INTRODUCTION

5G was designed to deliver significantly higher data rates than 4G, with a target downlink (DL) peak data rate of 20 Gbps [19]. To achieve this goal, 5G employs a variety of different mechanisms, many of which build on those that have been deployed in 4G networks. First of all, besides the frequency bands in the low-(< 1 GHz) and mid-band (1 GHz -7 GHz) range (frequency range 1 or FR1) that overlaps with the 4G frequency bands, 5G also utilizes high-band frequencies (24 GHz -60 GHz) in the mmWave range (frequency range 2 or FR2). Secondly, instead of a fixed 15 KHz sub-carrier-spacing (SCS) and a maximum channel bandwidth of 20 MHz, 5G introduces a flexible numerology to increase spectral efficiency, allowing 15/30/60 KHz SCS for FR1 bands and up to 100 MHz channel bandwidth, and 120/240 SCS and up to 400 MHz channel bandwidth for most FR2 bands. Thirdly, massive MIMO (multiple inputs, multiple outputs) may be used to increase the data rate by transmitting up to 4 (or in some cases 8) data streams simultaneously. In particular, carrier aggregation (CA), which combines multiple channels -each is referred to as "component carrier" (CC) -within the same band or across different bands ( §2.1), plays a crucial role in boosting 5G data rates to multi-Gbps. We use x CCs to denote Į number of channels being aggregated for simplicity.

While CA has been deployed in 4G networks, more diverse band combinations, together with flexible numerology, generally wider channel bandwidths, as well as MIMO enable 5G networks to achieve significantly higher data rates. To illustrate the comparative throughput performance of 4G vs. 5G networks, in Fig. 1, we plot the representative measurement results in the ideal channel condition. We see that the throughput performance of 5G networks is significantly higher than 4G networks. As of Jan 2024, we have observed up to 4 CCs aggregation in 5G low/mid-band and 8 CCs aggregation in 5G high-band, achieving an unprecedented 1.7 Gbps and 4.1 Gbps peak throughput performance in the wild, respectively. Moreover, with 5G stand-alone (SA) networks no longer relying on a 4G as an "anchor" cell, CA assumes significant importance in ensuring connectivity and throughput by simultaneously leveraging both low-band and mid-band channels. It has been recently reported [7,35] that Ericsson, Nokia, Qualcomm, and mobile operators have successfully aggregated 6 CCs in a 5G SA network, achieving a peak downlink data rate exceeding 3.6 Gbps using only 5G low-and mid-band channels in FR1.

Despite the importance of CA, there have been limited studies of CA in real-world 5G deployments from the research community (see §8). CA adds significant complexity to the analysis of 5G network performance (see §3). For example, there are multiple modes ("peaks") in both 4G and 5G throughput distributions in Fig. 2 and Fig. 24. These can be attributed to the effects of CA, where "peaks" correspond to areas under coverage of multiple 4G/5G bands/channels, thus multiple CCs with different combinations are available for CA. While previous measurement studies have more or less noted the highly variable [32,33] and "multimodal" nature of 5G throughput performance [48], our analysis reveals that carrier aggregation is one of the major contributors to such observed phenomena. If and when CA is activated, what and how many component carriers -and what band combinations -are used depends on various configurations and environmental factors, such as what bands/channels are configured and thus available in a given location, the channel conditions, and so forth. In addition, the capability of a 5G handset or user equipment (UE) also determines whether CA can be used for data transmission and what band combination is used. All of these affect the user perceived 5G throughput performance and application quality of experience (QoE). Using an XR (mixed/extended reality) application as an example (see §3.3), we show that while the significantly boosted 5G throughput enabled by 5G CA makes it possible to support applications with high bandwidth requirements, the application QoE may suffer due to high variability introduced by CA. Hence, in order to fully translate the throughput benefits brought by CA into the improvement of the application's QoE, it is imperative to take CA into account. The goal of our paper is four-fold: First, we carry out a comprehensive measurement study of CA deployments over three major US operators in two large US cities. We map out the prevailing characteristics of the current CA deployments in 4G/5G networks, quantify the impact, and discuss challenges CA poses ( §3). This is made possible via the use of a professional 5G measurement tool, Accuver XCAL [3], which allows a detailed collection of 5G New Radio (NR) PHY layer signals and RAN (radio access network) protocol messages; see §2.2 for our measurement platform setup and methodology. Second, through careful experiments and in-depth data analysis, we dissect the complexity of 4G/5G CA configurations and identify the key factors that affect when and how CA is used ( §4). The third and main goal of our paper is to develop a CAaware deep-learning framework, dubbed Prism5G, for predicting 5G network throughput performance ( §5) with the aim to aid application in QoE optimization. Through extensive evaluations, we demonstrate the benefits of CA-awareness in 5G throughput prediction ( §6). Last but not least, we consider two use cases to illustrate how Prism5G can help applications enhance QoE performance ( §7). Contributions. We summarize our key contributions and major findings as follows:

• We conduct a timely and comprehensive measurement of CA deployment in commercial 5G networks (as well as 4G networks) in the US. To the best of our knowledge, this is the first in-depth study that considers the impact of CA on throughput analysis and prediction. We map out the CA deployments by all three major US carriers in two cities (and surrounding suburban areas and nearby highways), including the 4G/5G channels and combinations.

• Our study shows that diverse channels and channel combinations have been used to form 5G aggregate channels of exceeding 100 MHz, with up to 4 CCs in the mid-band and up to 8 CCs in the high-band (mmWave), resulting in peak throughput of more than 1.7 Gbps and 4.1 Gbps, respectively. While most of current 5G CA deployments are concentrated in urban areas, all three US operators are gradually expanding CA deployment and coverage.

• While 5G CA significantly boosts throughput, the complexity of CA poses new challenges in analyzing and predicting 5G performance. We demonstrate that not only is the activation and deactivation of CCs that induce drastic changes in 5G throughput in a short period of time, but the aggregated channel also exhibits far higher variability than when no CA is used. All of these have crucial QoE implications for (end-to-end) bandwidth-adaptive applications. Using (scaled-up) XR application ViVo [16] as an example, we show that CA worsens overall QoE metrics comparatively.

• The above findings call for the need of a CA-aware 5G throughput predictor that can more effectively aid applications in fast and adaptive decision making. Toward this end, we dissect the key factors that influence CA configurations and affect their performance. We demonstrate the need to capture features associated with individual CCs and the importance of accounting for the complex feature interplay in predicting 5G throughput performance.

• We propose a novel CA-aware deep learning framework, Prism5G, which models individual CCs, conditions them, and fuses them based on the CA state to accurately predict 5G throughput. It utilizes features that can be collected from UE. Evaluations using real-world 5G traces demonstrate the efficacy of Prism5G, with around 14% improvements over the state-of-the-art. It tracks the 5G throughput transitions well when CCs are added or removed.

• To demonstrate the utility of Prism5G in aiding adaptive applications in enhancing QoE performance, we consider two use cases: 1) We show that Prism5G can help ViVo to attain near-optimal QoE metrics. 2) Using MPC [50] as a representative adaptive bit rate (ABR) algorithm for video streaming, we show Prism5G enhances the average bit rates and reduces stall times considerably. It greatly improves the stall time tail performance, reducing the 95% percentile tail performance by 33 seconds (a ≈ 37% reduction). Prism5G consistently outperforms the other 5G throughput predictors. • We make measurement datasets, main codes, and other relevant artifacts publicly availablefoot_0 .

BACKGROUND AND METHODOLOGY

This section introduces carrier aggregation (CA), followed by a description of our measurement setup and methodology.

