Real World Performance of LTE Downlink in aStatic Dense Urban Scenario – An Open Dataset

Vaclav Raida, Philipp Svoboda, Markus RuppInstitute of Telecommunications

Technische Universität WienVienna, Austria

{firstname.lastname}@tuwien.ac.at

Abstract—Open datasets with measurements in cellular mobilenetworks are rare. In this paper, we share 90 hours (over 600 000samples with time-resolution of 500 milliseconds) of smartphonemeasurements conducted in a dense urban environment in thedownlink of an operational LTE network. To capture, on the onehand, the short-term fluctuations of the measured key parameterindicators, and on the other hand, also time-of-day effects, wechose a static indoor setup. Despite the static configuration, thedataset offers a rich variability over time. The dataset can beappealing for machine learning applications (e.g., throughputprediction), simulations verification, or for educational purposes,as it contains detailed statistics throughout different LTE layers,which may help students to better comprehend the concepts ofcommunication in LTE downlink.

Index Terms—dataset, LTE, downlink, measurement, RSRP,RSRQ, RSSI, SINR, throughput, physical, MAC, RLC, BLER,PRB, CSI, CQI, rank, MCS, static, indoor

I. INTRODUCTION

Publicly available measurements in cellular mobile net-works are crucial for verifying the fidelity of simulation tools,statistical models, and machine learning predictions. Unlikein other communities (e.g., image processing), there are onlya few open datasets in the context of wireless networks.Moreover, the current datasets are often of poor quality andoffer rather short traces that last a few tens of minutes.

In this paper, we provide an open dataset [1] that comprises90 hours of static indoor measurements in LTE downlink (DL).The static setup is motivated by the fact that in the currentdatasets with different mobility profiles, it is often not clearwhether the variations of throughput and signal metrics arecaused by moving in space or only by a change in time.

Our measurement results show that even in a static scenario,most of the key parameter indicators (KPIs), on the one hand,fluctuate in the order of milliseconds. On the other hand, KPIsexhibit long-term time-of-day effects, which cannot be visiblein shorter measurements.

We conducted the measurements with Keysight’s NEMO [2]phone (Samsung Galaxy Note 4). From the hardware per-spective, the user equipment (UE) is a regular smartphone.The measurements thus represent well the quality of serviceexperienced by real end-users. However, Keysight softwarecollects KPIs directly from the chipset and thus offers extendedlogging capabilities (more KPIs available, higher time- and

amplitude-resolution) compared to what is available on theapplication layer (e.g., with Android Telephony API [3], [4]).

With a time-resolution of 500 ms, the dataset contains over6 · 105 samples. It provides signal metrics1 (RSRP, RSRQ,RSSI, SINR) for each receive antenna port, channel stateinformation (CSI) fed back to eNodeB, modulation and codingscheme (MCS) and details about physical resource blocks(PRBs) scheduled by the eNodeB, block error rate (BLER),hybrid automatic repeat request (HARQ) retransmissions,throughputs on different layers: physical, application, mediumaccess control (MAC), radio link control (RLC); and more.

A. Related Work

In this section, we summarize open datasets we are awareof, which contain measurements conducted in operational LTEnetworks with end-user UEs (smartphones). For each dataset,we cite the webpage with the downloadable files and thecorresponding scientific publication.

Bokani et al. [6], [7] offer LTE measurements that consistof GPS and throughput measured along a 4.7 km long routein Sydney, Australia, with a ten-second resolution. They con-ducted 72 repeated drives and collected over 15 000 samples.

Hooft et al. [8], [9] collected 40 LTE traces with durationsfrom 3 to 13 minutes, covering 5 hours of measurements inGhent, Belgium. They conducted walk-tests, as well as drive-tests with various vehicles (car, tram, bus, bicycle). The tracescomprise GPS and throughput with a one-second resolution.

