Performance Analysis of Sidelink 5G-V2X Mode 2 through an Open-Source Simulator

The Third Generation Partnership Project (3GPP) has recently published a new set of specifications to enable advanced driving applications in fifth generation (5G) vehicle-to-everything (V2X) scenarios, with particular effort dedicated to the sidelink resource allocation in the autonomous mode, named Mode 2. In this paper, we conduct a comprehensive analysis of Mode 2 performance via an open-source system-level simulator, which implements the 5G New Radio (NR) flexible numerology and physical layer aspects together with the newly specified sidelink resource allocation modes for V2X communications and different data traffic patterns. Results collected through extensive simulation campaigns, under a wide variety of vehicle density, data transmission settings and traffic patterns, showcase the effects of the new 5G-V2X features on the sidelink resource allocation performance and provide some insights into possible ways to further improve Mode 2 performance.


I. INTRODUCTION
Vehicle-to-everything (V2X) communication is a key paradigm for upcoming cooperative automated driving as it enables any vehicle to communicate with other vehicles and with any other V2X-enabled entity in the vicinity for sharing their local views and intentions, discover surroundings, and coordinate driving maneuvers.
The Third Generation Partnership Project (3GPP), starting from the Releases 14 and 15 of long term evolution (LTE), has included the support of V2X within the LTE-V2X standard that proved to sustain basic V2X safety applications. More recently, the enhancements to the fifth generation (5G) system and its new radio (NR) interface have been finalized in Release 16 to accommodate advanced V2X use cases within the 5G-V2X standard, sometimes called NR-V2X. The 3GPP is currently working on further improvements of 5G-V2X in Releases 17 and 18. Cellular V2X enables uplink and downlink communication between terminal nodes, namely user equipments (UEs), and base stations, namely gNBs, in the radio access network.
Sidelink communication is also enabled and refers to the direct communication between UEs, without conveying the data through the network. Leveraging sidelink communications, vehicles, road side units (RSUs), or handheld devices carried by pedestrians, which are all considered UEs, can communicate directly with each other.
One of the key aspects in the 5G-V2X sidelink is the resource allocation, which may be either decided by the network in a controlled way or directly by the single UEs through an autonomous selection procedure. In the controlled mode, referred to as Mode 1 in 5G-V2X and Mode 3 in LTE-V2X, the base station (BS) schedules the sidelink resources attempting an ideally interference-free allocation. Contrarily, in the autonomous mode, referred to as Mode 2 in 5G-V2X and Mode 4 in LTE-V2X, the UEs select sidelink resources on their own by using a channel sensing mechanism which is not immune to packet collisions.
The autonomous mode is unquestionably the more challenging of the two modes as the channel sensing and distributed resource selection mechanisms play a critical role VOLUME 4, 2021 for an efficient and effective sharing of the sidelink resources among vehicles. 5G-V2X compared to LTE-V2X adds further complexity in the autonomous mode configuration and operation due its higher flexibility in both the physical (PHY) and medium access control (MAC) layer operation. First, Mode 2 relies on a PHY layer numerology which is scalable and not fixed as for Mode 4 and determines the amount of sidelink resources to be shared: the bandwidth of the resource block depends, in fact, on the subcarrier spacing, which is given by the numerology, and therefore PHY layer settings and the adopted numerology have a joint impact on the amount of available sidelink resources. Second, a more flexible MAC layer supports a larger variety of data traffic generation patterns. Indeed, unlike LTE-V2X that only supported periodic messages in its initial design, in 5G-V2X also aperiodic traffic is considered from the beginning of its specifications.
As the autonomous 5G-V2X sidelink is no longer an idea under discussion, but rather a well-defined standard [1]- [3], a thorough performance analysis that takes into account all of the above aspects is necessary to understand the effects of the main PHY and MAC layer parameters and procedures, their interplay on Mode 2 sidelink communications under a number of different scenarios and settings.
The key contributions of this paper are as follows: • We introduce an open-source event-driven simulator, named WiLabV2Xsim, for the performance analysis of 5G-V2X. It extends LTEV2Vsim, an LTE-V2X simulator, by taking into account the flexible NR numerology and the PHY and MAC layer settings and procedures, as foreseen in 3GPP specifications [1]- [4]; 1 • We present a thorough performance analysis of 5G-V2X Mode 2 showing the impact of the new features of 5G-V2X, among which the flexible subcarrier spacing (SCS) of NR, the modulation and coding scheme (MCS) settings, the resource allocation and retransmission schemes, on the main performance metrics for V2X sidelink data exchange; • We provide helpful guidelines for an effective parameter setting that improves the performance of Mode 2 under different vehicle densities and data traffic patterns. The remainder of the paper is organized as follows. Section II presents the related work and details the gap that this paper is filling. Section III presents the key parameters related to the numerology and the physical layer of 5G-V2X Mode 2. Section IV introduces the Mode 2 resource allocation procedure and the main parameters taken into account within the proposed simulation framework. Section V presents the main building blocks of the implemented open-source simulator. Section VI provides simulation results with insights in the performance evaluation and hints for improvement. Finally, Section VII presents our concluding remarks. 1 The open-source simulator WiLabV2Xsim will be freely available at https://github.com/V2Xgithub/WiLabV2Xsim. It extends LTEV2Vsim, which is an open-source simulator currently available at https://github.com/alessandrobazzi/LTEV2Vsim.

