] Davide Scano, Francesco Paolucci, Koteswararao Kondepu, Andrea Sgambelluri,
Luca Valcarenghi, and Filippo Cugini. 2021. Extending P4 in-band telemetry to
and base station so that BS that consists of an advanced antenna user equipment for latency-and localization-aware autonomous networking with
system can direct the signal or beam in the direction of arrival from AI forecasting. Journal of Optical Communications and Networking 13, 9 (2021),
the user. D103–D114.
[10] Shree Krishna Sharma, Tadilo Endeshaw Bogale, Long Bao Le, Symeon Chatzino-
tas, Xianbin Wang, and Björn Ottersten. 2017. Dynamic spectrum sharing in 5G
5 CONCLUSION wireless networks with full-duplex technology: Recent advances and research
challenges. IEEE Communications Surveys & Tutorials 20, 1 (2017), 674–707.
The estimation of the direction of arrival is crucial to any adaptive [11] Tianqi Wu, Xuefeng Yin, Linjian Zhang, and Jie Ning. 2021. Measurement-based
beamforming strategy in wireless communications. Conventionally, channel characterization for 5G downlink based on passive sounding in sub-6
GHz 5G commercial networks. IEEE Transactions on Wireless Communications
beamforming algorithms run at the link layer of the network and 20, 5 (2021), 3225–3239.
this introduces additional latency when backhaul networks are con- [12] Xianyue Wu, Cheng-Xiang Wang, Jian Sun, Jie Huang, Rui Feng, Yang Yang, and
Xiaohu Ge. 2017. 60-GHz millimeter-wave channel measurements and modeling
sidered. With a view to making the networks more responsive, we for indoor office environments. IEEE Transactions on Antennas and Propagation
have proposed a new user-assisted beamforming strategy that ex- 65, 4 (2017), 1912–1924.
ploits the advantages of programmable data planes and in-network [13] Ming Xiao, Shahid Mumtaz, Yongming Huang, Linglong Dai, Yonghui Li, Michail
Matthaiou, George K Karagiannidis, Emil Björnson, Kai Yang, I Chih-Lin, et al.
computation using P4. Accordingly, the user equipment (UE) peri- 2017. Millimeter wave communications for future mobile networks. IEEE Journal
odically reports its location to the network and the computation on Selected Areas in Communications 35, 9 (2017), 1909–1935.
of beam parameters (the angle of arrival in this paper) are done by [14] Xuan Xue, Yongchao Wang, Xianbin Wang, and Tadilo Endeshaw Bogale. 2016.
Joint source and relay precoding in multiantenna millimeter-wave systems. IEEE
one of the programmable switches inside the network. Our prelim- Transactions on Vehicular Technology 66, 6 (2016), 4924–4937.
inary results show that the additional computation performed to [15] Kegen Yu and Y Jay Guo. 2008. Statistical NLOS identification based on AOA,
TOA, and signal strength. IEEE Transactions on Vehicular Technology 58, 1 (2008),
divide the cell area into binned grid provides angle approximations 274–286.
with error bounds and requires moderate per-stage resources. The
proposed scheme was also demonstrated on a Tofino ASIC.
As part of our future work, we aim to integrate the P4-based
angle approximation with a realistic radio propagation simulator,
enabling us to quantify the effect of our method on channel quality

66