A Quick Primer on CA

3GPP specifies a set of frequency bands within each of the lowband, mid-band, and high-band ranges that can be used to support 5G New Radio (NR). Each 5G band is designated with a number, prefixed with the letter "n" [45], e.g., n41 and n77 (C-band), both within the mid-band range. Similarly, 4G bands within the lowand mid-band ranges are also designated with a number [46], and we prefix them with the letter "b" to distinguish them from 5G bands. For each band, 3GPP also specifies what channel bandwidths can be supported, e.g., 15,20,40,60,70,80,90, or 100 MHz, and what subcarrier spacing (SCS) may be used, e.g., 15, 30, 60 or 120 kHz. We note that 4G bands use a fixed SCS of 15 kHz, and the maximum channel band is 20 MHz, compared to 100 MHz for 5G mid-band channels. For both 5G NR and 4G LTE, each band is specified to operate using either the TDD (Time Division Duplex) or FDD (Frequency Division Duplex) mode. In the TDD mode, both downlink (DL) and uplink (UL) data are transmitted using the same channel but in different time slots. Whereas in the FDD mode, a pair of channels (with the same frequency range) are dedicated to data transmissions, one for DL and one for UL.

At each 4G/5G base station, one or multiple channels (from the same or different bands) may be configured, depending on various factors, e.g., availability of spectrum, RAN capabilities, coverage, and bandwidth requirements of the service area. Each channel is often assigned a Physical Cell ID (PCI). The left panel in Fig. 3 depicts a base station with channels/bands/cells configured. We note that as radio bands have varying coverage ranges and different channel propagation characteristics, depending on the UE location, there may be one, two, or multiple channels/bands available. In a location where the UE is under the coverage of multiple cells, the cellular operator may opt to invoke CA by aggregating two or more channels to boost the throughput of data rates for the UE. Each channel is denoted as a component carrier and configured as a serving cell for CA. In the context of CA, the term (serving) cell is used interchangeably with CC. As illustrated in the right panel of Fig. 3, CA may be performed using contiguous or non-contiguous channels with the same bandthese are referred to as intra-band contiguous CA and intra-band non-contiguous CA, respectively. CA may also be performed using channels from different bands; this is called inter-band CA. In the latter case, the component carriers could be operating using TDD or FDD. In the case of 5G CA, each CC can also use different SCSs, e.g., 15 kHz or 30 kHz. For example, in the 5CC CA trial cited earlier [35], 3 TDD CCs and 1 FDD CC in the mid-band are used with the other 1 FDD CC in the low-band. In CA, one of the CCs is designated as the primary cell (PCell) over which radio resource control (RRC) messages are also transmitted, while all the other CCs are classified as secondary cells (SCells). SCells can be dynamically added or removed, depending on network conditions and other factors. CA is particularly important for 5G SA deployment as it is no longer anchored to 4G LTE, where low-band CCs can be aggregated to expand coverage. 3GPP specifies various band combinations that may be supported for NSA and SA CA operations. Lastly, we remark that CA is performed at the MAC layer where user data is multiplexed/demultiplexed across multiple CCs, as shown in Fig. 3. In a sense, NSA Dual Connectivity (DC) can also be viewed as a form of "CA" where traffic is split/merged between 4G LTE channels and 5G NR channels at the higher PDCP (Protocol Data Convergence Protocol) layer. We refer the reader to [2, 10] for more background and exposition on CA.

Measurement Methodology

To understand the CA deployment in real mobile networks, we conduct comprehensive measurements and summarize the statistical information of collected datasets in Table 1. Operators, Locations and Mobility. Our measurement campaign mainly focuses on three major US mobile operators: AT&T, T-Mobile, and Verizon. We survey two large U.S. cities, covering their urban downtown, surrounding suburban areas, and major city beltways. Besides measurements conducted under driving mobility, which create a comprehensive coverage map, we also experiment with stationary mobility at various city hot spots, such as bus stops in line-of-sight to base stations, providing the baseline results under an ideal channel condition. In addition to outdoor measurements, we perform indoor measurements under walking mobility, a recognized challenging scenario for 5G. Overall, we acquire a rich dataset that provides us with a representative view of the current state of CA deployment. When showing the measurement results, we use OpX, OpY, and OpZ to obscure the operator names. Tool setup. We use the consumer smartphones as the 5G probes listed in Table 5 and ensure their firmware has been updated to the latest version as the 5G network configurations constantly evolve. These phones are tethered to a laptop running the professional network diagnostic tool XCAL [3]. This tool is used to access the chipset diagnostics and log data, including RRC control messages and precise radio frequency (RF) layer information. We adopt Iperf3 with multi-threads data transmission for our throughput measurement and set up an AWS EC2 instance (m5n.xlarge) as the remote server, which is able to provide a 4.1 Gbps baseline throughput and up to 25 Gbps burst throughput.

Methodology. There are numerous practical challenges, marked as [C], that cannot be neglected. These challenges may impact our results and, therefore, require specific remediation.

[C1] We cannot directly control the carrier aggregation or easily select which channel becomes the component carrier. We resort to using the built-in function to force the technology and band that UE can use. For example, under the default coverage area of the band n71+n41 combination, we can force the UE to only use the channel within band n41 by entering the operator service code *#2263# for OpZ, while similar options are available on our professional tools XCAL [3] for the other operators.

[C2] The continuous large-volume measurement traffic makes our UEs potentially compete for radio resources with other users and face throttling issues despite subscribing to unlimited data plans. To mitigate this, we mainly conduct experiments at midnight when fewer people are on the streets, and utilize multiple SIM cards to avoid potential data caps. We cross-validate the measurement results collected by running experiments multiple times on different days and filtering out outliers to ensure the results are representative. On the other hand, we have also collected the data at different times of the day to capture the time diversity, see the discussion in the Appendix B.2.

[C3] The intensive data transmission with CA will quickly overheat the phone, leading to CA deactivation and a significant performance drop. We address this issue by engineering simple closed-loop liquid heat exchangers for cooling our smartphones and actively monitoring the phone's temperature during the measurement.

[C4] The data recorded by multi-phones can lead to many out-of-synchronization problems and thus unfair performance comparison among different operators or CA configurations. To solve it, we place the phones side-by-side and master them on the same laptop; see Fig. 22 for example. Altogether, we carefully design our experiments, which provide a revealing snapshot of the state-of-the-art CA deployments.

MEASUREMENT & QOE IMPLICATIONS

This section presents the main findings from our measurement study, highlighting the key benefits and challenges posed by CA, with additional measurement observations provided in Appendix A.

Using an XR application as an example, we also illustrate the impact of CA on application QoE.

CA Channel Allocation and Deployment

We start by discussing the CA channels, channel combinations, and CA deployment characteristics we have observed.