II. RELATED WORK AND PAPER SCOPE
Recently, surveys and tutorials providing a comprehensive overview of the 5G-V2X technology features have begun to follow one another, with focus on the architecture as well as the physical layer and how resources are allocated [5]- [8]. Other types of papers have started to quantitatively investigate the performance of some aspects of 5G-V2X. The work in [9] focused on the PHY layer performance of 5G-V2X, without exploring the MAC dynamics. The impact of the flexible 5G numerology on the MAC of the autonomous mode was preliminarily investigated in [10], but when considering the Mode 4 resource allocation mechanism and periodic traffic only, being this work published before the specifications of Mode 2 in Release 16. Aperiodic traffic has been recently considered in [11], where the misalignment between packet generation and resource allocation was investigated in case of Mode 4. The work in [12] extended the Mode 2 reservation scheme to specifically deal with aperiodic traffic, by inferring the probability of the packet generation intervals. In [13]- [15] the first studies on the performance of Mode 2 also appeared. In [13], the authors compare the use of Mode 2 with random allocation, providing an extensive description of the sidelink technology in 3GPP, and focusing on the effects of some key parameters used for the sensing-based resource selection. In [14], the authors mainly focus on the impact of the numerology to compare 5G-V2X Mode 2 and LTE-V2X Mode 4. In [15], the authors compare the behavior of Mode 2 in the presence of periodic and aperiodic traffic. All these studies do not elaborate on the relationship between packet size, MCS, and available resources nor analyze the key modifications introduced by the recently developed 5G-V2X Mode 2 compared to LTE-V2X Mode 4.
In this work, instead, with the aid of one of the first open-source system-level simulators on sidelink 5G-V2X, we disclose the effect of the peculiar features introduced by Mode 2 at both the PHY and MAC layers. For LTE-V2X, huge efforts have been devoted to the development of simulation platforms for the investigation of sidelink performance, with special attention to Mode 4. In most cases, the implementations are not shared publicly, hindering the reproducibility of results. However, a few solutions have also been made available as open-source, including as examples those presented in [16]- [18]. The first one [16] focuses on Mode 4 and is an extension of ns-3, based on the device-todevice (D2D) model presented in [19]. Another open-source simulator, called ms-van3t, extending ns-3 to include Mode 4 is presented in [17]. The solution presented in [18], also focusing on Mode 4, is a modification and extension of the SimuLTE [20] within OMNeT++; it is implemented in two versions, one integrating with the Artery 2 framework and the second one integrating with Veins [21] only. Very recently, a few implementations of some of the 5G-V2X features have been added to ns-3 and presented in [13]- [15].
Among the open-source simulation platforms, the first appeared for the study of LTE-V2X Modes 3 and 4 was LTEV2Vsim [22]. It is written in Matlab and since its publication in 2017, it has been extensively leveraged by several research teams to evaluate the performance of Mode 4 and its extensions, e.g., in [23], and more recently in [24]- [30], just to name a few. Moreover, besides LTE-V2X, the simulator provides support for evaluating IEEE 802.11p, and it was exploited for the investigation of their co-channel coexistence in [31]- [33]. We have enhanced such platform to include 5G-V2X PHY and MAC layers, with special focus on the autonomous mode, for the performance analysis presented in this work, which complements the current literature analyzing the 5G-V2X sidelink performance.

III. 5G-V2X PHYSICAL LAYER DESIGN
5G-V2X extends LTE-V2X with the features of the 5G NR physical layer and accommodates advanced V2X applications with a broad range of requirements, thanks to the multicarrier orthogonal frequency-division multiplexing (OFDM) empowered with the versatility of the NR numerology. The design of the PHY layer for the 5G-V2X sidelink, with its flexible numerology, time-frequency resource structure, and channel organization, is briefly summarized in this Section.

A. FLEXIBLE NUMEROLOGY
3GPP specifies two large frequency ranges for NR operation, namely frequency range 1 (FR1) and frequency range 2 (FR2). The former is what is usually referred to as the sub-6 GHz band, while the latter is referred to as the millimeter wave (mmWave) band. Depending on the frequency range, the maximum channel bandwidth and the space between OFDM subcarriers (the SCS) may vary. While in LTE the SCS was fixed as ∆f = 15 kHz, 5G NR introduces the concept of flexible numerology as the value of SCS can vary, as well as the OFDM symbol duration. In particular, SCS = 2 µ * ∆f and different numerologies are referenced by the parameter µ. Table 1 reports the numerologies specified by the 3GPP for the two frequency ranges [4].
Regarding the modulation, Release 16 5G-V2X supports Cyclic Prefix (CP)-OFDM not only for uplink and downlink but also for the sidelink [34]. The available subcarrier modulations are quadrature phase shift keying (QPSK), 16-quadrature amplitude modulation (QAM), 64-QAM, and 256-QAM. Different CP lengths can be associated with different SCSs in order to countermeasure different levels of inter-symbol interference (ISI) caused by the multipath fading. The NR standard gives also the possibility to use normal CP (NCP) and/or extended CP (ECP). The NCP is available for any value of SCS, whereas the ECP is possible only with SCS = 60 kHz, in both frequency ranges.

