Phy Abstraction for LTE

Physical layer abstraction
for
LTE downlink
PRESENTED BY
RAJ PATEL

Introduction
link level simulator simulates a single radio link
system level simulator takes into account
a complete cell: time consuming
Physical layer abstraction : process of modeling
the performance of the physical layer based on
the current channel state
and the physical layer parameters

Introduction
AWGN
MCS -> CQI
target SNR – 10% BLER
Plots : Target SNR vs CQI / MCS - linear

Introduction
Extrapolation of Reference curve to get effective SNR
choose MCS values belonging to same constellation.
Get the Target SNR value
•Calc. difference between the T.SNR values
We note down the effective code rate for the MCS used.
We use the reference curves to get the values of SNR
using the effective code rate of that MCS
•Calc. the difference between the SNR values

Observations
otheoretical difference and the difference calculated using interpolation are not the same
oPossible reason: C* = (TBS + CRC) / G. G: bits transmitted per second; C: Code Rate
o 40 <= Code Block Size(= TBS + CRC) <= 6144 ; CRC = 24 bits
oEg: 6126 bits TBC
6120 + 24 // 6 + 24 + 10 ; 10 : padding
Delta SNR from
Lookup table values
C = TBS / G
4.237 4.3203 1.4398 2.8805 4.7258 6.6409 2.6672 3.9737
Delta SNR from look
up table using
C* = (TBS + CRC) / G
4.1689 4.3423 1.4366 2.9057 4.7415 6.684 2.6877 3.9963
Delta SNR from log
BLER curve
2.86 3.446 0.788 2.668 4.2 3.742 2.412 2.33

Frequency Selective Fading
Coherence Bandwidth
Signal Bandwidth
Flat fading: Just attenuation, no distortion
Frequency Selective (much more realistic): Distortion
If the attenuation happens in different amounts for the different parts of the signal, it is a
distortion.
Condition: Coherence Bandwidth < Signal Bandwidth
Frequency selective fading channel model
Eg.: EPA

EPA : Extended Pedestrian A model
omultiple paths
osame signal copies arrive at the receiver
delayed and different attenuations
o-g E –M1 –R1 –N 100 –n 10000
o-M1: Abstraction flag
keeps channel coefficients constant over SNR range
o-R1: to reduce simulation time
o-g E: fading model
o-n: number of packets
o-N: number of channel realizations
oOUTPUT format:
SNR, 50 channel coefficients, BLER1

Abstraction Techniques
EESM
MIESM

EESM: Exponential Effective SINR Mapping
훾eff = 훽1 퐼−1 1
푁
푁 퐼
푛=1
훾푛
훽2
퐼 훾푛 = 1 − exp (−훾푛) ; 훾푛 is the instantaneous SNR
Aim: to calculate SINR effective
Noise_var = 1 / SNR_linear; inst_snr = 10*log10 (h^2/Noise_var);
1. Calculate the instantaneous SNR corresponding to each value of channel realization
2. Use the I function with the instantaneous SNR and average it over N
3. Use the inverse function of I to calculate the effective SNR

MIESM
Mutual Information Effective SINR Mapping
No closed form expression
Calculate the instantaneous SNR
Using lookup tables, calculate normalized capacity for each instantaneous SNR
Calculate average normalized capacity per SNR
Calculate the effective SNR using average normalized capacity with lookup table

MSE calculation
훾eff = 퐼−1 1
푁
푁 퐼(훾푛)
푛=1
*N stands for the number of values
of channel coefficients per SNR.
SNR interp: image of SNR effective on AWGN curve
푀푆퐸 =
1
푁
푁
푛=1
훾푖푛푡푒푟푝 BLER푐ℎ −훾eff
훾푖푛푡푒푟푝 BLER푐ℎ
2
*N here, stands for the number of SNR values.

