Message Prioritization for Autonomous Vehicle Communication

IN

DEGREE PROJECT TECHNOLOGY, FIRST CYCLE, 15 CREDITS

STOCKHOLM SWEDEN 2018 ,

Message Prioritization for Autonomous Vehicle

Communication

ERIK TÄLLBERG SIMON EDSTRÖM

KTH ROYAL INSTITUTE OF TECHNOLOGY

SCHOOL OF ELECTRICAL ENGINEERING AND COMPUTER SCIENCE

www.kth.se

INOM

EXAMENSARBETE TEKNIK, GRUNDNIVÅ, 15 HP

STOCKHOLM SVERIGE 2018 ,

Prioritering av meddelanden för kommunicerande autonoma fordon

SIMON EDSTRÖM ERIK TÄLLBERG

KTH

SKOLAN FÖR ELEKTROTEKNIK OCH DATAVETENSKAP

www.kth.se

Sammanfattning/abstract på Svenska och Engelska

Svenska:

För att autonoma fordon ska fungera säkert och effektivt behöver dem kommunicera med varandra och infrastrukturmoduler. Meddelanden som är kopplade till säkerhet är viktigare än de som relaterar till t.ex. underhållning men dagens kommunikationsprotokoll kan inte prioritera mellan dessa

meddelanden. I denna rapport undersöker vi möjligheterna för sådan prioritering genom att

vidareutveckla enkla medium access control protokoll för att kunna hantera detta. Särskilt analyseras protokollen slotted-ALOHA och frequency division multiple access. Prestandan för dessa modifierade protokoll utvärderas sedan och jämförs mellan varandra med hjälp av egenbyggda simuleringar. Vi drar slutsatsen att det är möjligt att prioritera mellan meddelanden av olika typer och att en

kombination av frequency division multiple access och slotted-ALOHA ger bättre prestanda än när man använder endast ena.

Engelska:

For autonomous vehicles to operate safely and efficiently they need to communicate messages with

each other and the roadside infrastructure. Messages related to safety are more important than

messages related to e.g. entertainment but the communication protocols of today do not allow for

prioritization between these message types. In this report we analyze the possibility of such

prioritization by extending simple medium access control protocols such as slotted-ALOHA and

frequency division multiple access to allow for this. The performance of these extended protocols are

then evaluated and compared using custom made simulations. We conclude that it is possible to

prioritize between message types and that a combination of extended frequency division multiple

access and slotted-ALOHA yields better performance results than one used separately.

D1: AUTONOMOUS CARS NETWORKING

Message Prioritization for Autonomous Vehicle Communication

Erik T¨allberg and Simon Edstr¨om

Abstract—For autonomous vehicles to operate safely and effi- ciently they need to communicate messages with each other and the roadside infrastructure. Messages related to safety are more important than messages related to e.g. entertainment but the communication protocols of today do not allow for prioritization between these message types. In this report we analyze the possibility of such prioritization by extending simple medium access control protocols such as slotted-ALOHA and frequency division multiple access to allow for this. The performance of these extended protocols are then evaluated and compared using custom made simulations. We conclude that it is possible to prioritize between message types and that a combination of extended frequency division multiple access and slotted-ALOHA yields better performance results than one used separately.

I. I NTRODUCTION

As we move towards more automated societies it is only natural that this transition happens within the automotive sector as well. In recent years there has been an increase in interest regarding self-driving cars and for these to operate efficiently and safely they need to communicate with each other [1]. Many traffic situations include many cars operating within the same area at the same time which poses a problem for the current methods of communication [2]. Whilst it is possible for cars to communicate in these situations it may not be with an acceptable performance to allow the many benefits of self driving cars. Therefore new communication solutions need to be designed, modeled and analyzed to allow seamless communication between autonomous vehicles [3].

In this report we aim to analyze basic communication pro- tocols, extend these protocols to also allow for prioritization between different message types, evaluate and compare these protocols with the help of analytical models and simulations.

Firstly, in section II, we present the requirements for au- tonomous vehicles, our choice of physical layer properties, some relevant MAC protocols and tools for modeling these.

Then we present our extended protocols in section III and how we model and simulate these in section IV. After that we present our results in section V to thereafter move into analysis and discussion in section VI and VII respectively. Lastly, we summarize and conclude this paper in section VIII.

II. L ITERATURE STUDY

The literature on the communication requirements for au- tonomous vehicles describe what types of messages an au- tonomous car may use [1] with respective performance re- quirements [4], [5], what physical layer properties we can expect [2], which different multiple access protocols are applicable [6] and how to model and evaluate these [7].

A. Types of messages used in autonomous vehicles and their requirements

From [1] it can be deduced that there are two main types of messages used in autonomous vehicles (AVs):

1) Messages which are essential to the vehicle’s operation and safety, containing information such as position, direction, speed, intention etc. Such messages are called

”control messages”. For the purpose of this report, all these messages will be referred to as ”high-priority messages” (HPM).

2) Messages which are ”nice-to-haves” but are not essential for the passenger’s safety. Such messages could include entertainment and other services. For the purpose of this report, these messages will be referred to as ”low-priority messages” (LPM).

The requirements of these message-types vary [4], but for the purposes of this paper, requirements and specifications where chosen from [5] and [8] according to:

1) High-priority - automated overtake:

• End-to-end latency (maximum delay): 10 ms.

• Reliability: 99.999% of all packets must make it through within the specified end-to-end latency.

• One packet is sent every 10 ms, yielding a high- priority packet (HPP) input rate of λ 0,HP = 100 packets/s.

2) Low priority - watching 1080p quality Netflix:

• Average bitrate of R _nx = 5 Mbit/s which yields a low priority packet (LPP) input rate of λ 0,LP = ^R _L

= 2083 packets/s, where L is the packet size.

Regarding the packet size of high and low priority messages, we use a packet size of L = 300 bytes. This is what is needed for a low degree of automation according to [5]. In reality, the packet size for low priority messages can be larger but for reasons of simplicity, we instead chose to increase the input rate of these packets so that the average bitrate becomes 5 Mbit/s as stated above.