Diverse Channels and Channel Combinations. In our measurement data, we observed a total of 44 unique 5G channels and 86 unique 4G channels used for CA by the three major US operators. Most 5G channels come from the mid-or high-band (mmWave) ranges, operating in the TDD mode, with only a few operating in the FDD mode (in the below 2 GHz spectrum). In contrast, most 4G channels come from bands below the 2 GHz spectrum, operating in the FDD mode. The main difference lies in the channel bandwidth:

The 4G channel bandwidth varies from 5 to 20 MHz. Whereas all 5G mid-band channels have a bandwidth of at least 20 MHz, most have a bandwidth of 40, 60 up to 100 MHz. In the case of mmWave channels, they are all 100 MHz. Table 2(a) provides representative samples of the 4G/5G bands/channels, and Table 6 in Appendix A.1 provides more detailed information. We see that many 5G channels share the same band with 4G channels -this is because cellular operators often "re-farm" their 4G spectrum for 5G services. Via CA, the individual 4G/5G channels may be combined in various ways to form an aggregated channel of higher bandwidth. As selected combinations highlighted in Table 2(b), in the 4G networks of all three operators, up to 5 channels may be aggregated to yield an aggregated bandwidth up to 100 MHz -which is the maximum (allowed) channel bandwidth of a single 5G mid-band channel. In 5G networks, both OpX and OpY support up to 2 CCs of low-and mid-band channels (in FR1), with an aggregated bandwidth of up to 120 MHz and 160 MHz, respectively. Both these operators also support up to 8 CCs using mmWave channels (in FR2), yielding an aggregate bandwidth of up to 800 MHz. In contrast, OpZ supports 120 120 140 160 160 175 180 BandWidth (MHz) 0 5 0 0 1 0 0 0 1 5 0 0 2 0 0 0 Throughput (Mbps) n41 a +n25 n77 a +n77 b n41 a +n41 b n77 c +n77 d n41 a +n25+n41 b n41 a +n71 a +n25+n41 b n41 a +n71 b +n25+n41 b 0 10 20 30 40 50 60 Timestamp 0 100 200 300 400 500 600 700 800 Throughput (Mbps) Avg. = 509.39 Mbps Std. = 47.68 Mbps Avg. = 211.49 Mbps Std. = 13.62 Mbps n41 n25 (a) Individual. 0 10 20 30 40 50 60 Timestamp 0 100 200 300 400 500 600 700 800 Throughput (Mbps) Avg. = 668.07 Mbps Std. = 132.39 Mbps CA:n41+n25 (b) CA: n25+n41. up to 4 CCs using channels from FR1 only, yielding an aggregated bandwidth of up to 180 MHz. See Table 7 in Appendix A.1 for sample channel combinations used in 5G CA. 4G/5G CA Deployment Prevalence. Our driving measurements in urban, suburban, and city beltway areas reveal that all three operators have widely deployed CA for both 4G and 5G networks, giving the mobile users a high likelihood of utilizing CA, as shown in Fig. 25. We observe that 4G CA covers almost the entire measurement area and provides seamless services, while 5G CA shows varying levels of prevalence across different areas, with averages of 24%, 44%, and 86% for OpX, OpY, OpZ, respectively. For the OpX and OpY, the deployment of mmWave 5G with up to 8 CCs aggregation is confined to densely populated areas. However, there is a notable ongoing expansion of their new 2 CCs aggregation in FR1.

In contrast, OpZ has aggressively deployed 5G CA by re-farming their radio resources, thus providing more diverse CA options (more details can be found in Appendix A.1) and wider coverages not only in urban area but also suburban and city beltways. Therefore, we frequently use measurement data from OpZ in our following study. Fig. 4 visualizes a sample spatial map of 5G CA deployment in an urban downtown area covering approximately 1 2 , with measurements conducted along various streets. The color schemes indicate the number of CCs observed. We see that as the UE moves along a route, the number of CCs may fluctuate, with either a new CC added or an existing CC removed. Below, we will proceed to examine the implications of these observations.

CA Benefits and Challenges

As shown in Fig. 1 and Fig. 2, CA can significantly boost the throughput performance of both 4G and 5G networks. This is particularly the case in terms of peak 5G throughput under ideal channel condition: by aggregating 4 5G mid-band channel components, OpZ attains a peak throughput of 1.7 Gbps; with up to 8 mmWave channel components, OpY attains a peak throughput of 4.1 Gbps. In addition, we provide more measurement observations in Appendix A.2 regarding different mobility and scenarios. Fig. 26 shows that throughput can be significantly increased, up to multiple times, due to the use of CA while driving. Fig. 28 demonstrates that using FDD low-band as PCell in CA helps improve 5G signal connectivity while walking indoors, thereby providing higher throughput.

However, such significant performance gains are achieved at the expense of much added configuration complexity and performance variability, which make analyzing and predicting 5G network performance far more challenging. In 4G networks, because of the much narrow component carrier bandwidth, which varies from 5 to 20 MHz (in our data we observe that CCs of 20 MHz are most frequently used by all three operators), the observed aggregate 4G throughput is closely correlated with the number of CCs used in 4G CA. This is not true in 5G networks, due to more diverse channels/bands and much wider and varied channel bandwidth. Fig. 5 shows the "violin" plots of the measured (aggregate) 5G throughput under 6 different 5G CA combinations from 2CCs to 4 CCs. We use the superscripts to distinguish different channels of the same band. Both with 2 CCs and an aggregated bandwidth of 120 MHz, the throughput performance of the n41 ė +n25 combination differs vastly from that of the n77 ė +n77 Ę combination (of two different n77 channels): the average throughput of the former is below 250 Mbps, about 1/2 of the latter, which is just below 500 Mbps. Both with an aggregate bandwidth of 160 MHz, the n77 ę +n77 Ě combination and n41 ė +n25+n41 Ę combination also exhibit quite different overall performance as indicated by the "fatness" of the contours, although they attain nearly the same peak throughput. Both with 4 CCs, the n41 ė +n71 Ę +n25+n41 Ě combination exhibit significantly higher throughput than the n41 ė +n71 ė +n25+n41 Ě , whose SCell uses the different channel within the same band. With 4 CCs and slightly wider aggregate bandwidth, the n41 ė +n71 Ę +n25+n41 Ě combination attains similar peak throughput as that of the n77 ę +n77 Ě and n41 ė +n25+n41 Ę combinations, but its overall throughput performance is more consistent, with higher average throughput than the 2 CCs and 3 CCs of 160 MHz. Compared to 4G CA, 5G CA in general introduces far higher performance variability, as noted in Fig. 2 and Fig. 24.

To better illustrate the performance variability introduced by 5G CA, in Fig. 6, we plot 60-second sample throughput trace segments of two 5G channels, n25 and n41, when both are used alone (i.e., no CA) as well as a sample throughput trace segment when they are combined as the 2 CCs aggregation (n41+n25). Data was collected at a fixed location with stationary UE, and the band was locked using the built-in function. First, we observe that the aggregate e throughput of n41+25 is not merely the sum of those of n41 and n25, sometimes at least 49.02% lower than the (theoretical) sum. In §4.3 we provide an explanation for this phenomenon. In fact, the channel characteristics and performance profile of an individual 5G channel can in general vary considerably from when it is used alone and when used in different CA combinations -this is because the configured power and MIMO layers may be altered (see §4.3 for an example using n41). In Fig. 7, we plot a 120-second sample throughput trace segment with up to 4 CCs, where the CCs are dynamically added or removed as the user drives in a downtown area. We see that the addition and removal of CCs introduce drastic fluctuations in 5G throughput. For example, due to CC removal, around the time instance 46 sec, the throughput drops by about 1/2, from 1.2 Gbps to around 600 Mbps within a second or so; whereas from the time instance 90 to 94 sec, the throughput increases quickly from 550 Mbps to around 950 Mbps, due to the addition of 2 CCs.

Besides SCell's activation/deactivation, PCell may dynamically switch from one band to another, introducing additional complexities, such as transitions from TDD band to FDD band with altered power allocation. For an example of this occurring when a user moves from outdoor to indoor, see Appendix A.2. Furthermore, during the periods without addition or removals of CCs, the throughput of the aggregated channel often fluctuates more significantly than when no CA is used. All in all, while 5G CA provides cellular operators with the ability and flexibility to better utilize fragmented 5G channels by combining them to form aggregate channels of much higher bandwidth and boost 5G network throughput performance, it also poses new challenges. In the following, we explore the implications of 5G CA on application QoE.