B. RESOURCE STRUCTURE
In NR, radio resources (RRs) span both time and frequency domains. In the time domain, RR are organized in frames, subframes, and slots. A frame has a 10 ms duration and is composed of 10 subframes of 1 ms duration. For sidelink transmissions, the minimum resource allocation unit in the time domain is a slot [35]. Depending on whether a normal CP or an extended CP is used, each slot consists of 14 or 12 OFDM symbols, respectively.
The slot time duration depends on the SCS and, in particular, its duration decreases as the numerology increases. For the 15 kHz SCS of LTE-V2X, the length of a slot is 1 ms, corresponding to a subframe. For higher numerologies, with 30 kHz, 60 kHz, and 120 kHz SCSs, the slot duration is 0.5 ms, 0.25 ms, and 0.125 ms, respectively. As a result, the number of slots per frame and per subframe changes according to the numerology, as indicated in Table 1. It is worth noting that larger SCSs correspond to shorter time slot duration, and so the use of higher numerologies is preferred under low-latency application requirements.
In the frequency domain, RRs are organized in resource elements (REs), physical resource blocks (PRBs), subchannels, and resource pools. Each RE is a subcarrier over an OFDM symbol, and each PRB is composed of 12 consecutive subcarriers with the same SCS. As the SCS changes, the bandwidth of a PRB varies accordingly. As a result, the number of PRBs within a fixed channel bandwidth depends on the SCS. The last column of Table 1 details the available PRBs for a given bandwidth of 10 MHz [36].
A subchannel consists of a group of M consecutive PRBs and represents the smallest allocation unit in the frequency domain for sidelink communications. M corresponds to the subchannel size, which is configured within a resource pool, and can take the following values, 10, 12, 15, 20, 25, 50, 75, or 100 PRBs [5]. A resource pool is a subset of available RRs that can be used by a group of UEs for their sidelink communication. A resource pool consists of contiguous PRBs and contiguous or non-contiguous slots.
In Fig. 1, the relationship between RE, subcarrier, PRB, and subchannel is provided with reference to the time and frequency domains.  The number of available subchannels varies as the SCS and the PRB bandwidth vary. For example, if subchannels with a 10 PRB-size are assumed, in a given channel bandwith of 10 MHz, 5 subchannels can be accommodated for SCS = 15 kHz, 2 subchannels for SCS = 30 kHz and a single subchannel with SCS = 60 kHz.
Each transmission is performed in one time slot and one or multiple contiguous subchannels, depending on the packet size, the number of PRBs per subchannel, and the MCS. More specifically, given the packet size and MCS, the number of PRBs needed for the packet to be transmitted is determined. Lower MCS indexes correspond to a higher number of PRBs. As a consequence, there is the risk that a very low MCS index could require more PRBs than those available in a given slot for a given channel size and subchannel size. More details will be given in Section VI.

C. SIDELINK PHYSICAL CHANNELS
A sidelink physical channel is defined as a set of REs carrying user data and other control information originating from higher layers. The transmission data is organized in transport blocks (TBs), which contain the data and is associated with a sidelink control information (SCI). In 5G-V2X, each TB is associated to a SCI, which is in turn transmitted in two stages to reduce the complexity of resource sensing. The first-stage SCI is primarily used for channel sensing and is decodable by any UE. The remaining scheduling information is carried by the second-stage SCI. Then, in contrast to LTE-V2X, where only broadcast is supported, the advent of the second stage SCI in NR-V2X allows for a customizable SCI architecture that supports unicast, groupcast, and broadcast transmissions.
The transmission data, the SCI and other information are mapped onto sidelink physical channels. For sidelink communications, NR defines four sidelink physical channels: • Physical Sidelink Shared channel (PSSCH): it conveys sidelink user data, synchronization information, control configuration data, and the 2nd stage of the SCI; • Physical Sidelink Control Channel (PSCCH): it conveys the 1st stage of the SCI, which transports sidelink scheduling information. The information of the PSCCH must be decoded by any UE for channel sensing purposes; • Physical Sidelink Feedback channel (PSFCH): it carries feedback related to the successful or failed reception of a sidelink transmission; • Physical Sidelink Broadcast Channel (PSBCH): it conveys information related to synchronization and is sent periodically within a sidelink synchronization signal block, but not on slots of a resource pool.
Note that the SCI is mapped onto two distinct sidelink physical channels. The 1st stage SCI is transmitted onto the PSCCH while the 2nd stage SCI is transmitted along with its relative TB onto the PSSCH. The PSSCH can span over one or multiple subchannels according to the packet size and the adopted MCS, while the PSCCH is always located in a known position within the subchannel. The TB and its relative SCI are always transmitted in the same time slot.

D. SLOT STRUCTURE
Each data transmission, which covers in 5G-V2X one slot and one or more subchannels, starts with one symbol used for automatic gain control (AGC), which carries a copy of the second symbol. Then, in addition to the physical channels, a number of demodulation reference signals (DMRSs) is transmitted with known data to allow performing channel estimation for the correct demodulation of the physical channels. DMRSs are transmitted along each of the sidelink physical channels at configurable positions inside the slot. Different DMRS positions and time-densities can be used to cope with different scenarios (i.e., depending on the speed as this changes the dynamics of the radio channel). Higherorder numerology may require less DMRS transmissions as they are associated with shorter slot duration. The position of the DMRS inside the slot is referenced in the SCI. An empty guard symbol concludes the slot to allow transmission to reception switch and time adjustments.
The 3GPP specifications permit high variability in the configuration and Fig. 2 shows two examples, where the PSCCH entirely occupies the used subcannel for three symbols and two symbols are dedicated to the DMRSs.