Our goal here is to modify existing link layer communi-

cation protocols to prioritize between the messages of high

priority and messages of low priority. This should be done in

a way so that the high-priority messages exactly match their

requirements (with an added safety factor) and then use the

remaining resources for low priority messages.

D1: AUTONOMOUS CARS NETWORKING

B. Physical layer properties

The research and development on ”vehicle-to-infrastructure and vehicle-to-vehicle communication” (V2X) the last years has led to efforts in standardization for V2X technologies.

The European Telecommunications Standards Institute (ETSI) has in its European Release 1 defined that the most relevant access technology for V2X communication is ITS-G5 which is similar to IEEE 802.11a [2]. Therefore, in this analysis, we will use the IEEE 802.11a data rate of 54 Mbit/s.

C. Relevant multiple access protocols

Two main protocol types where identified: Slotted- ALOHA (SALOHA) and Frequency Division Multiple Access (FDMA). Other protocols such as Time Division Multiple Access (TDMA), polling and Carrier Sense Multiple Access (CSMA) where also looked at briefly but discarded because of their complexity to implement for our problem (autonomous vehicle communication) and difficulty to prioritize between message types. However, later on, we will compare the chosen protocols with the industry standard and thus a brief explana- tion of CSMA is also given.

1) SALOHA: All of the allotted bandwidth is shared be- tween message types. All users using the network transmit data-packets immediately when they have something to send.

This creates collisions between the packets and if two or more packets collide, all of their information can be assumed to be lost. The term ”slotted” here refers to the fact that users are only permitted to start sending their packets at certain time- intervals. This is to avoid a small part of the start or end of the message colliding, as even a small collision can destroy the entirety of the message. When a packet collides, the transmit- ting user sends the packet again after a certain time interval called a ”backoff time”. When a packet has collided and is awaiting retransmission we call that packet ”backlogged”. One of the main benefits of SALOHA is the fact that no central controller is necessary for message-coordination as is the case with e.g. TDMA or CSMA; its throughput, delay and stability are however lower than that of other protocols. If the channel becomes oversaturated with messages then it becomes unstable and no messages are transmitted of the channel, a more formal definition of this is discussed later in this report. [7]

2) FDMA: The available frequency band is split up into multiple sub-bands. These frequency sub-bands then act inde- pendently from each other, working as separate communica- tion channels. Each channel is allotted to a single user which tunes its filters to send and receive only in its own channel.

Its simplicity is advantageous but when some users don’t have anything to send some of the available bandwidth resources is wasted. [6]

3) CSMA/CA: Carrier sense multiple access with collision avoidance is a protocol that is similar to (unslotted) ALOHA.

However, the main difference is that a user starts by listening to the channel before transmitting and if it is not busy it transmits. If the channel is perceived as busy it has to wait and perform a backoff procedure. [7]

D. Mathematical modeling of multiple access protocols The performance of FDMA and SALOHA is discussed in [7]. We have chosen to focus on throughput, S, and delay, D, for the different protocols. First, however, a common nomenclature must be established for the analysis:

• λ HP , λ LP : amount of packets being sent per second on average for high-priority and low-priority packets respectively.

• R B : available bitrate for the physical channel in bits per second.

• L: the size of one packet in bits.

• t p : amount of time it takes for one packet to send in seconds, also called ”slot-time”.

• N _c : amount of cars using the channel.

For all applications, the slot time t p is always given by t _p = L

R B

(1) 1) SALOHA: Two main equations from [7] are given for the throughput and delay respectively:

S _SALOHA = G(n)e ^−G(n) (2) and

D _SALOHA = E[n _retx ]E[t _retx ] + 3

2 t _p (3)

which can be expressed as

D SALOHA = (e ^G(n) − 1)( 3 2 t p + 1

β ) + 3

2 t p (4) An explanation of the parameters within these equations follows.

Because of the fact that the channel both has new packets as well as old packets being retransmitted, after colliding the total input to the channel can be modeled as:

λ _t (n) = λ + nβ (5)

where n refers to the number of backlogged packets, 1/β to the average amount of time before a packet is retransmitted and λ as the external input to the system. In [7], backoff times are chosen at random according to the probability density function (PDF):

p _t

(a) = βe ^−βa , a ≥ 0 (6) where β is a system parameter here chosen as β = 10t _p . G(n) refers to the average amount of packets that should be transmitted per time-slot and is given by

G(n) = λ t (n)t p (7)

This, however, raises an issue for modelling as the number of backlogged users is not known. However, through manip- ulation of equations in [7] such as:

G(n)

S _SALOHA = λ t (n)

λ = 1

P _succ = 1

e ^−G(n) (8) we can express λ as a function of G(n) according to:

λ = G(n) t p

e ^−G(n) (9)

D1: AUTONOMOUS CARS NETWORKING

To find G(n) not as a function of the number of backlogged packets, n, but instead as a function of the external input, λ, we simply have to invert equation 9 which gives us:

G(n) = −W ₀ (−t _p λ) = G(λ) (10) where W ₀ is the principal branch of the Lambert function.

There are two more things to note, the first is that from equation 3, E[n retx ] refers to the average amount of packets that are being retransmitted and E[t retx ] refers to the average time it takes for retransmitted packets to successfully transmit.

The second is that the total external input to the system, λ is given as

λ = (λ 0,HP + λ 0,LP )N c (11) where λ 0,HP and λ 0,LP and refers to the average number of packets being sent per second for one car for high- and low priority messages respectively. This gives us enough information to express S SALOHA and D SALOHA as functions of our known inputs. For easier analysis, D SALOHA can be normalized and expressed in amount of slot-times instead of amount of seconds according to:

D = e D

t _p = (e ^G(λ) − 1)( 3 2 +  g 1

β

 ) + 3

2 (12)

2) FDMA: The two main equations for delay and throughput as stated in [7] are:

S F DM A = λt p (13)

and

D _{F DM A} = N _c

λ (S _{F DM A} + (S _{F DM A} ) ²

2(1 − S F DM A ) ) + τ _p (14) where τ p is the propagation delay here set to 0.