Application QoE Implications

To illustrate the impact of 5G CA on application QoE, we use ViVo, an immersive XR application developed in [16] as an example. ViVo employs 3D point clouds to represent objects and the environment. To stream objects and their environment over networks, two key mechanisms are employed: a) ViVo first predicts the viewer's field of view 150 ms ahead to determine visible and unobstructed objects, shaping the 3D frame that must be delivered in the next 150 ms. b) ViVo adjusts the quality level (defined by point density) of the 3D frame to meet bandwidth constraints and the 150 ms delivery deadline. Similar to conventional video streaming ABR (adaptive bit rate) algorithms, ViVo uses past bandwidth measurements to estimate the available bandwidth in the next 150 msfoot_1 . Application QoE is therefore measured using two metrics: i) (average) quality level measured frame by frame; ii) (average) stall times, where a stall occurs if a 3D frame cannot be delivered within 150 ms.

We consider two scenarios. 1) ViVo over a 5G channel without CA: The average throughput is 355 Mbps with a standard deviation of 161 Mbps, and the peak throughput is 759 Mbps over various experiment runs. A segment of the throughput trace is shown in Fig. 6. 2) ViVo over an aggregate 5G channel with 4CCs 3 , where CCs may be dynamically added or removed: The average throughput is 700 Mbps with a standard deviation of 331 Mbps, and the peak throughput is 1732 Mbps. Refer to Fig. 7 for a segment of the throughput trace. In case 1), the maximum resolution of 3D point clouds and frame rate requires a maximal bit rate of 375 Mbps, slightly above the mean channel throughput. In case 2), to leverage the much higher aggregate channel bandwidth, we scale up the maximum resolution of the 3D point clouds and frame rate accordingly to 750 Mbps, also slightly above the mean throughput of the aggregate channel. In both cases, ViVo adapts to the fluctuating channel throughput by adjusting quality levels. To underscore the impact of 5G CA and motivate the need for CA-aware throughput prediction, we also consider an ideal version of ViVo, where the actual throughput in the next 150 ms interval (instead of estimated based on past measurements) is used.

Fig. 8 shows a number of representative results using various 5G traces with (a) no CA (i.e., case 1) and (b) with (upto) 4 CCs (i.e., case 2). Using the ideal ViVo as the baseline, the results are shown as the percentage of (average) quality degradation and the percentage of increases in (average) stall times. We see that without CA, there are multiple instances where both quality and stall times have degraded more than 5%. In the case of 4CC, the performance of most instances is visibly worse. While in several instances, the quality gradation is kept at 5%, this is achieved with significantly worsening performance in terms of stall times.

The above results underscore the fact that the current applications struggle to fully utilize the 5G network capabilities. It calls for advanced 5G performance prediction models to aid applications in more effective and adaptive decision-making, as in the case of ideal ViVo.

DISSECTING KEY FACTORS AFFECTING CA FOR THROUGHPUT PREDICTION

This section explores the complex interplay among various radio parameters that shape CA configurations and performance. Our goal is to develop an effective throughput prediction algorithm, so our focus lies on UE-collectible parameters or "features" (cf. Table 3), e.g., via Android APIs [6]. We use OpZ as the primary mobile 3 We note that the 4 CCs can aggregate up to 180 MHz bandwidth and do not equate to a fourfold increase in throughput as each channel has a different bandwidth (see Fig. 5 and Table 2). n71 n25 n41 n77 n261 Channel 0 2 4 6 8 10 bps per Hz 3.8 9.5 7.4 7.6 4.5 Figure 10: Selected 5G channel efficiency with good channel condition (CQI>12).

operator, given its extensive CA coverage and diverse channel combinations.

Key 5G PHY Radio Parameters

From the physical layer (PHY) perspective, 5G throughput (theoretically) depends on several key parameters. With CA, aggregated throughput is the sum of individual CC throughputs, making the number of CCs crucial. For each CC, its throughput depends on several key factors: the maximum channel bandwidth which determines the (maximum) configured resource blocks (RBs) and thus affects the number of RBs that may be allocated to each UE (in the frequency domain); the number of symbols allocated to the UE per slot (in the time domain); and the number of MIMO layers (#Layers) used. The frequency and time domain radio resource allocation together yield the number of resource elements (REs) allocated to the UE per slot. The number of bits carried in each RE and thus the total number of bits carried in each slot -referred to as the transport block size (TBS) -is determined by the modulation and coding scheme (MCS). Hence, the throughput (Tput) of each CC is a function of these parameters, namely, = ( , # , # ) (see Appendix B for the theoretical calculation of PHY throughput). The sample results in Fig. 9 show the relationship among the PHY throughput, measured in terms of TBS (bits), the number of symbols allocated per slot, and employed MCS, where the number of MIMO layers is fixed to 2.

Both MCS and the number of MIMO layers hinge on channel conditions (as well as other factors such as the amount of buffered user data). UE periodically feeds back channel state information, such as channel quality indicator (CQI) and rank indicator (RI), to aid the base station in deciding on the MCS and the number of MIMO layers for data transmissions. CQI itself is a function of the reference signal received power (RSRP) and quality (RSRQ), signal-to-noise-ration (SINR), etc., measured at the UE.

As channels in different bands have distinct radio propagation characteristics and may be subject to varying environmental factors, not all channels/bands are equal. As an example, Fig. 10 shows the spectrum efficiency (defined as bits per second (bps) per Hz) for five channels from the low-, mid-and high-band ranges, measured under an ideal scenario (the best channel condition using the highest MCS and full RB allocation). Whether CA is invoked (configured) -and when it is invoked, the number of CCs used -also hinges upon the channel conditions (among other factors).

Need for Modeling Each Channel/Band

We argue the need for capturing the channel characteristics or "features" of individual component carriers separately, especially when they come from different bands, in order to predict their throughput accurately. We use the Synchronization Signal RSRP (SS-RSRP) as an example feature to illustrate the point. We consider two cases involving CA with two intra-band CCs vs. two inter-band CCs: 1) one n41 CC of 100 MHz bandwidth (PCell) and another n41 CC of 40 MHz (SCell); and 2) one n41 CC of 100 MHz bandwidth (PCell) and another n25 CC of 20 MHz.

In Fig. 11 and 12, we present the correlation (measured by the Pearson coefficient) between reported RSRP and observed throughput of PCell and SCell for intra-and inter-band scenarios. Fig. 11 (a) and (b) show the correlation between PCell RSRP and throughput in the case of intra-band CA using 2 n41 CCs. In Fig. 12 (a) and (b), we show the same results in the case of inter-band CA using 1 n41 CC and 1 n25 CC. In both cases, there are strong correlations (Corr > 0.6) between the throughput of each CC and its RSRP. In contrast, in Fig. 11 (c) & (d) and respectively Fig. 12 (c) & (d), we show the correlation of the RSRP of one CC with the throughput of another CC. We see that in the case of the intra-band CA, the correlation is still above 0.6, whereas in the case of the inter-band CA, the correlation has dropped significantly, to only 0.5 in terms of PCell-RSRP and SCell-Tput and to 0.55 in terms of SCell-RSRP and PCell-Tput.

In Fig. 13a and Fig. 13b, we further compare intra-band CA vs. inter-band CA by examining the correlation between the RSRP of the PCell and that of SCell. We see that for intra-band CA, the RSRPs of two CCs are highly correlated and track each other very well over time. Whereas for inter-band CA, the RSRPs of two CCs are not strongly correlated over time. Therefore, for inter-band CA, simply using the reported RSRP of one CC (say, the PCell) to predict the throughput of another CC (e.g., one of the SCells) may lead to a suboptimal result.