E. RETRANSMISSIONS
Sidelink retransmissions can be used to improve the communication reliability for 5G-V2X use cases with stringent requirements. In Mode 2, this can be done either through blind retransmissions or through HARQ-feedback. In the blind retransmission, a pre-configured number of retransmissions can be performed for each message without any feedback; blind retransmissions allow to increase the probability of correct reception not only because more attempts are performed to decode the message, but also because the signals received can be combined at the receiver and increase the overall signal-to-interference-plus-noise ratio (SINR). In the presence of HARQ-feedback, the transmission over the PSFCH is performed by a receiving UE as response to a transmission over the PSSCH previously received. For broadcast transmissions, only blind retransmissions are supported, while HARQ-feedback is also available for groupcast and unicast communications.

IV. AUTONOMOUS RESOURCE ALLOCATION IN 5G-V2X
In Release 16, the 3GPP re-defines the scheduling of resources with additional features and mechanisms. In this section, the scheduling of resources for the LTE-V2X Mode 4 (i.e., before Release 16) and the 5G-V2X Mode 2 (i.e., from Release 16 on) are presented.

A. LTE-V2X MODE 4
In Mode 4, UEs autonomously select the resources for their transmission using the sensing-based semi-persistent scheduling (SB-SPS) specified in Release 14 under the assumption of periodic traffic. The procedure for the selection of resources consists of two phases: (i) channel sensing and (ii) resource selection. During the channel sensing phase, the UE senses the radio channel and detects ongoing transmissions by decoding the SCIs transmitted by other users. Together with SCI decoding, the UE measures also the reference signal received power (RSRP) and calculates its average value over a sensing window of 1 s preceding the slot at which new resources must be selected. The average is done to include all transmissions that were performed in the past and that correspond to a periodicity that can conflict with the allocation that is being performed.
To identify a suitable resource for the transmission, the UE starts from all the possible allocations excluding those for which it lacks information (for example because it was not able to sense the medium due to half duplexing) and then builds a shorter list of candidate resources from which the selection can occur. This list of resources, namely L2, consists of a percentage M of the resources that have been least interfered and therefore, are less prone to collisions. By default, M is equal to 20%. The UE then randomly selects a resource from the L2 list. The resource corresponding to the same subchannels is then periodically reserved every resource reservation interval (RRI), which is a time interval assumed equal to the packet generation periodicity and constrained to some specific values (i.e., 20 ms, 50 ms, or any multiple of 100 ms up to 1 s). The number of transmitting periods before the resource is reevaluated is equal to the reselection counter, which is a discrete number initialized randomly, e.g., between 5 and 15 if RRI ≥ 100 ms. When the reselection counter reaches zero, the resource is changed with probability 1 − p k (p k is set by the operator between 0 and 0.8); with probability p k , instead, the resource is kept for another random interval. The only alternative to the periodic allocation described above can be done by setting the RRI to 0, which means that there are no subsequent reservations.

B. 5G-V2X MODE 2
In Mode 2, analogously to Mode 4, the UEs autonomously select the sidelink resources for their transmission, without the support from the network. In addition to the SB-SPS, 5G-V2X explicitly introduces the dynamic scheme for the allocation and scheduling of aperiodic traffic. With the dynamic scheme, the selected resource is only used for one transmission, and resources can be reserved for its own retransmission only. Therefore, with the dynamic scheme, at each transmission, new resources must be selected. The resource selection procedure is the same for both the semi-persistent and dynamic scheme and it involves, like in Mode 4, the sensing and resource selection phase. The sensing window in the 5G-V2X Mode 2 case can have a duration of either 1100 ms or 100 ms according to the configuration that is set for each resource pool. During this period of time, the UE senses and decodes the SCIs sent by other UEs on the sidelink channel. The decoded SCIs are stored together with the measurement of the RSRP and this information is used to determine which resources must be excluded when a new resource selection is required. In Mode 2, resources can be excluded from being possible candidate resources if and only if the associated information is not available (e.g., measurements were not performed because the station was transmitting) or they are reserved by a prior SCI with an associated RSRP above a given threshold. After the exclusion process, a resource is selected randomly from the set of the remaining available resources. Analogously to LTE-V2X Mode 4, also in 5G-V2X Mode 2, when the SB-SPS is used, the resource is used for a number of consecutive periods according to the reselection counter, with the same rules defined for Mode 4. The time period between the selected resources for consecutive transmission is determined by the RRI reported in the SCI, which can be selected in 5G-V2X in a larger set of values (i.e., any integer number of milliseconds between 1 and 99 ms or any multiple of 100 ms up to 1 s). Blind retransmissions, if present, are also advertised in the SCI. Each SCI can reserve up to two retransmissions, with full flexibility in frequency and time in a window of 32 slots.  transmissions reserve the same resource after a period equal to the selected RRI. Transmissions can occupy one or more adjacent subchannels. A past transmission of the sensing UE is represented in violet. No sensing information is available during the time at which UEs transmit, due to the half-duplex limitations. Lastly, the remaining available resources after the exclusion process are represented in orange.
It is important to highlight that the described Mode 2 reengineers Mode 4. In particular, 1) The averaging operation of the RSRP measurements over the sensing window is removed (only the value corresponding to the TB directly associated to the SCI is used); 2) The L2 list is not built nor considered. The removal of the averaging operation for the RSRP measurements has two main effects. On the one hand, a positive effect is that without the averaging operation, the more recent measurements are more important than the less recent ones, which might be related to obsolete reservations; this is particularly true in highly dynamic environments. On the other hand, without the averaging operation, there could be possible issues if the last reservation has been lost, which might occur even with exactly periodic traffic. In such a case, the system is in fact not able to detect the busy status of a resource.
The removal of the L2 list instead implies that partly and possibly highly interfered resources are passed to the higher layers for selection, contrarily to Mode 4 that uses the L2 list and therefore selects the resources among the least interfered ones.
The drawbacks of the average and L2 removal appear clear in the case where the SB-SPS is enforced but the generation of the packets does not perfectly correspond to the periodic resource allocation, which we call incoherent traffic generation, as opposed to a perfect correspondence, called coherent traffic generation. Fig. 4 shows an example of coherent and incoherent traffic generation. In Fig. 4, a resource is selected for being used over time with a given RRI and either 1) the packet generation period is coherent with the RRI (top of the figure) or 2) the packet generation is not coherent with the RRI (bottom of the figure). In the incoherent case, which assumes in the example one new packet every two RRIs, due to the absence of the averaging operation of the RSRP and of the L2 list, the resource is considered empty after the time instants b and d, with higher probability of collisions on resources c and e.