Raida et al. [10], [11] provide traces of three Austrianmobile network operators (MNOs) for two repeated, 160 kmlong train drives. The measurements have 500 ms resolutionand contain GPS, frequency band, channel, cell ID, and thefollowing LTE signal indicators: pathloss, RSRP, RSRQ, RSSI.

Meixner et al. [12], [13] shared 377 LTE traces collectedwith a custom Android application in Germany, Netherlands,and in the USA with an approximate duration of 5 minutesper trace. Traces contain environment and mobility profile an-notations, GPS, cell information (area code, cell ID, frequencyband), RSRP, and RSRQ. The advertised sampling period is

1RSRP = reference signal received power. RSRQ = reference signalreceived quality. RSSI = received signal strength indicator. SINR = signalto interference and noise ratio. In the standard [5], the last metric is called“reference signal-SINR” (RS-SINR). We use a shorter name SINR. Someauthors call it only SNR. Android developers call it RSSNR [4].

100 ms. However, as the samples are pulled from AndroidTelephony API [4], which does not refresh the reported KPIsso often, we found many duplicate values in the dataset.It seems that the KPIs are updated only once per 2 s. Thepublished files contain additional columns (e.g., RSSI), whichare unfortunately filled with invalid values. The availabilityof quantities provided by Android API is device-specific. Theused UE thus probably does not support reporting of all KPIs.

Raca et al. [14], [15] offer 135 measurements collectedwith Android application G-NetTrack Pro2 in operational LTEnetworks of two Irish MNOs in different scenarios: static,pedestrian, bus, car, train. Traces contain RSRP, RSRQ, RSSI,SINR, channel quality indicator (CQI), and DL and uplink(UL) throughput with one-second resolution and an approx-imate duration of 15 minutes per trace. Traffic is generatedby downloading a file of size > 50 MB over TCP/IP. Unfor-tunately, GPS quality is very poor. For example, a 25-minutecar drive car/B_2018.01.18_07.30.40.csv betweenCork and Glanmire, which must have at least ten kilometers,contains only 11 distinct GPS points.

B. Paper Outline

Section II presents our measurement setup. Section IIIintroduces the individual KPIs that are available in our opendataset [1]. Section IV summarizes some of our observationsand contains hints regarding possible KPI averaging andresampling. Section V concludes the paper.

II. MEASUREMENT SETUP

We conducted static indoor measurements between 16:00on 8 February and 10:00 on 12 February 2019 in a lockedclassroom3 (the measuring UE was placed on a desk) at theInstitute of Telecommunications, TU Wien, Vienna, Austria.In February, there is a semester break in Austria, and thus nopersons were entering the classroom.

We measured with Keysight’s NEMO smartphone [2], [16]in an operational LTE network (i.e., we shared the cell withother real users) of one of the Austrian MNOs. The phonewas Samsung SM-N910F (Galaxy Note 4). Keysight providesspecial software that enables extended scripting and loggingcapabilities. KPIs are recorded directly from the chipset.

To generate DL traffic, which allows measurement of avail-able throughput, we scheduled an HTTP download of a 40 GBfile from our server at the Institute of Telecommunications, TUWien. Every time, the whole file was retrieved (or when aninterruption error occurred), the measurement script immedi-ately restarted the download.

The measuring UE was locked to a single cell in LTE band20 (frequency division duplex, center frequency ≈ 800 MHz).Channel bandwidth was 20 MHz, which corresponds to 100

2https://www.gyokovsolutions.com/3The classroom is located on the first floor; windows are oriented in

the southwest direction facing the inner courtyard. There is no line-of-sight connection to the base station. Approximate GPS coordinates: latitude48.196519, longitude 16.371119.

TABLE IOVERVIEW OF KPIS PROVIDED IN OUR OPEN DATASET.