Need for Accounting for Complex & Dynamic Feature Interplay

In the above, we see that the channel's RSRP is strongly correlated with its observed throughput. However, the RSRP of a channel alone is not sufficient to predict its throughput. We will use examples to illustrate. The examples demonstrate the need to account for the complex and dynamic interplay among various radio channel features, and thus make the case for sophisticated deep learning methods for CA-aware throughput prediction. We first consider and compare the throughput of a channel when CA is not invoked vs. that of the same channel when it is used as part of 3CC CA. Fig. 14 shows the measured throughput results for an n25 channel with and without CA at the same location. In both cases, the reported RSRP and CQI of the channel are similar (and the number of allocated RBs is also similar). We see that there is a significant difference in observed throughput: without CA, the throughput of the n25 channel is above 200 Mbps on average, whereas with CA, its throughput drops to only slightly above 100 Mbps on average. This is due to the fact that with CA, the number of MIMO layers used for the n25 is reduced from 3 to 1 (likely due to the reduced transmission power for the n25 channel by the base

(a) PCell RSRP v.s. PCell Tput. (b) SCell RSRP v.s. SCell Tput. (c) PCell RSRP v.s. SCell Tput. (d) SCell RSRP v.s. PCell Tput. CA_n25 NonCA_n25 1 0 0 1 5 0 2 0 0 2 5 0 Tput (Mbps) CA n41_n25_n41 NonCA n25 RSRP -68.1dBm -69.6dBm CQI 12.2 12.1 MIMO 1 3 #RB 102.4 103.4 Total Tput 862 Mbps 212 Mbps Figure 14: Throughput of the same channel with and without CA, using n25 as an example. station). In Fig. 15, we compare the throughput of the same channel (an n41 channel of 40 MHz) that is used in different CA band combinations: 1) n41-n41 intra-band CA (with the first n41 channel of 100 MHz being the PCell) and 2) n25-n41-n41 inter-band CA (with the n25 channel being the PCell). We see that the throughput of this n41 channel (of 40 MHz) differs significantly in these two cases, despite the fact that it is used as the SCell in both cases, and the RSRP and CQI are also similar. Furthermore, the number of MIMO layers used is also the same. The difference in the observed throughput can be attributed to the number of RBs allocated. With the other CCs having 120MHz bandwidth, the additional SCell may be throttling out in the service busy area. The above examples show the importance of taking into account the diverse band combinations as well as the complex and dynamic interplay among various channel parameters. Therefore, CA-awareness is critical to 5G throughput prediction.

CA-AWARE TPUT PREDICTION MODEL

This section introduces Prism5G, a novel CA-aware deep learning framework for 4G/5G throughput prediction. It models and conditions individual CCs, fusing them based on CA state for accurate 5G throughput prediction (see Fig. 16), using the UE-collectible features (see Table 3).

Overview of Prism5G

Challenges. The widespread adoption of CA in 4G/5G has significantly improved achievable throughput for end users. However, accurately modeling the performance of throughput on the UE side becomes particularly challenging in the presence of CA, especially within the realm of 5G.

(1) Heterogeneity: The diversity of 5G channels, and their combinations, exhibit distinct characteristics. Additionally, the availability and configuration of these channels may vary across locations and network deployments.

(2) Complexity: The intricate interplay and correlations among channels within a CA configuration escalate the complexity, rendering the explanation of the ML model challenging. (3) Data Scarcity: The laboriousness of measurements and commercial constraints impede telecom companies from scaling the open-sourced datasets obtained from their commercial networks, reducing the data available for training and evaluating ML models. (4) Lightwight: The limited compute resources on mobile UE and need for real-time inference necessitate lightweight ML models.

Our Design. To address these challenges, we design a deep learning framework, denoted as Prism5G, which exhibits adaptability, flexibility, and explainability for predicting 5G network throughput. Distinguishing itself from existing approaches [4,28,32] that blindly predict overall throughput, Prism5G endeavors to harness the above measurement observations and domain knowledge of 5G networks to explicitly account for the impact of the CA mechanism. Fig. 16 shows the overall design of Prism5G. It consists of three core principles: (1) modeling of each CC (blue), (2) monitoring the signaling events (green), and (3) fusion learning for the interactive correlation (orange). Specifically, Prism5G adopts a weights-shared neural network to predict the future throughput of each CC. These individual predictions are then aggregated to obtain the overall throughput. Such explicit consideration of CA enables Prism5G to make predictions at a fine-grained level (per CC), thereby achieving a certain degree of flexibility and explainability compared to directly modeling the overall throughput. Meanwhile, Prism5G takes the signaling events and distinct capabilities of different chipset modems as domain knowledge for each prediction. This knowledge is transformed into masks that explicitly adjust the states of carrier components. By incorporating the domain knowledge, Prism5G achieves faster adaptation to network environment changes and reduces complexity by eliminating the need for indirect learning from the physical layer. Finally, Prism5G also explicitly considers the interaction between CC under different channel combinations. Collectively, these designs empower Prism5G with the capability to effectively and accurately predict the throughput performance of 4G/5G networks. Detailed information for each module is provided below.

Model Module Explained

Per CC Modeling. Let denote the total number of carrier components and X ę denote the features of each carrier component . These features (see Table 12) include information from the past T time steps, such as throughput and physical channel quality, i.e., Đ ) . Based on these input features, we employ an RNN module to predict the future throughput of each carrier component:

, where 1 denotes the trainable parameters of the RNN layers. The RNN modules share weights to leverage shared knowledge and reduce the number of parameters, thereby decreasing the overall complexity of Prism5G. The type of RNN module is configurable, and we use LSTM in our experiments. Although we use it as our building block when developing Prism5G, the design of Prism5G does not require a specific DNN architecture. In other words, it can be easily replaced by other similar (or more advanced) building blocks, such as transformers. CA Event Monitoring. To capture the dynamics of the channel combination over time, Prism5G translates the signaling control messages obtained from the Radio Resource Control (RRC) protocol into a binary mask vector, denoted as I ∈ (ÿ,Đ ) . This vector is responsible for activating and deactivating RNN modules (i.e., carrier components): X ′ ę = X ę » I. Furthermore, in order to provide the fusion learning module with a richer context of the current channel combination, Prism5G utilizes an embedding layer to transform the sparse binary mask vector I into a dense embedding E. Fusion Learning. This module fuses the RNN hidden states of all carrier components to extract the interplay and correlations among different channels, taking into account the current channel combination condition: ℎ Ĝ = Ă 2 ( [ℎ 1 , . . . , ℎ ę , . . . , ℎ ÿ , E]), where 2 represents the trainable parameters of the fusion learning module. Aggregated Throughput Prediction. Prism5G first aggregates the RNN hidden state ℎ ę of each component carrier and the overall channel correlation ℎ Ĝ , i.e., ℎ ′ ę = ℎ ę + ℎ Ĝ . This aggregated information is then inputted into an MLP module to predict the future throughput of each carrier. The overall throughput for an end user is obtained by aggregating all the predicted throughput for each carrier. The whole process is presented as

where ĦĨěĚ denotes the predicted aggregated throughput. We jointly train all the aforementioned modules. The optimal parameters Θ = [ 1 , 2 , 3 ] of Prism5G are obtained by minimizing the prediction errors:

EVALUATION

This section evaluates Prism5G's performance, highlighting a 14.0% reduction in root-mean-square-error (RMSE) compared to the best baseline, with efficient training and inference times. Notably, Prism5G displays remarkable adaptability to network changes and strong transferability across diverse settings.