The fixed rate R b can be divided in different ways, e.g.

into a number of equally large channels or into two channels with a channel split ratio of γ ⊂ [0, 1]. For the first, each of the channels get R b /N _c for N c number of cars and for the second, each channel gets the two channels get R b γ and R _b (1 − γ). Equations 13 and 14 are for the case when the available frequency band is divided into N _c equally large sub- bands.

III. E XTENSION OF BASIC PROTOCOLS

The basic protocols outlined in the previous section cannot inherently differentiate between different types of signals. To achieve this, the protocols must be extended.

A. SALOHA

The SALOHA protocol described in [7] uses a backoff algorithm whereby once a packet has collided, its backoff time is chosen according to equation 6. This is however fairly unstable according to [7] which is also shown later in section IV-A3. A better approach is to make the backoff time be randomly chosen from a time-window with uniform probability. We call the start of this window t b1 and the end of the window t b2 . Therefore, once a packet collides it will

land somewhere between these two at random as illustrated in figure 1. This method is still generally unstable but if we double t b2 for the packet every time it crashes the system becomes much more stable. We call such a system one of

”quadratic backoff windows”.

Fig. 1. A graphical illustration of an example system using the extended SALOHA protocol. It is illustrated that HPP and LPP can both collide and not collide, that if they collide they try to retransmit and that the HPPs backoff window is smaller than the LPP backoff windows (t

< t

).

Now, to differentiate between different types of packets we can vary the initial window size, t b20 , as this will consid- erably affect the following window sizes of the packet as it continuously collides before finally transmitting correctly.

For high-priority packets we use a smaller t b20 than for low priority packets, hopefully leading to a smaller delay and higher throughput.

B. FDMA+SALOHA

The FDMA protocol described in [7] allots a channel to one specific user no matter how the frequency band is divided. To instead allot the channel to a specific message type that can be sent from all users, a modification of the protocol to handle the traffic in each channel is needed. We chose to do this by combining FDMA and SALOHA as illustrated in figure 2.

Fig. 2. A graphical illustration of the extended protocol FDMA+SALOHA.

It is illustrated that the total available frequency is divided into two sub-bands with split ratio γ. The upper sub-band is allotted to HPM and receives γR

of the total bandwidth while the lower sub-band is allotted to LPM which

receives (1 − γ)R

. Within each sub-band there is a SALOHA protocol. Of

note is that the two sub-bands will have different slot times unlike in the

figure as they will have different bitrates.

D1: AUTONOMOUS CARS NETWORKING

In our case, we have two types of messages and choose to divide the systems fixed rate R _b into two channels with the split ratio γ ⊂ [0, 1]. For each channel, there is a SALOHA protocol that handles the traffic that arises when many cars are using the channel. All messages are of the same type in each channel but they are sent from different users, meaning that packets can still collide and have to be retransmitted. In this case, unlike for the extension of pure SALOHA, the initial window size for the backoff times are the same for all packets and message types but qaudratic backoff windows are still used.

IV. M ODELING AND SIMULATION METHOD

A. SALOHA

To evaluate how our protocol with quadratic backoff times works we can either use analytical models as described in section II or we can build a simulation to see what actually happens when we implement our protocol. Both of these have their respective advantages and disadvantages.

1) The simulation software: A simulation software was created in MATLAB. In order to better understand what happens during a simulation, a graphical representation was developed. A simple demonstration with explanatory text can be seen in figure 3. The real simulated system more closely resembles that in figure 4 which uses a simulation time (t end ) of 100 slots. For the real simulation, the system is simulated for hundred or thousands of slots. An unstable system is shown in figure 5. We define this to be the case when the external input λ is greater than than the throughput S for a normalized offered load G greater than 1, see figure 6. This can essentially be described as the input of the system (λ) being greater than the output of the system (S when G > 1).

The simulation can be used to look at the effect of various parameters such as t b20 , γ, simulation runtime t end , message length L, the maximum number of collisions before resetting t _b2 etc. and combine them with different amount of cars to observe system performance. It also allows choosing from different backoff-protocols to compare their performance.

Once the simulation environment is set up we measure performance indicators of interest. The software gives the following outputs, anything can however be studied within the simulated system if one chooses:

• Normalized system throughput, S.

• Normalized delay for high- and low priority packets, D 1

and D ₂ . Only packets which are successful have their delay time recorded.

• Successful packets within the simulation time, P 1 and P 2 .

• Number of collisions for high- and low priority packets, n c1 and n c2 .

• Average number of backlogged packets at every time slot, m nblp

• Average normalized offered traffic, G

• If the system was stable or not according to our previous definition, P stable

• The percentage of packets that succeeded within the required end-to-end latency, P e2e , as defined in section II-A.

Fig. 3. A graphical illustration of what happens during a simulation runtime, here for a simple case with low λ. New packets have brighter colors while resent packets get darker the more times they have collided. At slots 1, 3 and 4, three LP packets arrive and instantly succeed. At slot 8, a HP and LP packet arrives at the same time and collide. The HPP waits for 1 slot and then resends, and by chance (25%) the LPP waits the same amount.

At slot 10, another LPP arrives. The two HPP and LPP packets from slot 8 have now both collided twice meaning they will wait anywhere from 1 to t

2

= 1 ∗ 2

= 4 and 1 to t

2

= 4 ∗ 2

= 16 slots respectively. This works out well in this specific simulation snapshot as all packets arrive at slightly different times from slots 13 to 15, with a new LP packet arriving at slot 16. To clarify further, six new LP and one new HP packet has arrived in this simulation runtime. For further simulations, the backoff window start (t

) is equal to 0 and not 1 as it is here. For this snapshot, S = 7/18 and G = 12/18.

Fig. 4. A graphical illustration of a simulation with a higher λ and longer simulation time. This system has a high load but is not yet unstable.

Because of the fact that the results vary for each simulation, an average from N sim number of simulations is always used.

The simulation takes a long time to run and to lower this

calculation time, the simulation can be tuned not to run if

it detects an unstable configuration or if it knows that a

simulation will be unstable.