File Quantity Unit

signal.csv RSRP,RSRP0,RSRP1 dBmRSSI,RSSI0,RSSI1 dBmRSRQ,RSRQ0,RSRQ1 dBSINR, SINR0, SINR1 dBpathloss dB

csi.csv Sample duration msWideband CQI –Requested throughput bit/sHistogram: rank 1 vs. rank 2 %

csi_subband.csv 13 subband CQIs (0–12) –

pla.csv sample duration msPRB utilization %TB average size bitTB maximum size bit

pla_prb.csv Utilization of each PRB (0–99) %

pla_mcs.json List of MCSs for codeword 0 –List of MCSs for codeword 1 –Histogram: (MCS0,MCS1) %

phy.csv Throughput Rphy, Rphy,0, Rphy,1 bit/sBLER,BLER0,BLER1 %Number of received TBs –Histogram: CFI ∈ {1, 2, 3} %

mac.csv MAC throughput Rmac bit/sNumber of received TBs –Number of new (1st attempt) TBs –TBs 1. retransmission %TBs 2. retransmission %TBs 3. and more retransmission %Residual BLER %

rlc.csv RLC throughput Rrlc bit/sRLC BLER %Number of received SDUs –

app.csv Application throughput Rapp bit/sProtocol string

PRBs. The eNodeB had two antennas and supported open-loop spatial multiplexing. The UE also has two antennas, thusallowing for 2×2 MIMO (multiple-input and multiple-output).

In LTE, the transmission time interval (TTI) is 1 ms. The re-porting period of most of the KPIs collected by the NEMO UEis approximately 500 ms. Only for some higher layers (RLC,application), the sampling period is longer—approximatelyone second. KPI reported at time ti represents an average overthe reporting period (ti−1, ti]. For some KPIs, a more detailedsummary in the form of a histogram is available.

III. THE DATASET: KEY PARAMETER INDICATORS

In cellular mobile networks, we witness a considerablevariability of measured KPIs, even in a static scenario. Fig. 1shows the raw samples and 15-minute averages of severalselected KPIs. Table I summarizes LTE DL KPIs providedin our dataset [1]. Each sample is stored with a correspondingUnix timestamp. In the following Sections III-A–III-G, wegive more details about the individual KPIs.

−99−98−97−96−95−94

RSRP/dBm

Raw samples 15-minute averages

−66−65−64−63−62−61

RSSI/dBm

−15−14−13−12−11−10

RSRQ/dB

3

6

9

12

15

18

SIN

R/dB

6789

101112

CQI

0

20

40

60

80100

PRB/%

1020304050607080

Rph/(M

bit/s)

5

8

11

14

17

20

BLER/%

16:0

018

:00

20:0

022

:00

9Fe

b2:

004:

006:

008:

0010

:00

12:0

014

:00

16:0

018

:00

20:0

022

:00

10Fe

b2:

004:

006:

008:

0010

:00

12:0

014

:00

16:0

018

:00

20:0

022

:00

11Fe

b2:

004:

006:

008:

0010

:00

12:0

014

:00

16:0

018

:00

20:0

022

:00

12Fe

b2.

004:

006:

008:

0010

:0010

203040506070

Rmac/(M

bit/s)

Fig. 1. Part of the KPIs collected during our static NEMO measurements in February 2019 in a classroom at the Institute of Telecommunications, TU Wien.

A. Cell-Specific Reference Signal

At predefined time-frequency coordinates, each cell broad-casts cell-specific reference signal (CRS) [17, Section 6.10.1],based on which UE measures and calculates the physical layerindicators: RSRP, RSSI, RSRQ, and SINR; see [5].

RSRP serves as the main indicator for cell selection duringthe initial connection and during handovers. RSSI measuresthe total wideband power. RSRQ—which is calculated fromRSRP and RSSI—depends on the cell load [18] and providesadditional information in the handover decision process [19].SINR is essential for the derivation of the CQI (Section III-B).

If multiple transmit and receive antennas are available, UEmay estimate RSRP, RSRQ, RSSI, and SINR for each antennaport. The final reported value is given by the maximum [5]among all ports. In our case, the UE has two receive antennas;there are thus three values available for each of the fourmentioned KPIs: One global (final reported) value, and onevalue for each of the receive antenna ports (0 and 1).