V. THE OPEN-SOURCE SIMULATOR IN BRIEF
The open-source discrete event simulator LTEV2Vsim was designed for the investigation of resource allocation techniques for V2V connectivity, with focus mainly on sidelink LTE-V2X, through the support for both Mode 3 and Mode 4. Later improved with the support of IEEE 802.11p, it is here extended and renamed as WiLabV2Xsim to include the new features of Mode 2 of 5G-V2X, including also the NR numerology and the proper PHY layer settings. All relevant parts of the protocol stack are accurately reproduced, with a realistic generation of the messages at the facilities layer, a careful implementation of the MAC level protocol, a detailed definition of the radio resources at the PHY layer based on the various settings (e.g., numerology, channel bandwidth, subchannel size, MCS, packet size, etc.) and the modeling of packet reception as hereafter detailed. Fig. 5 shows a concise block diagram of the simulator in its use for the investigation of 5G-V2X. The first step is the initialization procedure, which sets the starting position of the vehicles, the available physical radio resources (including the available subchannels), and an initial assignment of resources. The position of the vehicles can be taken either from a given theoretical model or from realistic traffic traces. For the former approach, both highway and Manhattan grid scenarios are implemented, in alignment with [37].
Once the simulation is started, the simulator processes one after the other the occurring events: position update, packet generation, and slot-start/slot-end. More specifically, at regular intervals vehicle positions are updated, which implies an update of path-loss and large-scale fading (shadowing) values. Asynchronously from the position update, packets are generated following the statistics of interest. Also asynchronously with respect to the previous events, the slots are finally processed, with actions taken at their beginning and end, where the end of the slot corresponds to the last instant before the time gap used as guard interval (see Section III). Before the slot-starts, new resource allocations are performed if needed. Then, the nodes that reserved a resource and have a packet waiting in the queue, start their transmission; all the other vehicles act as possible receivers. At the slot-end, the correctness of the transmissions is evaluated, with an update of the output metrics and an update of the measurements used for the SB-SPS procedure. The core part of the evaluation consists in detecting whether each transmission succeeded or not. Denoting the transmitting vehicle as i, the receiving vehicle as j, and the slot under consideration as t, the average SINR is obtained as where P ti is the power transmitted by i, h ij,t is the large-scale fading contribution to the link from i to j in slot t, L(d ij ) is the path-loss from i to j as a function of the distance from i to j, P n is the noise power, and I ij,t is the average interference. In (1), the numerator represents the useful received power, whereas the denominator is the sum of the noise power and the interference, assumed Gaussian with zero mean. I ij,t is in turn defined as: where V t is the set of the nodes transmitting in slot t, and η ki is a multiplying coefficient, between 0 and 1, that quantifies how much power is sent by k in the subchannels used by i, related to the transmission power of k. η ki is 1 if k uses exactly the same subchannels as i, and is lower than 1 if its signal does not overlap or it overlaps only partially. The calculation of η ki takes into account the in-band emission (IBE) in alignment with the specifications in [36]. When blind retransmissions are considered, if the first transmission fails, the corresponding average SINR is saved and summed to the average SINR of the second transmission, thus implementing maximal ratio combing (MRC).
Once the average SINR is calculated, either it can be used to assess statistically the outcome of the transmission using link-level curves detailing packet error rate (PER) vs. SINR, or it can be compared with a SINR threshold (calculated as detailed in Appendix A) to assess whether the transmission is successful or not. The small-scale fading is implicitly included in the curves or threshold. Whereas the former approach is slightly more accurate, the latter allows to simulate any combination of packet sizes and MCS indices without the need of storing a large number of link-level curves and to avoid resorting to interpolation subroutines.
Based on the outcome of each transmission, the simulator provides output performance metrics, including, among others, packet reception ratio (PRR), end-to-end delay (EED), and packet inter-reception (PIR).