D1: AUTONOMOUS CARS NETWORKING

Fig. 5. A simulation which yielded unstable results, showing what happens when λ becomes too large for the system to handle. Here, the green/black mass of long blocked packets can be seen growing larger and larger when the simulation time increases, indicating a buildup. It is clear here that G > 1 and that λ ≈ 1 is greater than S = 8/101.

Fig. 6. A figure taken directly from [7] describing the stability of SALOHA.

λ can be seen as the input and S can be seen as the output when G is larger than 1. When S becomes smaller than λ the system is no longer stable as the input is greater than the output. This happens at G = G

. Note that this figure is for the analytical equations outlined in section II-D and not using quadratic backoff windows.

2) Modeling vs. Simulation: An analytical model provides a deeper understanding what is going on and allows us to easily change the system configuration as well as exploring a whole space more exactly whilst not taking a long time to calculate. However, the analytical model is only precise in a few cases and requires a lot of approximation, thus making it unsuitable for e.g. large values of λ.

The simulation gives us a more realistic scenario making it more applicable but takes a long time to calculate and can be very dependent on how long it is allowed to run. It also comes with the issue that we only simulate within a short time which becomes an issue when e.g. the backoff time is relatively large compared to the simulation runtime.

Some packets will be recorded as not being successful purely

because they did not succeed within the simulation time, even if they would have succeeded if the simulation would have run longer. Therefore when the load becomes high, the delay measurement sometimes looks to become lower but this is only because the (sometimes very few) packets that made it through are recorded - the rest are lost.

To analyze and compare the model with the simulation, known equations for delay and throughput were taken from [7] and compared with the same parameters within the simulation.

This gave a throughput and delay which can be seen in figure 7 and figure 8 respectively. The backoff times here are not chosen from a window but instead follow the PDF from equation 6.

Fig. 7. System throughput for model from [7] compared to built simulation.

The simulation took averages from N

= 100 runs for each data point.

Fig. 8. Average packet delay for model from [7] compared to built simulation

We see that the model holds true for small values of λ but does not hold true as λ increases. The reason for this becomes very clear as we look at the number of backlogged packets.

From equations 5 and 7, the number of users for the analytical

D1: AUTONOMOUS CARS NETWORKING

model can be found to be modeled as:

n =  G(λ) t p

− λ 

/β (15)

Recording the average amount of backlogged users in the simulation and comparing them to equation 15 gives figure 9.

Fig. 9. Average number of backlogged users for model from [7] compared to built simulation. The ”jag” here at λ = 0.24 is from an anomaly in the simulation.

This shows that the model holds true for small enough values of λ but does not work when λ becomes too large.

This is expected and noted in [7] but no explicit threshold for λ is mentioned. Because of the fact that a central issue within this report is ”how much can we increase the system input λ before our requirements are no longer satisfied?” a choice was made to base the analysis and final results on simulations only.

3) The choice of quadratic backoff window: The reasoning for choosing quadratic backoff windows was partly motivated by the fact that it will become easier to prioritize between message types but in principal the same approach could have been taken with the given exponential distribution. The main reason, as stated earlier, is that the stability of quadratic backoff windows far greater than that of the exponential distribution as shown in figure 10.

4) Correcting for simulation artifacts: As outlined earlier, simulating does not come without its problems. One such issue is that the average- and maximum packet delays are affected by how long the simulation is allowed to run. This is illustrated in figures 11 and 12. This is because as the simulation is allowed to run longer and longer the probability of earlier sent packets succeeding increases. This effect is amplified the larger λ is which is clearly seen when comparing figures 11 and 12 as λ HP < λ LP .

To compensate for this, the largest 1% of all delay values are removed from the calculation of delay as only there are only a small handful of outliers which is seen in 11. Even though there are not just a few outliers but instead very many which

become an issue for LPM, this is not an issue since there is no requirement for low priority delay and as such, this is not of the same importance. This will of course affect the accuracy of the delay values but this is a better approach than allowing t end to have a large impact.

Fig. 10. Stability of quadratic backoff windows for two different values of the initial backoff windows size, t

compared to an exponential distribution.

Fig. 11. How different sending times affect delay for high-priority packets.

A few peaks can be seen but if the largest 1% of them are removed the values are not too disturbed. This effect is however amplified the larger the input rate for HPM.

B. FDMA+SALOHA

Our FDMA+SALOHA protocol can be seen as two separate channels that each use a SALOHA protocol. Therefore, the same simulation method as described earlier is used to analyze this protocol as well. However, some small modifications are needed. Firstly, we set the simulation to only have one message type per channel as they do not interfere for this protocol.

Secondly, the variable being tested has to be set to γ and not

t b20 . Each channel is then analyzed individually as a system

with a normal SALOHA protocol.

D1: AUTONOMOUS CARS NETWORKING

Fig. 12. How different sending times affect delay for low priority packets.

C. Finding a γ and t b20,LP to match the requirements for high priority messages

To find the optimal γ and t b20,LP for our protocols we investigate how the protocols perform in relation to the re- quirements given in section II-A. The limiting requirement is the reliability - that 99.999% of packets must succeed within the specified end-to-end latency. For HPM this requirement can be considered as 100% meaning that every packet has to be successfully transmitted. This is because we would need to run the simulation for 100 000 packet slots before we even allow a single packet to fail and the simulations here only run for a few thousand slots at most.

To find what sub-band split ratio, γ, fulfills the reliability requirements we begin by looking at all HPM data points (combinations of N _c and γ) that are stable. Then, out of these, we look at all points that fulfill the reliability requirement.

But this is not enough since some of these data points belong to the unstable area of the LPM. We have to take this into consideration and remove all of the unstable LPM data points.

What we get is a map of all data points that are in the stable area and that achieve the requirements.

To optimize HPM performance for extended SALOHA we need to find the optimal initial window size for LPM, t b20,LP . This is however a more difficult issue as increasing t b20,LP will increase the performance of HPM and decrease the performance of LPM without end but at a diminishing rate.