Experiment Setups and Implementations

Datasets. We extract and process self-collected datasets at 10ms and 1s time granularity, obtaining a total of 6 sub-datasets for machine learning evaluation, each comprising thousands of valid data records. These sub-datasets encompass diverse operators, channel combinations, and scenarios. Each scenario contains 10 traces with 0 25 50 75 100 125 150 175 200 Timestamps 2 0 0 4 0 0 6 0 0 8 0 0 1 0 0 0 1 2 0 0 1 4 0 0 1 6 0 0 Throughput (Mbps) Real Prophet 140 145 150 155 2 0 0 .0 4 0 0 .0 6 0 0 .0 8 0 0 .0 (Z2) 70 75 2 0 0 .0 4 0 0 .0 6 0 0 .0 8 0 0 .0 (Z1) (a) Prophet. 0 25 50 75 100 125 150 175 200 Timestamps 2 0 0 4 0 0 6 0 0 8 0 0 1 0 0 0 1 2 0 0 1 4 0 0 1 6 0 0 Throughput (Mbps) Real LSTM 140 145 150 155 2 0 0 .0 4 0 0 .0 6 0 0 .0 8 0 0 .0 1 0 0 0 .0 (Z2) 70 75 2 0 0 .0 4 0 0 .0 6 0 0 .0 8 0 0 .0 (Z1) (b) LSTM. 0 25 50 75 100 125 150 175 200 Timestamps 2 0 0 4 0 0 6 0 0 8 0 0 1 0 0 0 1 2 0 0 1 4 0 0 1 6 0 0 Throughput (Mbps) Real Prism5G 140 145 150 155 2 0 0 .0 4 0 0 .0 6 0 0 .0 8 0 0 .0 (Z2) 70 75 2 0 0 .0 4 0 0 .0 6 0 0 .0 (Z1) (c) Prism5G. 690 710 730 750 770 790 Timestamps (ms) 6 0 0 7 0 0 8 0 0 9 0 0 1 0 0 0 1 1 0 0 1 2 0 0 1 3 0 0 Throughput (Mbps) (Z1) Real Prism5G (a) Throughput drop. 1400 1450 1500 1550 1600 Timestamps (ms) 0 2 0 0 4 0 0 6 0 0 8 0 0 1 0 0 0 Throughput (Mbps) (Z2) Real Prism5G (b) Throughput boost. 300 to 600 data samples per trace. In Appendix C, Table 11 summarizes the statistical information of these sub-datasets, and Table 12 illustrates the description of each data field. Each trace uses a moving window to create data pairs, consisting of a historical window and a future window. Those data pairs are then split into training, validation, and test sets based on a specified ratio. Baseline Setups. We compare Prism5G with six baseline models, which can be categorized into three groups: (1) Statistic-based time-series forecasting: Prophet [44];

(2) Classical machine learning (widely used for throughput prediction due to their explainability): GDBT [32], RF [4]; (3) Deep learning-based: LSTM [28], TCN [9], and Lumos5G [32] foot_2 . For all the evaluations, we set the input and output sequence length to 10, signifying the 100ms (or 10s) prediction horizon, depending on the dataset time granularity. We use root-mean-square-error (RMSE) as the loss function and report the optimal model based on its validation set performance.

For the concrete training and evaluation strategy for Prophet and classical ML, refer to Appendix C.1. Runtime. Compared with LSTM, Prism5G introduces an additional 34.1% training time on average. The training duration typically ranges from 5 to 30 minutes, depending on the dataset and hyperparameter selection. As for inference, although Prism5G incurs an extra 23.2% time on average, yet remains below 1 ms per sample, well below decision time (10ms for ViVo and 1s for ABR in §7).

Prediction Results and Comparisons

Overall Accuracy. Table 4 shows the overall performance of Prism5G and baselines performance measured in RMSE on the collected 5G traces. The experiments are conducted in both the short period using 10 ms time scale with 100 ms prediction horizon, and the long period using 1 s time scale with 10 s prediction horizon. We can see that: (1) Prism5G is consistently superior to the current best baseline, with an average of 14% and a maximum of 22% reduction in RMSE.

(2) Prism5G outperforms other algorithms in both time scales with an average of 17.4% and 10.6% reduction in RMSE for the short and long time scales, respectively. (3) Purely time series prediction algorithms (e.g., Prophet) do not work well on 5G datasets, having the highest RMSE across all the other algorithms for nearly all datasets.

We visualize the prediction results in Fig. 17 by plotting the first predicted point in the horizon window. For simplicity, we only select two representative baselines (Prophet and LSTM) for comparison with Prism5G. We delve into two critical instances: area Z1 is marked by a significant throughput drop due to SCell deactivation and worse channel quality, and area Z2 is characterized by a notable performance boost due to SCell activation and better channel quality. We can notice that in Z1, Prophet and LSTM are overestimating the throughput, while underestimating it in Z2. Prism5G has the closest prediction to the real throughput in both

0.0 2.5 5.0 7.5 10.0 Stall Increase (%) 28 26 24 22 0 Quality Degrade (%) Ideal QoE ViVo ViVo+Prophet ViVo+LSTM ViVo+Prism5G 30.0 32.5 35.0 37.5 40.0 Stall Time (s) 465 470 475 480 485 Average Bitrate (Mbps) Better QoE MPC MPC+Prophet MPC+LSTM MPC+Prism5G cases. We refer the reader to Appendix C.2 for additional results, where Fig. 33 and Fig. 34 show each CC's predicted results by Prism5G, where each cell is well modeled. Transition Point. We also visualize the 10ms short-period prediction results of the representative transition point areas Z1 and Z2 in Fig. 18. This allows us to see that Prism5G exhibits a faster reaction at the transition points and closely matches the real throughput. In turn this is indicative of its availability to assist application decisions at a finer time granularity. Ablation Study. We perform an ablation study to demonstrate the necessity of two key mechanisms in our model: (1) the state trigger mechanism and (2) the fusion mechanism. As shown in Table 13 in Appendix C.2, without the state trigger mechanism, the average and maximum RMSE increase by 5.3% and 7.1%, respectively. Similarly, without the fusion mechanism, these increase by 6.2% and 9.5%. Generalizability. Expanding on our previous evaluation, where each trace contributes to both training and test sets, we delve deeper into assessing the generalizability of the proposed Prism5G. We use the sub-dataset of OpZ with the walking mobility at 1s time scale for this study. We first evaluate the Prism5G on the same route but different runs. To do this, we split the dataset based on the traces instead of spli ing them as in the previous evaluation. Table 14 showcases the consistent superiority of the Prism5G, achieving an average 9.4% lower RMSE compared to the best baselines. Additionally, we extend our evaluation to include new traces collected on different routes, not part of the original subdataset. We observe that the Prism5G outperforms all baselines, maintaining strong performance with an average RMSE 12.5% lower than the baselines.

In summary, Prism5G outperforms the current state-of-the-art algorithms and has the highest accuracy for predicting 5G throughput for the aggregate and individual cells. It quickly adapts to transition points when an SCell is activated or deactivated with high accuracy.