VI. PERFORMANCE EVALUATION
In this section, simulation settings and metrics are first described and then the results achieved through the implemented open-source simulator are analyzed focusing on Mode 2. First, the effect of the numerology, MCS and retransmissions is investigated, followed by an insight into the impact of two key modifications to the resource allocation process, i.e., the L2 list removal and the new calculation of the received power within the sensing window.

A. SIMULATION SETTINGS AND METRICS
In this subsection, the settings of the simulations are detailed in terms of evaluation scenario, channel modeling, and data traffic model. A summary of the main settings is also provided in Table 2.

1) Evaluation Scenario
We consider a Highway scenario with road configuration parameters aligned to [37]: a 2 km-long straight highway with 3 lanes per direction and wrap-around (i.e., a vehicle exiting on one side of the scenario, enters from the other side in the same lane). Each vehicle moves at a speed that is a Gaussian random variable with 70 km/h average and 7 km/h VOLUME 4, 2021

2) Power Settings and Channel Model
The UEs transmit within the 5.9 GHz band in a 10 MHz channel bandwidth with a constant spectral power density of 13 dBm/MHz. They have an antenna gain equal to 3 dBi at both the transmitter and receiver sides. The noise figure of the receiver is assumed equal to 9 dB. The path-loss model follows the WINNER+, scenario B1, with correlated log-normally distributed shadowing, characterized by a standard deviation of 3 dB and a decorrelation distance of 25 m, as in [38]. Given the highway scenario, line-of-sight (LOS) conditions are assumed.
The correct reception of each packet is detected as discussed in Section V, based on a SINR threshold obtained as detailed in Appendix A. The RSRP threshold is set to -126 dBm when not differently specified.

3) Data Traffic Model and Link Level Settings
Without loss of generality, packets of 350 bytes [39] are hereafter considered when not differently specified, either periodically generated every 100 ms or with a variable generation interval. As an example, but with the aim to adopt a realistic synthetic model, when a variable packet generation interval is adopted we derive such interval from the rules for the generation of cooperative awareness messages (CAMs), like in [11]. Some of the results have been collected also assuming larger packets to model the information-rich nature of advanced services; in particular, we have considered packet size of 1000 bytes with a 100 ms generation interval, resembling the worst case generation pattern of collective perception messages (CPMs) [40] in a highway scenario, according to [41].
At the access layer, subchannels of 10 PRBs are assumed, corresponding to current regulations in Europe for LTE [42]. The number of subchannels in the 10 MHz bandwidth depends on the numerology and corresponds to 5 subchannels with µ = 0 (SCS = 15 kHz), which is used hereafter when not otherwise stated. Unless differently indicated in the text, MCS 4 is used, which corresponds to QPSK modulation, coding rate R c = 0.3 ; under these settings, all 5 subchannels are required to accommodate a 350 bytes-long packet.

4) Performance Metrics
We evaluate the performance in terms of PRR, which is derived as the average ratio between the number of neighbours correctly decoding a message at a given distance and the total number of neighbours at the same distance. It corresponds to PRR type 1 in [37].
Additionally, we measure the range, defined as the maximum distance from the transmitter at which the PRR remains above a certain value, say 0.9 [43].