In theory one can increase t b20,LP to allow any number of cars but as shown in section V, this is not possible. As such, the question at hand is how many cars each protocol can handle in regards to our HPM requirements and what is acceptable for LPM despite there being no strict requirements.

V. R ESULTS

A. How γ affects performance for FDMA+SALOHA

The sub-band split ratio, γ, plays a large role in the performance of our FDMA + SALOHA protocol. As seen in figure 13 and 14 it directly affects the throughput for both

HPM and LPM respectively. What we see is that the maximum throughput follows a linear relationship between γ and N _c . A larger γ, which means that HPM get most of the bandwidth resources, increases the HPM throughput for a large number of cars. For LPM it is the other way around, a small γ increase the throughput for many cars. We could have guessed this from equation 1 and 13 since R b depends on γ. However, for a smaller number of cars, this is not the case. This is due to the fact that throughput is a measure of saturation for G < 1.

For a constant number of cars, an increase of the bandwidth makes the channel less saturated and therefore the throughput is smaller.

Fig. 13. Throughput for HPM as a function of sub-band split ratio γ and number of cars N

. The dark blue area is where the protocol becomes unstable and no simulations have been done.

Fig. 14. Throughput for LPM as a function of sub-band split ratio γ and number of cars N

. The dark blue area is where the protocol becomes unstable and no simulations have been done

The sub-band split ratio, γ, affects packet delay for both

HPM and LPM as illustrated in figures 15 and 16. It also

affects stability which is seen in the dark blue corner areas as

well as in figure 17. For the stable area, darker colors indicate

a smaller delay.

D1: AUTONOMOUS CARS NETWORKING

Fig. 15. Delay for HPM as a function of sub-band split ratio γ and number of cars N

. The top left area has infinite delay as the system in unstable there.

Fig. 16. LPM delay as a function of sub-band split ratio γ and n.o. cars N

.

Fig. 17. Stable (bright) and unstable (dark) data points for HPM.

We see that the maximum delay, in the same way as throughput, follows a linear relationship with γ and N _c . For HPM and a constant number of cars, a larger γ yields smaller delays, for LPM it is the other way around.

For all FDMA + SALOHA results, simulation parameters of N sim = 150 and t end = 1000 where used. These parameters where chosen from many tests to minimize computation time whilst preserving statistical accuracy.

B. How t _b20,LP affects performance for the extended SALOHA protocol

The initial window size affects both the SALOHA protocol delay and throughput performance. In figure 18 and 19 it can be seen that, for a constant number of cars, the throughput increases and the HPM delay decreases as t b20,LP increases.

Fig. 18. Throughput as a function of LPM initial backoff window size t

and number of cars N

. t

is here set to 1.

Fig. 19. HPM delay as a function of LPM initial backoff window size t

and number of cars N

. The area to the top left has infinite delay and is highly

unstable.

D1: AUTONOMOUS CARS NETWORKING

It can also be of interest to look at how the LPM delay is affected by increasing t _b20,LP , this is illustrated in figure 20.

To analyze which combinations of N c and t b20,LP that match our requirements we must also look at how many packets succeed within the required end-to-end latency as specified in section II-A. This is shown in figure 21. However, many of the points illustrated in these figures do not belong to stable configurations which can be seen in figure 22. Despite shorter simulation runtime, we chose to keep a large portion of the specified area to analyse as this gives a clearer picture what is going on. This is especially true for figure 21 where it is very clear that more cars and a smaller value of t b20,LP lead to fewer HP packets succeeding within the required delay.

Fig. 20. Mean normalized delay for LPM expressed in amount of packets.

Note that for larger values of N

the delay is displayed as decreasing but this is only because of the fact that very few packets are successful. The area to the top left has infinite delay and is highly unstable.

Fig. 21. Percentage of high priority packets that succeed within the end-to-end latency requirement.

Fig. 22. Stable (bright) and unstable (dark) data points for SALOHA.

C. Finding t b20,LP and γ that match the HPM requirements The optimal sub-band split ratio, γ, for our FDMA+SALOHA protocol is determined as the peak of the bright area in figure 23. It is approximately γ = 0.7 and yields a capacity of roughly N c = 20 cars.

Fig. 23. Map of HPM data points that are stable and fulfill the delay and reliability requirements (bright) and points that do not (dark). This graph is produced by setting the requirement that no packets can fail within the end-to- end latency and that the LPM channel must be stable. One roughly achieves this graph by setting the maximum delay to 225 packets (10 ms for R

=54 Mbit/s) from figure 15 and then applying the LPM stability requirement from figure 16.

The optimal initial window size for LPM, t b20,LP , cannot

be determined as easily. To find the amount of allowed cars

we must take which data points from figure 21 that match

our requirement of 100% succeeding packets which gives us

figure 24.

D1: AUTONOMOUS CARS NETWORKING

Fig. 24. Data points fulfilling the requirements for the extended SALOHA.

This graph is produced by setting the requirement that no packets can fail within the end-to-end latency and that the system must be stable.

The problem now becomes that the only negative effect of increasing t b20,LP is that the delay for LPM is also increased which in turn has no requirement. This means that in theory it is possible to have as many cars as one wishes using the extended SALOHA protocol simply by increasing t b20,LP

but the issue is that the LPM delay will grow very quickly in regards to the amount of cars that the system can handle.

This is because the allowed number of cars follows a log 2

relationship with t b20,LP which can be seen from the fact that G = 1 is linear in figure 25 which has a logarithmic x-axis. As explained earlier in section IV-A1, for high loads it is at G=1 where the system will become unstable. Seeing as figures 22 and 24 are closely linked, it is G = 1 which can be assumed to say where the line goes for how many cars can be handled for different combinations of N _c and t _b20,LP .

Fig. 25. Normalized offered load, G, in relation to t

, for SALOHA.

This is for the special case G=1 after which point we can assume instability because of the high total input rate of the SALOHA channel.