The dataset signal.csv also contains DL pathloss, whichUE estimates by comparing the received CRS power with thetransmit CRS power that is signaled by eNodeB [20, Section6.3.2, PDSCH-Config].

B. Channel State Information (CSI)

Based on the estimated channel matrix and measured SINR,the UE signals to the eNodeB CQI, rank indicator (RI), andprecoding matrix indicator (PMI).

The CQI value encodes a preferred code rate and modula-tion [21, Tab. 7.2.3-1 and Tab. 7.2.3-2]. The dataset csi.csvcontains a “requested throughput” field, which is calculatedfrom transport block (TB) sizes given by wideband CQIvalues, assuming that all PRBs would be scheduled to themeasuring UE and that BLER would be zero. CQI selection isnot standardized by 3GPP; it is vendor-specific. For example,under the same channel conditions, a better, more complexreceiver can report higher CQI [22, Section 2.3.6].

RI indicates the number of useful data streams. With tworeceive antennas, we have RI ∈ {1, 2}. RI chooses betweentransmit diversity (RI = 1) and spatial multiplexing (RI = 2).

PMI is transmitted in closed-loop spatial multiplexing andinforms eNodeB about the optimal precoding matrix the UEhas selected [22, Section 2.3.3]. In our case, the cell was usingopen-loop spatial multiplexing, PMI is thus not available.

Apart from a global “wideband” CQI, there are also multi-ple “subband” CQIs available (csi_subband.csv), whichallow eNodeB to schedule users on favorable resources [23,Section IV]. Fig. 2 shows a sequence (approximately 150 slong) of wideband and subband CQIs from our dataset.

C. Packet Link Adaptation (PLA)

After receiving CSI from all active UEs, the eNodeBperforms the radio resource management: For each user, itchooses specific PRBs, MCS, TB size. The table pla.csvcontains for each reporting period the average TB size, themaximum TB size, and average PRB utilization that themeasuring UE experienced.

5

6

7

8

9

10

11

Wid

eban

dC

QI

0 30 60 90 120 150 180 210 240 270 30002468

1012

Sample index

Subb

and

inde

x

0

5

10

15

Subb

and

CQ

I

Fig. 2. Example of 300 consecutive samples of the wideband CQI and 13subband-CQIs. The reporting period is approximately 500ms per sample.

0

20

40

60

80

100

PRB

utili

zatio

n%

0 30 60 90 120 150 180 210 240 270 3000

20

40

60

80

Sample index

PRB

inde

x

0

20

40

60

80

100

Util

izat

ion

%

Fig. 3. Overall PRB utilization (upper plot) and utilization of individual PRB-indexes (lower plot). For some PRB-indexes (0–3, 24–27, 48–51, 72–75), theaverage utilization never reaches 100%. A possible explanation is that certainPRBs are reserved for some narrowband services (e.g., narrowband IoT).

PRB utilization tells us what percentage of all eNodeB’sPRBs were scheduled to the measuring UE during the re-porting period. Remaining PRBs were scheduled for otherusers. The PRB utilization is also available for individual PRBindexes (pla_prb.csv). In our case of a 20 MHz channel,there are 100 PRB frequency-indexes (0–99). See Fig. 3.

Finally, for each reporting period, the file pla_mcs.jsoncontains a histogram of modulation and coding schemes thatwere applied by eNodeB. An example of a single reportingperiod is plotted in Fig. 4. In some cases, a different MCS wasused for each codeword. A NaN value of MCS 1 means thatonly one data stream was transmitted (i.e., rank = 1, transmitdiversity was used instead of MIMO).

D. Physical Channel Throughput

The file phy.csv contains the physical downlink sharedchannel (PDSCH) throughput Rphy, BLER, and the number ofreceived TBs within the reporting interval.

Throughput Rphy is calculated from all received TBs (in-cluding TBs with CRC4 failure). BLER measures the ratio

4Cyclic redundancy check.