PRISM5G USE CASES

This section employs two use cases to showcase Prism5G's efficacy in enhancing application QoE through smart decision-making. We also compare Prism5G with Prophet and LSTM. XR Immersive Content Delivery. Recall from §3.3 that ViVo employs a simple mechanism to estimate future network bandwidth based on past data and use it to decide on the quality level of 3D frames that are 150ms ahead. The decisions are made frame-byframe at a rather fast time scale (10's ms). We replace this simple mechanism with Prism5G, which predicts 5G throughput at a fast time scale (10s ms) with a short horizon (100s ms). We denote this as ViVo+Prism5G. We also compare with ViVo+Prophet and ViVo+LSTM. Experiments are done using 2300+ 5G traces with (up to) 4CCs with average aggregate throughput from 223.86 to 627.69 Mpbs. We use the scaled-up ViVo application as in §3.3. The results are shown in Fig. 19. We see that ViVo+Prism5G attains near-optimal performance, with QoE results very close to those of the ideal ViVo. In contrast, ViVo+LSTM also yields improved QoE performance, but it is far from near-optimal. While ViVo+Prophet also improves the average quality levels, but the improvements come at the expense of slightly worsening average stall time performance. UHD Video-on-Demand Streaming. In contrast to ViVo, ABR algorithms used in video streaming make decisions at a longer time scale (at the level of chunk length, typically a few seconds long). In words, decisions are made to prefetch several video chunks ahead at certain quality level, and a large client buffer is used to store prefetched data to accommodate network bandwidth fluctuations. We use MPC [50], a widely used throughput-based ABR algorithm, as an example 5 . MPC uses a simple predictor based on Harmonic Mean which estimates the future bandwidth using data from the recent past. We replace this predictor with Prism5G to predict future 5G throughput over a time horizon of 10 seconds at the time scale of seconds. We emulate streaming of 16K videos over 5G networks. Each video is encoded using H.264 codec in multiple quality levels: [1.5, 2.5, 40.71, 152.66, 280, 585] Mbps, corresponding to resolutions [360p, 480p, 2K, 4K, 8K,16K]. Evaluations are performed using the 5G CA traces collected. The results are shown in Fig. 20.

We see MPC+Prism5G yields significant QoE improvements over MPC: while it increases the average bit rate modestly from below 468 Mbps to around 472 Mbps (i.e., a 0.71% improvement), it reduces the average stall time significantly from around 39 seconds to below 31 seconds (i.e., a 19.06% reduction). In contrast, while MPC+Prophet and MPC+LSTM increase the average bit rate considerably (i.e., by 2.5%), these improvements are accompanied by only slight (i.e., 2.8%) reductions in the average stall time. This is largely because during the transition periods when CCs are deactivated, both Prophet and LSTM significantly over-estimate 5G throughput during these periods (cf. Fig. 18). While such overestimates increase the overall bit rates, they can induce significant stall events. Furthermore, what the average QoE statistics do not show are the significant improvements afforded by Prism5G in stall time tail performance (especially during the transition periods when 5G throughput drops significantly, due to CC de-activation). As shown in Fig. 21, MPC+Prism5G improves 99%, 95%, and 90% stall time tail performance by 50.8, 33.0, and 16.0 seconds, respectively, far better than those of MPC+Prophet and MPC+LSTM.

RELATED WORKS

Deployment of commercial 5G services around the world since 2019 has attracted a flurry of measurement studies of 5G networks. These studies have revealed crucial insights concerning coverage, latency, throughput, and application performance [8, 11, 12, 14, 17, 26, 31-34, 36, 37, 39, 40, 47, 49]. In the following we will focus our attention on 4G/5G carrier aggregation and 3G/4G/5G throughput prediction related research. 4G/5G Carrier Aggregation & Measurement. Although 3GPP has released the technical specification of CA in 5G networks since 2019 [1], the commercial deployment of 5G CA is still in its early stages, with fairly limited literature studies. Lin et al. [24] discussed the importance of CA from the technical specification perspective, particularly in the context of operators' co-construction and sharing. Liu et al. [27] conducted a measurement study on the general 5G experience with three US operators and proposed a patch solution for configuring radio resource control and making better choices to select service cells. Li et al. [23] examined the signaling process of CA access from the control plane perspective primarily using commercial 4G measurement datasets, as commercial 5G-CA on mid-band had not been widely deployed at that time. Narayanan et al. [33] performed measurements of mmWave 5G (FR2) deployments but only studied the impact of CC number on peak throughput. Fezeu et al. [11] further study video streaming performance in the context of different frequency bands and examine the impact of channel variability through scaled variability metrics. Other works [15,18,38,40,48] primarily focused on investigating the capacity of 5G channels and achievable performance with limited in-depth discussion toward CA. Additionally, telecommunication companies have conducted preliminary CA measurements [7,35] but only in controlled trial environments rather than on large-scale commercial deployments. Unlike existing studies, we conduct a timely and comprehensive measurement of CA deployment in commercial 5G networks and 4G networks in the US, with the consideration of CA impacts on throughput analysis and prediction. 3G/4G/5G Throughput Prediction. There is a large amount of literature on throughput prediction [21,22,29,41,42,51] for 3G/4G networks, utilizing both machine learning-based and analytical approaches. 5G throughput prediction is notably more challenging than 3G/4G due to diverse bands, complex technologies, and various other factors involved. Lumos5G [32] employed GDBT and Seq2Seq to predict mmWave 5G throughput based on UE-side contexts. Mei et al. [28] employed the LSTM model to capture the temporal patterns of bandwidth evolution. Minovski et al. [30] trained ML models in LTE networks to predict throughput using lower-layer radio-related metrics. They further fine-tuned these models for nonstandalone 5G networks. A hybrid prediction method is proposed in [25], which utilizes an ARMA time series model for intra-cell bandwidth prediction and a Random Forest (RF) regression model for cross-cell bandwidth prediction. However, previous studies have not explicitly integrated CA into ML throughput prediction models, making them inadequate for CA-enabled 5G networks. In contrast, Prism5G stands out as a flexible framework that accommodates various DNN modules. It incorporates measurement observations and domain knowledge of CA into its design, enabling accurate 5G throughput prediction with per-component carrier modeling and accounting for carrier interactions.

CONCLUSIONS

We have carried out a timely in-depth study of carrier aggregation (CA) in commercial 5G (and 4G) networks. Through comprehensive measurement-based analysis, we not only demonstrate how CA can significantly boost 5G network throughput performance, but also illustrate the new challenges posed by CA and their implications on application QoE performance. Our findings lead us to pursue the development of a CA-aware 5G throughput prediction framework that can effectively aid applications in fast and adaptive decisionmaking. To this end, we have identified the key factors influencing whether or when CA is deployed and what band combinations are used. Building on these insights, we have developed Prism5G, a novel deep learning predictor for 5G throughput prediction which explicitly considers the complexities introduced by CA, thus is CAaware. Using real-world 5G traces, we have extensively evaluated the efficacy of Prism5G and compared its performance with existing 5G throughput prediction algorithms. Our results demonstrate that the Prism5G outperforms the state-of-the-art algorithms by over 14% on average in terms of prediction error. Through two use cases, we further demonstrate the utility of Prism5G in aiding applications (and networks) to optimize QoE performance. Future Directions. The performance of CA under multi-user competition scenarios has not been well studied and requires further exploration. We plan to evaluate the trade-offs of adapting various learning models, such as transformers, to the Prism5G framework. Future efforts will also focus on implementing a real-world benchmark platform with downstream applications for mobile throughput prediction evaluation and optimizing system overhead. Ethical Considerations. The measurement study was conducted by the research team in compliance with wireless carriers' agreements. No human subjects were involved, nor was any personally identifiable information collected. A.2 Impact of CA on 4G/5G Performance.