For large values of t b20,LP in relation to the system runtime, t _end , G is not accurate. This is because once a LPP has collided in a system with parameters t b20,LP = 1024 and t end = 1000 it will never be sent within the simulation runtime again. To analyze this further, a simulation of much greater t end was done which shows the true nature of G = 1 which can be seen in figure 26. G = 1 closely matches the stability of the system which in turn closely follows the amount of allowed cars that the system can handle.

Fig. 26. G = 1, in relation to t

and for t

= 8000.

Figures 25 and 26 are contour plots of how the normalized offered load (which can be seen as the input from both new and retransmitted packets) changes in regards to our parameters.

G = 1 is especially interesting to follow as it shows when the system most likely becomes unstable as previously discussed.

Looking at figure 26 we clearly see that for small t b20,LP

in relation to t end follows a predictable nature and as such analytical equations can be derived. Even though different values of t end give different results in regards to G and there- fore different number of allowed cars for extended SALOHA, relevant connections between variables can still be made as all simulation parameters are kept constant when comparing.

VI. A NALYSIS

Comparing the results for our extended SALOHA and FDMA+SALOHA protocol from section V-C reveals that it is generally better to divide the frequency band into two sub- bands if we want to prioritize between two types of messages.

The extended FDMA+SALOHA protocol with the optimal split ratio can manage roughly 20 cars whilst the extended SALOHA protocol only manages a handful of cars as seen in figure 23 and 24 respectively. However, we can further increase the amount of cars the extended SALOHA protocol can handle be continuously increasing t b20,LP but this is very slow. But how slow? From figure 26 we can approximate the relationship between the amount of allowed cars and t b20,LP to follow the equation:

N _c {G = 1} = 0.18 log ₂ (t _b20,LP ) + 2.9 (16)

D1: AUTONOMOUS CARS NETWORKING

This means that we need a t b20,LP of 2 to allow for 3 cars, 70 to allow for 4 cars, 3300 to allow for 5 cars, 155 000 to allow for 6 cars, 7 300 000 to allow for 7 cars and so on. This would mean a delay of 6 minutes or so for the case of 7 cars which is generally not considered good performance. If you however consider the our specific use case, watching Netflix at 1080p quality, delay does not actually have that much of an impact as once the entertainment starts rolling all packets are delayed by the same amount simply meaning that everything takes a long time to change. Even in this case though, users would most likely get frustrated if they had to wait 6 minutes to change movie and as such a feasible limit can be drawn to a LPM delay of half a second giving a total amount of allowed cars of 5.

This can be compared to the case of FDMA+SALOHA which at its optimum of 20 cars has a LPM delay of around 4.6 milliseconds. It is worth noting that because of the reasons outlined in section IV-A4 these values or not to be interpreted directly but rather in relation to each other.

From a principle standpoint, we could say that extended FDMA+SALOHA decreases bandwidth and therefore maxi- mum bitrate whereas extended SALOHA increases delay of LPM. For the case of watching Netflix, this could mean lowering quality for FDMA+SALOHA whilst waiting longer between changing movies for SALOHA. However, using the numbers outlined above, it is clear that FDMA+SALOHA is better in essentially every way.

The reason for this is that the requirements for HPM are our limiting factor meaning we can push their input to the limit and be able to disregard what happens with the LPM.

When the two are mixed up, despite the added amount of bandwidth, the HPM are not protected in any way from the huge amount of LPM and so are lost in the crowd, greatly decreasing performance.

VII. D ISCUSSION

Our simulation study suggests that the solution with split frequency band performs better for a larger number of users when it comes to the requirements for HPM. The results presented does compare the performance of our two protocols but not much more. In reality however, the number of cars these protocols will have to manage is generally greater than 20. Networks could be split up into many sections of 20 users each but it is more desirable to allow for more users using the same protocol.

A. Comparing our protocols to the forefront in industry - WAVE

As mentioned in section II-B the access technology for AVs issued by ETSI is ITS-G5. This technology utilizes the multiple access protocol CSMA with collision avoidance (CSMA/CA) [9]. A performance comparison between this protocol and slotted ALOHA is made in [7]. There, we can see that CSMA/CA has a larger throughput than SALOHA, but also a larger delay for larger throughput. This indicates that its reliability is not as not as good as SALOHAs when the traffic

increases. This is also strengthened by [9] who writes that CSMA/CA does not perform well in congested situations. We believe it is safe to say that our FDMA+SALOHA protocol has the potential to outperform CSMA/CA regarding the reliability of HPM in congested situations.

B. Implications of our findings for society

Our findings suggest that basing vehicle communication technology on the physical layer solution IEEE 802.11a with MAC protocol FDMA+SALOHA can manage approximately 20 vehicles. In reality, this number of vehicles is fairly small.

In dense city environments, the numbers are more often than not larger than this which of course implies some challenges.

In situations like this, we can no longer ensure 100% reliability or small enough delays. This increase the risk of accidents and precautions that decrease the efficiency enabled by this vehicular communication has to be taken.

However, there are many further solutions to explore, eg.

that not all cars has to share its data, drop HPM which transmission time is longer than the delay requirement or infrastructure-based solutions.

C. 5G - Performance potential of next generations network system

Even if we can apply further solutions to improve the performance of our protocols with the IEEE 802.11 technology there will still be problems regarding congestion.

This is due to that the available spectrum is limited for IEEE 802.11 technology. There are further problems with this technology. As mentioned, the wireless system for AVs has to support a diverse set of applications. For the IEEE 802.11 standard to cope with this it requires a wide spread of infrastructure-based solutions called roadside units (RSUs), this may not always be practical. The next generation wireless system, 5G, aims to solve these problems by both augmenting existing solutions and introducing new ones. For example, today’s widely accepted system 4G LTE supports integration of WiFi and the unlicensed spectrum but 5G will extend this to also support the integration of 3G, 4G, Bluetooth and others. This augmentation will allow AVs to connect to the most suitable network for different types of messages. [10]

Furthermore, 5G can be seen as an implementation of a set of building blocks. The most relevant ones for vehicular communication and AVs are listed in [10] and are here shortly explained:

1) Proximity service: Proximity service (ProSe) provides awareness for AVs by discovering devices and services in its proximity. It enables AVs to receive ad hoc location information and to communicate without an infrastructure. By avoiding the transmission from passing through the core of the network this provides high data rate transmissions and efficient utilization of resources.