––

24–

25–

26–

27–

1515

1616

1818

1919

2020

2121

2220

2222

2320

2322

2323

2421

2424

2525

2626

2727

5%

10%

15%

20%

MCS 0:MCS 1:

Perc

enta

geof

time No data received

Rank 1Rank 2

Fig. 4. Histogram of modulation and coding schemes of codewords 0 and 1during a single reporting period.

of the TBs with failed CRC to the number of all receivedTBs. Rphy and BLER are also available for each codewordseparately (Rphy,0, Rphy,1, BLER0, BLER1).

The control format indicator (CFI ∈ {1, 2, 3}) defines howmany OFDM5 symbols in each subframe are allocated forthe PDCCH.6 Higher CFI means fewer resources for thePDSCH. See [24, Section 5.3.4] for reference, and [25] forvisualization, in which a user can set different CFI values.

E. Medium Access Control (MAC) Layer ThroughputAccording to manual [16], MAC throughput Rmac is cal-

culated from successfully received service data units (SDUs);we thus have

Rmac ≈ (1− BLER)Rphy. (1)

Or, if the first option is not available, Rmac is estimated fromsuccessfully received TBs (slightly higher due to MAC headerand control data). In this case, (1) becomes equality.

Other available KPIs are the number of all successfullyreceived TBs in the reporting interval, excluding redundantretransmissions,7 and the number of successfully received newTBs. The “new TBs” do not include successfully receivedHARQ retransmissions that belong to failed TBs from oneof the previous reporting intervals.

HARQ retransmissions are documented by three fields:Ratio of first retransmissions to all received TBs, ratio ofsecond retransmissions to all received TBs, and ratio of third(and more) retransmissions to all received TBs.

Residual BLER is the percentage of TBs that HARQ wasnot able to correct. From 641 314 MAC layer samples in ourdataset, the residual BLER is nonzero in just 0.14% of cases(915 samples). The highest residual BLER reported in a singlereporting interval is 0.5%.

F. Radio Link Control (RLC) Layer ThroughputRLC throughput is calculated from SDUs that are success-

fully transferred through the RLC layer. The reporting intervalis longer than for the previous layers (ca. 1.2–1.3 s).

Since our DL traffic is generated by HTTP download (i.e.,TCP over IP traffic), the RLC works in acknowledged mode[26].8 Thus, RLC utilizes the ARQ procedure to correct

5Orthogonal frequency-division multiplexing.6Physical downlink control channel.7E.g., due to ACK message being corrupted on its way back to eNodeB.8For example, for voice over IP, RLC would apply unacknowledged mode.

20 30 40 50 60 700

0.2

0.4

0.6

0.8

1

R/(Mbit/s)

Em

piri

cal

CD

F

(a) Throughput distributionson different layersPhy.MACRLCApp.

0 2 4 6 8 10 12

Retransmissions %

(b) Distributions of HARQretransmission percentages

Retr. 1Retr. 2Retr. 3+

Fig. 5. Empirical CDFs of (a) throughput and (b) HARQ retransmissions.All time-series were resampled to 4 s bins before the CDF reconstruction.

residual erroneous TBs obtained from the MAC layer.

G. Application Layer Throughput

For the application layer, the reporting interval is evenlonger—approximately 1 to 3 seconds. For each sample, weobtain application protocol (in our case HTTP) and applicationthroughput, which does not include TCP/IP headers.

IV. OBSERVATIONS AND REMARKS

Fig. 1 from top to bottom concisely summarizes whathappens in LTE downlink. UE measures the radio channelbased on pilot signals (CRS) broadcasted by the eNodeB,to which it feeds back the observations in the form of CSIchosen such that the resulting BLER would be around 10%.Based on the UE feedback, the eNodeB chooses MCS, thenumber of parallel data streams (rank), in case of closed-loopspatial multiplexing, also a precoding matrix, and schedulesparticular resource blocks to our UE. All of this constitutesthe physical throughput Rphy, which is, however, reducedby transmission errors (BLER). Thus, approximately 10% ofresources have to be dedicated to HARQ retransmissions,which carry incremental redundancy that helps the MAC layerto decode the erroneous blocks.