B. IMPACT OF NUMEROLOGY AND MCS
In the first simulation campaign, the performance of different numerologies (SCS=15, 30, 60 kHz, see Table 1) and MCSs is evaluated. The effect of retransmissions is also investigated, by considering a single blind retransmission.
Before getting into the simulation results, an insight into the resources that follow different MCS and SCS settings is provided. To this aim, the number of orthogonal resources available in one 10 MHz channel and in a period of 10 frames (i.e., 100 ms) is shown for packets of 350 bytes (Fig. 6(a)) or 1000 bytes (Fig. 6(b)), when varying the MCS and SCS. 3 It can be observed from Fig. 6 that SCS = 30 kHz and SCS = 60 kHz can only be used after a certain MCS index. For example, SCS = 30 kHz requires an MCS equal to or larger than 11 with 350 byte packets and 22 with 1000 byte packets; if SCS = 60 kHz is adopted, at least MCS 19 is needed with 350 byte packets, whereas none of the available MCSs allow to carry a 1000 byte packet. With a lower MCS, the available frequency resources are in fact not enough to accommodate a packet in a single slot. As a consequence, if lower MCSs are needed to improve robustness, such as in the case of V2X broadcast communications, then a lower SCS must be adopted or a larger channel bandwidth must be selected. If both the channel bandwidth and the MCS cannot be increased, then higher order SCS might be relegated to the transmission of small packets only. Fig. 7 shows the PRR as a function of the distance from the transmitter. For the sake of a fair comparison with different SCS settings, MCS = 21 is used; it is the lowest MCS where a similar number of orthogonal users can be allocated   assuming 350 byte packets, as derived from Fig. 6(a). With MCS = 21 each packet of 350 bytes occupies 10 PRBs, which can be allocated using 1 subchannel. Note that when SCS = 15 kHz, there are 5 subchannels available in each slot; hence, up to 5 orthogonal transmissions from different UEs can be allocated on the same slot. 4 For SCS = 30 kHz and SCS = 60 kHz the possible orthogonal allocation resources per slot are reduced to 2 and 1, respectively. So the interference between different UEs using different resources but on the same slot, namely the IBE, is significantly reduced for SCS = 30 kHz and completely removed for SCS = 60 kHz. The reduction of the IBE implies a significant performance improvement, as observable in Fig. 7 and in agreement with what noted in [10]. To corroborate this statement, we also provide in Fig. 7 the results by neglecting IBE for the cases with SCS = 15 kHz and SCS = 30 kHz; as it can be 4 In the absence of IBE. observed, when the IBE is not considered the results for different SCS are comparable. It is however to note that, as observed discussing Fig. 6, the advantage of having a higher SCS and thus a lower IBE is constrained by the number of subchannels available for the given bandwidth and SCS, which might be insufficient for a reliable transmission (i.e., adopting a low MCS) of large packets. Fig. 8 shows the range when varying the MCS and the density of vehicles, both with or without blind retransmissions. As the MCS increases, on the one side each packet occupies a decreasing number of subchannels, which means that the number of resources available for selection increases (see Fig. 6) and therefore potentially the average interference decreases; on the other side, the SINR required for correctly decoding the packet increases, possibly reducing the reliability of the communication. Overall, under the considered settings it can be observed that the best configuration for the PRR is the one with the lowest MCS as it shows the highest robustness. Furthermore, the figure confirms that adding a retransmission increases congestion and it is a favorable choice only when a small portion of the resources is used, i.e., only when the vehicle density is relatively low with a small MCS (i.e., 50 v/km), or when a high MCS is used. At higher vehicle densities, when low MCS settings (i.e., 4) are considered, retransmissions cause an increase in congestion which generally leads the performance to get worse.
Without loss of generality, in the following we consider MCS = 4 as it showed the best performance in Fig. 8. Since blind retransmissions are shown to be effective only under a lightly loaded scenario, they are disabled in the rest of the simulations.

C. IMPACT OF THE L2 LIST REMOVAL
As a further study, we evaluate the effects of the L2 list removal for packet size equal to 350 and 1000 bytes. For 1000 byte packets an MCS equal to 11 is considered, which VOLUME 4, 2021 lets a packet occupy the same number of resources as a 350 byte packet sent with MCS equal to 4 (i.e., 5 subchannels). In Fig. 9, the PRR performance of the Legacy Mode 2, i.e., without the L2 list, is compared to the case where the L2 list is adopted. In the latter case, different percentages of least interfered resources used as candidate ones are considered, i.e., M = 20% (corresponding to what used in LTE-V2X Mode 4) or M = 50%. Results are also reported for two values of the RSRP sensing threshold: -110 and -126 dBm, where the former corresponds to the value commonly used in studies dealing with LTE-V2X Mode 4 (e.g., [44], [45]), and the latter is the minimum possible value. As observable, the curves have similar trends in the two subfigures, with a lower range when 1000 bytes are assumed due to the less reliable MCS.
The figure shows that when the threshold is set to −110 dBm, the case with the L2 list and M = 20%, outperforms both the case with M = 50% and the case without the L2 list (intuitively corresponding to M = 100%). The reason is that by increasing M , more resources that are already used by other vehicles in the neighborhood are considered as available, thus causing a higher probability of collisions. Differently, when the threshold is reduced to −126 dBm, all the curves are comparable and the effect of the L2 list is negligible, for any M settings and for both packet size settings. Indeed, in this case it is more likely that the resources indicated as busy by the SCI have also an associated power exceeding the sensing threshold.
To further showcase the impact of the L2 list and of the threshold setting, results in terms of the range metric are reported in Fig. 10. For example focusing on 350 byte packets ( Fig. 10(a)), it can be observed that when the L2 list is removed and the threshold is set to −110 dBm, the range metric is 110 m, compared to the 170 m that can be reached with the adoption of the L2 list. When the threshold is decreased, the effect of the L2 list removal is mitigated. Indeed, with a −126 dBm threshold, the performance with and without the L2 list is comparable.
The results shown in Figs. 9 and 10 reveal that the removal of the L2 list in 5G-V2X Mode 2 makes the RSRP threshold become a critical parameter, which needs to be kept small in order to avoid the (blind) selection of interfering resources that are already in use. This is different from LTE-V2X, where the value of the RSRP threshold has a minor impact on the performance [45]- [47].