2) Mobile edge computing: Mobile edge computing (MEC)

moves some of the core functionalities closer to the user (the

edge). This enables the extremely low latencies required for

AVs.

D1: AUTONOMOUS CARS NETWORKING

3) Network slicing: Network slicing eases management of all available access networks that 5G embraces by logically separating them. For AVs, different network slices can be based on the type of message. HPM can be specified as a slice that requires high reliability and very low latency while LPM can be specified as a slice that enhances infotainment user experience by ensuring high-bandwidth content distribution.

Together, these building blocks solve many of the short- comings of today’s standards for vehicular communication.

Together they can ensure safety by fulfilling the requirements for HPM in a resource efficient way. ProSe also increases the security for AVs by being able to determine the source of hack- ing attacks. MEC enables intensive computing applications to deliver results on time by moving cloud services closer to the edge of the network. Network slicing can designate slices based on the diversity of requirements for different messages.

Note that network slicing is similar to our FDMA+SALOHA solution. The difference is that our solution slices the frequency band to meet the requirements of HPM while network slicing designates a slice of the available networks to do the same.

Even without these special building blocks, 5G will simply be able to increase the maximum bitrate by a factor of more than 100 comparing with todays standards [11]. As shown with our analysis of FDMA+SALOHA, the relationship between allowed number of cars and bitrate is linear. Therefore, using our only our simple protocol outlined in this paper, the amount of allowed cars could be closer to 2000 if we used the physical layer properties of 5G. This should be an acceptable number to split the network communications cell in.

To summarize this section we can conclude that 5G will implement many different improvements to existing solutions.

This enables the network systems to utilize their resources in a highly efficient way and thus fulfilling all requirements for AV communication even during highly congested situations.

D. Critique of our method

It is safe to say that our analysis has provided a result that one of our two proposed protocols performs better in regards to AV communication requirements. It is a very narrow analysis as it only investigates the MAC in the data link layer and does not dig deeper into the physical layer. The assumptions made for AV communication in section II-A limits this analysis since there are many more scenarios and requirements for AVs.

However, this analysis is sufficient to compare the protocols since we base the analysis of each protocol on the same assumptions and use the same simulation software with the same parameters.

1) The issues that come with simulation: The main benefit of simulation is that it more closely resembles that of a real system and allows a brute-force method for finding answers.

This is because very few assumptions have to be made.

Instead, the drawback factor comes from the fact that only a limited runtime can be simulated. This has a number of effects:

• As the system generates randomly every time, many iterations have to be run to achieve a reliable result.

This leads to very high computation times (many of the graphs in this report took 10+ hours to compute) without necessarily giving extremely accurate results. An example of this is the allowed number of cars ”edge” in figures 23 and 24.

• For large values of initial backoff time in relation to simulation runtime the results become inaccurate. This is seen when comparing figures 25 and 26.

• The system starts with 0 packets every time. Therefore, we are actually simulating the start of a communication process N s im times and not a static steady state system.

For large values of t end the effects of this becomes smaller but they are never the less always there.

• It is difficult to change how the system works in com- parison to using analytical models. Models can be easily changed and give a deeper understanding whilst simu- lations require a lot of work to implement and the user only sees input and output without understanding what is actually happening. This is however necessary when the systems become too complex.

• The values for packet delay become very dependent for large values of λ. This is shown in section IV-A4. Whilst this is not a very big issue for comparing two protocols using the simulation with the same parameters, it does make it difficult to compare to other protocols. In combi- nation with the fact that new protocols can be difficult to implement, this is can be a problem. Furthermore, because of the fact that the amount of allowed cars is affected by the delay values, this can be an issue when determining the total amount of allowed cars for our protocols.

• Considering the fact that the main reasoning behind using simulations instead of models was that models where not accurate for large λ, the simulation may be not as useful as one first thinks. This is because, as outlined in the previous point, for large λ the results for delay are not easily studied. However, by studying how parameters are affected by t end it is possible to outline analytical equations that describe these relationships. A brief analysis of this was done and a graph of throughput depending on t end can be seen in figure 27.

2) Defining stability: Stability in this report is defined using analytical models outlined in [7]. Since we can only see a snapshot of a simulation, we can only use approximative values of G and S which is what defines stability. For the simulations used in this report, a data point is considered unstable if a single one of the N sim number of simulations was unstable as this indicated that the system should be unstable if allowed to run for a long time.

3) The premise for real SALOHA: As outlined in [7], real SALOHA uses acknowledgement messages (ACKs) to communicate to the sending user that their message has been received correctly. If the user does not receive such a message then that package is considered not to have been successful.

That we have ignored this is not a major issue as this entire

backoff procedure can take place within a single SALOHA

slot. However, this does strain the assumption that the packet

size L can be considered as 300 bytes.

D1: AUTONOMOUS CARS NETWORKING

Fig. 27. How different values of t

affect system throughput. We see here a negative log

relationship, indicating that the system throughput becomes smaller with larger t

but that this effect diminishes quickly.

4) Other assumptions: The assumption of our inputs, that is 100 packets per second for HPM and 5 Mbit/s for LPM is a very specific case. We have also assumed that neither of these can be changed. For instance, it seems more likely that a better approach would be to simply lower the quality of the Netflix stream to allow more cars rather than have a constant quality at all times. This can be generalized to the fact that our protocols do not in any way listen to the environment that they are communicating in which in all likelihood greatly decreases their performance.

We also assume the absolute best conditions for our physical layer properties that is, a maximum bitrate of 54 Mbit/s whilst this will be lower in reality. Other factors such as the speed of the cars using the network and the distance between the cars also play a major role. whilst we have essentially studied the case for still cars that are right next to each other.

E. Further analysis

1) Creating more stable protocols: As outlined in [7]

there are many ways to make protocols more stable. The main way is to make the system listen to how congested the current channel is and adjust backoff windows and input rates according to this. The introduction of a controller, that is a unit which coordinates traffic can further increase stability.