Fig. 5 (a) presents cumulative distribution functions (CDFs)of throughput on different layers. MAC layer throughput is by≈ 10% smaller than the physical throughput due to blockerrors. In most of the cases, a single HARQ retransmission isenough to decode the erroneous blocks, see Fig. 5 (b). Theresidual MAC BLER is negligible; there is thus no visibledifference between the MAC throughput and the RLC through-put. Finally, the further reduction from RLC throughput toapplication layer throughput (without TCP/IP headers) is dueto header overhead.

A. Averaging of Logarithmic Quantities

In the case of logarithmic quantities s(dB), we should alwaysperform averaging in linear scale, when further postprocessingthe data (binning, smoothing, resampling):

s̄(dB) = 10 log

(1

N

N∑i=1

10

(s(dB)i /10

)). (2)

0 10 20 30 40 50 60 70 80 90 10030

40

50

60

70

t/s

Rap

p/(M

bit/s) Raw samples Resampled

Fig. 6. Volume-preserving throughput resampling. The plot shows the original(blue) and the resampled (red) application layer throughput samples.

B. Volume-Preserving Resampling of Rates

Because the duration of the reporting interval varies, wemight want to resample some time-series—e.g., to calculateCDFs. In the case of quantities that characterize a rate(throughput or BLER), we should take care that the resamplingis volume-preserving.

Volume-preserving resampling (or rebinning) is illustratedin Fig. 6. The area of each new, larger red bin is the sameas the area under the blue curve. The piecewise constantstaircase function emphasizes that each rate sample is notrelated only to a specific timestamp, but to the whole previousreporting interval. With a different kind of resampling (e.g.,linear interpolation), we would modify the total volume (totaldata volume in case of throughput, or the total number oferroneous blocks in case of BLER).

V. CONCLUSION

We have thoroughly described and partially visualized theprovided static indoor LTE DL dataset [1], which contains over600 000 KPI samples collected during 90 hours throughoutmultiple layers (physical, MAC, RLC, application). Despitethe lack of mobility, the dataset features rich variability andthus poses a nontrivial challenge for machine learning taskssuch as throughput prediction.

The dataset captures real-world end-user performance ex-perienced in an operational network, in a dense urban envi-ronment. It is thus particularly useful for researchers that aredeveloping data-driven models and cellular mobile networksimulators. There is a countless number of other possiblescenarios. We thus hope to see also other research groupspublishing detailed, high-quality open datasets in the future.

We believe that our dataset is useful also for educationalpurposes: It contains details about rank, MCS, transport blocksizes, CRC failures, and HARQ retransmissions; it can be usedto visualize PRB scheduling and subband CQIs. The studentsthus can explore the dataset to understand some of the LTEDL procedures better.

ACKNOWLEDGMENT

This work has been funded by the ITC, TU Wien. The research hasbeen cofinanced by the Austrian FFG, Bridge Project No. 871261.We thank Kei Cuevas for the proofreading.

REFERENCES[1] V. Raida, P. Svoboda, and M. Rupp. (2020) LTE DL static urban

indoor dataset. [Online]. Available: http://doi.org/10.23728/b2share.2367746915e34d1ca4bda84193f8056b

[2] Keysight Technologies. (2020) Nemo Handy handheld measurementsolution. [Online]. Available: https://www.keysight.com/en/pd-2767485-pn-NTH00000A/nemo-handy

[3] Android Developers. (2020) CellIdentityLte. [Online]. Available: https://developer.android.com/reference/android/telephony/CellIdentityLte

[4] ——. (2020) CellSignalStrengthLte. [Online]. Available:https://developer.android.com/reference/android/telephony/CellSignalStrengthLte

[5] 3GPP, “Physical layer; Measurements,” TS 36.214, 2020, v. 15.5.0.[6] A. Bokani, M. Hassan, S. S. Kanhere, J. Yao, and G. Zhong, “Com-

prehensive mobile bandwidth traces from vehicular networks,” in Pro-ceedings of the 7th International Conference on Multimedia Systems.Association for Computing Machinery, 2016.