We now present the impact of CA on 5G (and 4G) performance for all three surveyed mobile operators under various mobility and scenario settings. Ideal Channel Condition. Fig. 1 and Fig. 23 report the achievable throughput for all three mobile operators under stationary mobility with the line-of-sight to the base station. For the downlink, 4G and 5G mmWave (FR2) throughput increases exponentially due to each SCell using a bandwidth similar to that of PCell. In the 5G

50 100 150 200 250 300 350 Throughput(Mbps) 0 0 . 2 0 . 4 0 . 6 0 . 8 1 PDF(%) Avg. = 107.73 Mbps Std. = 69.64 Mbps low-/mid-band (FR1), OpZ also achieves an average of 1.5 Gbps by integrating three smaller bandwidth channels as SCells, resulting in more than double the throughput of not using CA. Even though a lower number of SCells are configured, OpX and OpY still achieve an average of 1.3 Gbps and 1.6 Gbps throughput with the aggregation of C-band. Despite news [5] of its deployment, we only observe CA on the uplink for 5G mmWave, which has 321.5 Mbps uplink throughput in limited locations.

Outdoor Driving. Fig. 25 shows the prevalence of CA observed from our driving experiments. We see that CA is widely used for 4G services by all three carriers, whether it is urban, suburban, or along highways. As for 5G services, OpZ has deployed CA aggressively, not only in urban environments but also in around 75% of the suburban areas surveyed and along with many areas along highways. Meanwhile, OpX and OpY's 5G CA deployment has also achieved significant progress, covering around 25% and 54% of the surveyed urban areas, where their high-band aggregation takes up 6% and 25%, respectively. For a comparison of the 4G and 5G throughput using the same driving measurement data, we refer to Fig. 26. We see that CA boosts the 4G throughput for all mobile operators to nearly, and occasionally well above, 100 Mbps in the urban and suburban areas, as well as along the highways. Surprisingly, with CA in FR1 channels, OpZ can nonetheless boost its average 5G throughput to more than 700 Mbps, 600 Mbps, and 350 Mbps in the urban, suburban, and highway settings, respectively. While less CA is observed while driving, urban locations for OpX and OpY have approximately double the throughput when compared to their suburban locations at nearly 450 Mbps and 840 Mbps, respectively. Indoor Walking. We have also conducted indoor walking experiments, which create a more challenging channel environment. From Fig. 27, we see that for all three mobile operators, the indoor downlink throughput experiences a significant drop compared to the results obtained under ideal channel conditions. Meanwhile, OpX and OpY have a high probability of dropping back to 4G. In contrast, OpZ uses the low-band FDD channel as PCell (which extends coverage due to it experiencing less radio path loss and thereby receives higher power, as shown in Fig. 28) and another mid-band TDD channel as SCell (which increases bandwidth) to form FDD-TDD CA. OpZ can expand its indoor 5G coverage and achieve good throughput performance. Impact of UE Capability. CA not only depends on the availability of channel combinations, but also on the capability of UE. This is particularly true when the number of 5G CCs increases. To account for this, we conduct experiments using three Samsung smartphone models. Fig. 29 shows the throughput and percentage of CC detected on the walking dataset for Samsung S10, S21 and S22 phones. Impact of CC Changes. During our driving experiments in urban, suburban, and highway scenarios, we observed that the addition

B THROUGHPUT FACTORS B.1 Theoretical Channel Capacity

Each transport block size (TBS) can be approximated as

where Ċ re is the number of resource element. Ď denotes the coding rate, the ratio between the number of information bits and the total number of bits sent, and č ģ denotes the maximum modulation order (e.g., 6 for 64QAM and 8 for 256QAM). Lastly, Ĭ is the MIMO layers, representing the spatial multiplexing. They altogether decide the number of information bits, i.e., Ċ info . The čīėĤĪğİěĨ will further consider PHY layer processing, such as encoding, and add round bits to fit the system. See [13] for more details.

B.2 Temporal Dynamics.

Although the key measurement results presented in this paper are primarily collected during the cell low-traffic periods to avoid traffic thro ling. We also conduct measurements at different times of the day to show the impact of temporal dynamics. We find that the cell's overall performance remains stable over weeks or months since operators typically do not frequently change hardware or configurations. When examining temporal dynamics on a daily scale, we observe that the signal strength of each carrier component (connecting to the same PCI band) converges, as shown in Table 8. However, we also observe user numbers will cause temporal dynamics and impact throughput. To investigate this issue, we picked two locations on campus with good and bad signal coverage. For simplicity, we visualize one component carrier's throughput traces in Fig. 31 and Fig. 32. The traces are collected at different times of the day, including the rush hour when thousands of students move among buildings between two classes (T1) and non-rush hours (T2 and T3) The color represents the throughput, with green indicating higher throughput and red indicating lower throughput.

Our data reveals that UE will experience significantly lower throughput during rush hour (T1), particularly in locations with suboptimal signal strength. This is because the cell may allocate (a) TBS vs MCS. (b) TBS vs #symbols. fewer resource elements/blocks. For detailed statistics, see Table 9 and Table 10. On the other hand, the other PHY layer parameters, such as channel conditions (CQI), modulation and coding rate (MCS), show li le variance. This validates that the temporal dynamic of 5G throughput can be captured by those features, which makes our modeling feasible. In the CA cases, the total number of aggregated resource blocks decreases at the peak hour, while the signal strength and channel quality (CQI) of each cell remain

(a) T1. (b) T2. (c) T3. consistent. Interestingly, we also find that PCell and SCell occasionally decrease at varying degrees, and SCell may be dropped. We believe this is more related to RAN-side scheduling algorithms, and will leave it for future studies. Table 10: Statistic values of cases in Fig. 32. T1: rush hour T2: non-rush hour T3: non-rush hour CQI 6.2 ± 1.5 6.1 ± 1.4 6.3 ± 1.4 MCS 11.1 ± 2.8 11.2 ± 2.7 11.2 ± 2.7 #RB 64.4 ± 22.9 88.1 ± 22.1 95.0 ± 32.6 In this section, we provide additional information about the machine learning experiment setups and results discussed in §6.

C.1 Experiment Setups

Datasets. Table 11 summarizes the dataset used for machine learning. Table 12 describes the data field. Implementation. We implement Prism5G using PyTorch framework. Each RNN and MLP module has a two-layer architecture with 128 hidden units. The input sequence length is set to 10, while the output sequence length is 10. We normalize the dataset using the min-max scaler and randomly divide the dataset into training, validation, and test sets using a ratio of 0.5:0.2:0.3 for all experiments.

Prism5G and all the deep learning-based baselines are trained using the Adam optimizer [20] with a learning rate of 0.01, a batch size of 128, and a max epoch of 200. We utilize root-mean-square-error (RMSE) as the loss function. The best ML model is selected according to its performance on the validation dataset. All data processing Train. & Eval. strategy for classical ML. The GBDT (gradientboosted decision trees) and RF (random forest) only take the feature for the prediction and don't have an intrinsic design for the time series task. We combine all historical data into a single feature (i.e., Ď (Đ ,ġ ) -> Ď (Đ ×ġ,1) ) and feed it into the algorithm as a whole. We choose the regression tree as our aim is to predict the throughput.

Ablation study of Prism5G. Various components of Prism5G were removed for an ablation study to demonstrate their necessity. The performance of these simplified models in RSME compared to the full model are shown in Table 13.

C.2 Additional Results

Performance of baselines at transition points. Fig. 33 and Fig. 34 display the results of inference performed by Prism5G on the urban downtown driving dataset for 4G and 5G compared to real observed values. A comparison of the performance of Prophet and the LSTM against the real values to show their difficulty in modeling around transition points is shown in Fig. 35 and Fig. 36, respectively. Generlizability. Table 14 presents the results of evaluations conducted on (1) the same route but different runs as the training dataset and (2) entirely new routes not included in the training