We have here adhered to techniques which are relatively simple to implement. This is an active choice but performance should be able to be greatly increased if other methods are approached.

2) Changing t b20 for HPP: In this report, we have only analyzed what happens for increasing initial backoff windows for LPM. The graphs outlined in section V however, show that the system should perform better if HPM have a slightly larger value of t b20 .

3) Changing t _b1 : For the protocols outlined in this paper, the start of the backoff window does not change and is always set to the packet slot that is right after packet collision.

Changing this could lead to a difference in performance.

4) Changing the backoff window in other ways: Our pro- tocols use a flat probability distribution for when the packets should arrive within the packet window. Other probabilities such as normal or poission distributions within the window itself could also make a difference in performance. One could also imagine the window being divided and spread out in such a way as to allow certain time-sections to be reserved for high priority messages whilst the remaining time is available for low priority messages.

We have here chosen to only study how quadratically increasing window sizes affect system performance. The window size could however be modified to change in a different manner, say in such a way that t b2 increases exponentially.

5) Splitting networks into small cells: As outlined in this report, a communication protocol can handle a handful of cars. Even advanced protocols of today can only handle a few hundred cars [2]. This means that the cars regardless of protocol need to be split up into multiple ”communication cells” that each can handle the specified amount of cars. This could then be applied to a general traffic simulation to further analyze how systems would work in the real world.

6) Further work in a broader sense: To fully analyze what MAC protocols that are optimal for AV communication many different versions and combinations of the entire wireless system has to be analyzed and compared. As seen in the discussion about the potential and performance of 5G, a lot of different solutions are proposed to be utilized for different types of use-case requirements in different situations. All of these solutions has to be carefully analyzed and perfectly optimized.

There are many more MAC protocols to analyze, FDMA, slotted ALOHA and their combinations are only some of many. How can CSMA be modified to prioritize between dif- ferent types of messages? Might HPM interrupt a transmission of LPM and how would that affect the performance of a CSMA protocol? These types of questions can be asked for all MAC protocol solutions.

Further analysis that could be made in the broader case of AV communication could be research on how new innovative solutions affect the performance of the communication. How could unmanned aerial vehicles boost communication perfor- mance in AV scenarios? Or what if AVs use another physical communication medium?

VIII. S UMMARY AND C ONCLUSION

In this report, we aimed to investigate how MAC protocols

can be extended to allow for message prioritization in the

context of autonomous vehicles. Two different modifications

to existing protocols have been proposed, analyzed and

D1: AUTONOMOUS CARS NETWORKING

compared in regards to industry requirements. This was done by building a simulation software in MATLAB and simulating the scenario of two types of messages, low priority and high priority. Our proposed protocol FDMA+SALOHA performed better in terms of reliability and delay for a larger number of cars than what our modified SALOHA protocol performed.

Our results show that FDMA+SALOHA can manage 20 cars while fulfilling the requirements, mean while our modified SALOHA only can manage around 5 cars under the same requirements. We conclude that these protocols would not suffice in reality since the wireless systems has to be able to manage more than 20 cars at a time in some situations.

Moreover, we also concluded that simulations produced more accurate results than that of models and that a SALOHA system with quadratically increasing backoff windows is more stable than one with exponentially distributed backoff times.

Lastly, we saw that that our proposed protocol with split frequency band probably outperforms the CSMA/CA protocol regarding relibability for HPM in congested situations.

A CKNOWLEDGMENTS

The authors would like to thank supervisor Viktoria Fodor, course coordinator Anita Kullen as well as Joakim Lilliesk¨old, Hans Sohlstr¨om and Gabriella Hernqvist. We would also like to thank the vehicle engineering program director, Susann Boij, for making it possible to study this bachelor thesis course after requests from the student body.

R EFERENCES

[1] K. Zheng, Q. Zheng, H. Yang, L. Zhao, L. Hou, and P. Chatzimisios,

“Reliable and efficient autonomous driving: the need for heterogeneous vehicular networks,” IEEE Communications Magazine, vol. 53, no. 12, pp. 72–79, Dec 2015.

[2] L. Hobert, A. Festag, I. Llatser, L. Altomare, F. Visintainer, and A. Ko- vacs, “Enhancements of v2x communication in support of cooperative autonomous driving,” IEEE Communications Magazine, vol. 53, no. 12, pp. 64–70, Dec 2015.

[3] B. W. Khoueiry and M. R. Soleymani, “An efficient noma v2x communi- cation scheme in the internet of vehicles,” in 2017 IEEE 85th Vehicular Technology Conference (VTC Spring), June 2017, pp. 1–7.

[4] 3rd Generation Partnership Project, “Technical Specification Group Services and System Aspects;Enhancement of 3GPP support for V2X scenarios; Stage 1 (Release 15),” 2017.

[5] E. U. Commission, “5G Automotive Vision,” Oct. 2015.

[6] R. Rom and M. Sidi, in Multiple Access Protocols Performance and Analysis, ser. Telecommunication Networks and Computer Systems.

Haifa, Israel: Springer Verlag, 1990.

[7] N. Benvenuto and M. Zorzi, in Principles of Communications Networks and Systems, 3rd ed. New York: Wiley Telecom, 2011, ch. 3, pp.

816–818.

[8] How-To Geek. (2018, Jan.) How much data does netflix use? Accessed:

April 2018. [Online]. Available: https://www.howtogeek.com/338983/

how-much-data-does-netflix-use/

[9] E. G. Strom, “On medium access and physical layer standards for cooperative intelligent transport systems in europe,” Proceedings of the IEEE, vol. 99, no. 7, pp. 1183–1188, July 2011.

[10] S. A. A. Shah, E. Ahmed, M. Imran, and S. Zeadally, “5g for vehicular communications,” IEEE Communications Magazine, vol. 56, no. 1, pp.

111–117, Jan 2018.

[11] InternationalTelecommunicationUnion, “Minimum requirements related

to technical performance for IMT-2020 radio interface(s),” 2017.