[7] A. Bokani. (2016) GitHub – aubokani/Bandwidth-Dataset: Mobilenetwork bandwidth traces. [Online]. Available: https://github.com/aubokani/Bandwidth-Dataset

[8] J. van der Hooft, S. Petrangeli, T. Wauters, R. Huysegems, P. R. Alface,T. Bostoen, and F. De Turck, “HTTP/2-Based Adaptive Streaming ofHEVC Video Over 4G/LTE Networks,” IEEE Communications Letters,vol. 20, no. 11, pp. 2177–2180, 2016.

[9] J. van der Hooft et al. (2016) 4G/LTE bandwidth logs. [Online].Available: https://users.ugent.be/~jvdrhoof/dataset-4g/

[10] V. Raida, P. Svoboda, and M. Rupp, “Modified dynamic time warpingwith a reference path for alignment of repeated drive-tests,” in 2020IEEE 92nd Veh. Technol. Conf., Oct. 2020, submitted for review.

[11] V. Raida et al. (2020) RP-DTW data set. [Online]. Available:http://doi.org/10.23728/b2share.b7f93ada6c6f46b296cd3c39665982c4

[12] B. Meixner, J. W. Kleinrouweler, and P. Cesar, “4g/lte channel qualityreference signal trace data set,” in Proceedings of the 9th ACM Mul-timedia Systems Conference. Association for Computing Machinery,2018, p. 387–392.

[13] B. Meixner et al. (2018) 4G/LTE channel quality reference signaltraces data set | Zenodo. [Online]. Available: https://doi.org/10.5281/zenodo.1220256

[14] D. Raca, J. J. Quinlan, A. H. Zahran, and C. J. Sreenan, “Beyondthroughput: A 4g lte dataset with channel and context metrics,” in Pro-ceedings of the 9th ACM Multimedia Systems Conference. Associationfor Computing Machinery, 2018, p. 460–465.

[15] D. Raca et al. (2018) Beyond throughput: a 4G LTE datasetwith channel and context metrics | Zenodo. [Online]. Available:https://doi.org/10.5281/zenodo.1219679

[16] Keysight Technologies, “Nemo file format specification,” Manual, 2018,version 2.37.

[17] 3GPP, “Physical channels and modulation,” TS 36.211, 2020, v. 15.9.0.[18] V. Raida, M. Lerch, P. Svoboda, and M. Rupp, “Deriving cell load

from RSRQ measurements,” in 2018 Network Traffic Measurement andAnalysis Conference (TMA), 2018.

[19] J. Kurjenniemi, T. Henttonen, and J. Kaikkonen, “Suitability of RSRQmeasurement for quality based inter-frequency handover in lte,” in 2008IEEE International Symposium on Wireless Communication Systems,2008, pp. 703–707.

[20] 3GPP, “Radio resource control (RRC); Protocol specification,” TS36.331, 2020, v. 15.9.0.

[21] ——, “Physical layer procedures,” TS 36.213, 2020, v. 15.9.0.[22] J. C. Ikuno, UMTS Long-Term Evolution. John Wiley & Sons, Ltd,

2011, ch. 2, pp. 29–53. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/9781119954705.ch2

[23] S. Schwarz, C. Mehlführer, and M. Rupp, “Calculation of the spatialpreprocessing and link adaption feedback for 3GPP UMTS/LTE,” in2010 Wireless Advanced 2010, 2010.

[24] 3GPP, “Multiplexing and channel coding,” TS 36.212, 2020, v. 15.9.0.[25] L. Guillemot. (2020) LTE resource grid. [Online]. Available: https:

//www.sqimway.com/lte_resource_grid.html[26] 3GPP, “Radio link control (RLC) protocol specification,” TS 36.322,

2019, v. 15.3.0.