D. IMPACT OF AVERAGING IN THE SENSING PROCEDURE
The last simulation campaign is aimed to evaluate the effects of the removal of the averaging step from the sensing procedure. As described in Section IV and through example 2 of Fig. 4, this is particularly relevant in the case of incoherent traffic generation, which is a possible situation when realistic traffic is assumed.
More specifically, instead of a constant and uniform generation of the packets, here the messages are generated following the ETSI rules for CAM messages [48]. This means that CAMs are generated based on the vehicle changes of position or speed. As a consequence, each vehicle will have messages generated periodically, but with a period that changes from vehicle to vehicle. The average generation period is approximately 207 ms [11]. Under these conditions, it was demonstrated in [11] that assuming an RRI of 100 ms and leaving empty some of the reserved resources was a good approach for LTE-V2X Mode 4. Therefore, a similar approach is here evaluated for 5G-V2X Mode 2.
In Fig. 11, the PRR when varying the distance is shown for a number of different configurations. In particular, the following cases are compared assuming the incoherent traffic generation: (i) the legacy Mode 2; (ii) Mode 2, with the L2 list with M = 20% and performing the average of the RSRP over the last 1 s (to resemble the behaviour of Mode 4); (iii) Mode 2, performing the average of the RSRP over the last 1 s; (iv) Mode 2, with the L2 list with M = 20%. As an additional benchmark, it is also shown what happens with the legacy Mode 2 if coherent traffic generation is assumed; in order to have a similar data traffic pattern, 200 ms for both generation period and RRI is assumed in such case.
It can be observed in Fig. 11, that the performance of Mode 2 worsens significantly when incoherent traffic generation is assumed compared to the case in which generation period and RRI are aligned. This is because a UE would wrongly detect the reserved but unused resource over the last (100 ms-long) sensing window as free. Interestingly, it can be noted that a relevant improvement can only be achieved by reintroducing both the average of the RSRP and the L2 list. In that case, and only in that case, the performance in fact approximates the one achieved under coherent traffic generation. Indeed, considering a longer sensing window over which RSRP values are averaged and selecting the resource among those actually detected as less interfered provides a less myopic allocation under variable packet generation. Differently, when even only one of the two mechanisms is removed, then the scheduling is no longer able to detect future resources occupations and a lower PRR is observed, confirming the strict interplay among the two mechanisms as conceived in Mode 4.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.  All in all, these results show that having removed the averaging step and the L2 list, reserving resources for periodic allocations without using them could have a worse impact on performance of 5G-V2X Mode 2 compared to LTE-V2X Mode 4.

VII. CONCLUSION
In this paper, we analyzed the performance of sidelink Mode 2 with special focus on the novel features introduced in 5G-V2X compared to the previous 3GPP releases, i.e., the flexible NR numerology and modifications to the resource selection mechanism encompassing different procedures and settings for the sensing and the identification of the candidate resources. Results have been achieved through a properly overhauled open-source system-level simulator under a wide range of vehicle density, MCS, and data traffic pattern settings.
The study provides a set of helpful guidelines for understanding how the newly added features to the autonomous mode by 5G-V2X affect the reliability performance of packets exchanged over the sidelink. The effects have been measured for each feature both individually and jointly with others, while also disclosing the impact of other crucial settings. VOLUME 4, 2021 With reference to physical layer features, in order to leverage large SCSs, which provide PRR improvements, either a higher channel bandwidth should be available or high MCS settings. However, benefits of a large SCS can be still achieved for smaller packets even with a lower MCS, which has been proven to ensure a higher robustness. Furthermore, blind retransmissions are shown to be especially effective at low densities, otherwise they risk to further increase the load on the channel. This confirms the need for enforcing more sophisticated load-aware retransmission policies.
Interesting insights about the most appropriate setting of the RSRP sensing threshold are also provided to make the best of the Mode 2 resource selection mechanism. This tuning is specifically needed because Mode 2 does not distinguish the level of interference of resources detected as occupied over the recent past.
All the above findings apply to strictly periodic traffic. The study also focused on realistic (variable) packet generation patterns. It was proved that the persistent reservation regardless of the actual packet generation interval, with potential unused resources, which was previously shown to be quite effective in Mode 4, is instead surprisingly ineffective in Mode 2. In such a case, either a new resource selection per each message or an RRI setting which is more in agreement with the traffic generation needs to be performed by the UE.

APPENDIX A
In this Appendix, the calculation of the SINR threshold used to assess the correctness of each transmission is provided, which depends on the adopted MCS m and packet size B (in bytes). Given m and B, we derive: 1) the modulation order, with the corresponding number of bits per symbol b symbol (m) from [35, Table 5.1.3.1-1]; and 2) the transport block size b TBS (m, B), in bits, according to the procedure described in [35,Clause 8.1.3.2]; for the sake of precision, it can be noted that the calculation of b TBS (m, B) is also influenced by the configuration of the physical channels (e.g., the number of the DMRS symbols). With the obtained values, the actual coding rate R c can be calculated as: where n RE is the number of REs dedicated to the transmission of the PSSCH except those used for the second stage of the SCI. Consequently, the number of data bits per second per Hz carried by the given MCS, denoted by b Hz (m, B), can be calculated as: b Hz (m, B) = n sypslot · n scpPRB · b symbol (m) · R c t slot f PRB , where n sypslot is the number of sidelink symbols present in a slot, which is always 14, n scpPRB is the number of subcarriers in the frequency domain per PRB, which is always 12, and t slot and f PRB are the duration of the slot and the bandwidth of a PRB, respectively, which depend on the numerology (although their product is always 180).
By inverting the Shannon's normalized capacity formula for the Gaussian channel with a parametric factor loss Φ, which accounts for non-Gaussian signaling, finite-length coding, imperfect decoding process and other protocol's overheads, the SINR threshold can be obtained as: The factor loss Φ is here set to 0.6 following the recommendations of [49]. As examples, γ * is reported for B = 190 and B = 350 bytes in Table 3, together with N PRB which is the number of PRBs required to transmit a packet of a given size, with a given MCS.