MSE results
MCS MSE EESM using
'linear','extrap'
NORMALIZED
Linear, log
MSE_MIESM
'linear','extrap'
NORMALIZED
Linear, log
3 58.695, 0.3663 108.92, 0.2975
15 1.5247, 0.4958 0.3202, 0.3395
15 _n = 1000, N =1000 1.1699, 1.3596 0.3403, 1.9242
20 * 0.3869, 0.2304 0.1067, 0.5900
23 0.2551, 0.4954 0.0823, 0.3636
25 0.0897, 0.7444 0.0672, 0.7858

MSE –With 훽1, 훽2
훾eff = 훽1 퐼−1 1
푁
푁 퐼
푛=1
훾푛
훽2
푀푆퐸argmin
훽1,훽2
=
1
푁
푁
푛=1
훾푖푛푡푒푟푝 BLER푐ℎ −훾eff 훽1,훽2
2

MSE Results –With 훽1, 훽2
MCS B values MSE EESM
calibrated
3 [0.0334,0.6226] 0.7683
15 [3.975e+02,4.7833e+03] 0.0037
15 _n = 1000, N
[3.991e+02,5.581e+03] 0.0041
=1000
20 (erroneous) [41.3997,58.1240] 0.0466
23 [6.862e+02,1.241e+04] 1.64e-04
25 [7.469e+02,1.318e+04] 1.20e-04
MCS B values MSE MIESM
calibrated
3 [0.2051,17.348] 0.9835
15 [0.7490,0.6111] 0.2887
15 _n = 1000, N
[0.7903,0.7440] 0.3339
=1000
20 (erroneous) [0.6041,0.7456] 0.0430
23 [0.8813,0.7282] 0.0567
25 [0.8398,0.8028] 0.0645

EESM – calib.
MCS- color
3-Red, 15- Yellow, 20*- Sky blue,
23- Blue, 25- Pink

Conclusions and Observations
Calibration factors work better with EESM
The resultant MSE after using calibration factor with EESM are around 10^3 times better
Where as for MIESM, it is 10 times better.
MCS 25: EESM MIESM
MSE Without calibration 0.7444 0.7858
MSE With calibration 1.20e-04 0.0645

Calculations done in the log scale don’t make
푀푆퐸argmin
훽1,훽2
=
1
푁
푁
푛=1
2
Division in log scale?
MCS MSE EESM using
'linear','extrap'
NORMALIZED
Linear, log
MSE_MIESM
'linear','extrap'
NORMALIZED
Linear, log
3 58.695, 0.3663 108.92, 0.2975
15 1.5247,0.4958 0.3202, 0.3395
20 (erroneous) 0.3869, 0.2304 0.1067, 0.5900
23 0.2551, 0.4954 0.0823, 0.3636
25 0.0897, 0.7444 0.0672, 0.7858
NOTE: Calculations in Linear scale show a gradual
Decrease in MSE value, unlike the log scale
Thus operate with linear values
if we are using Normalization
But why does Lower MCS have weird
MSE values?

Issues with the lower MCS values any ideas??
Working on Linear scale, why is it that the Lower MCS has higher values of MSE compared to
higher MCS values?
Reason: Normalization while calculating MSE
푀푆퐸argmin
훽1,훽2
=
1
푁
푁
푛=1
2
훾푖푛푡푒푟푝 BLER푐ℎ − 훾eff 훽1, 훽2 : more or less remains the same, say around 5-10 dB
But, 훾푖푛푡푒푟푝 BLER푐ℎ changes according to MCS value, stays close to -2 to 2 dB

MCS MSE EESM using
'linear','extrap'
NORMALIZED
Linear
MSE_MIESM
'linear','extrap'
NORMALIZED
Linear
3 58.695 108.92
15 1.5247 0.3202
15 _n = 1000, N =1000 1.1699 0.3403
20 (erroneous) 0.3869 0.1067
23 0.2551 0.0823
25 0.0897 0.0672
Table with the calculations done in Linear scale.

For 15 _n = 1000, N =1000 case, the calculations are not in synchronization with the other cases.
Reason: too many values: may be it gives us a better estimate.
calibrated
3 [0.0334,0.6226] 0.7683
15 [3.975e+02,4.7833e+03] 0.0037
15 _n = 1000, N
[3.991e+02,5.581e+03] 0.0041
=1000
20 (erroneous) [41.3997,58.1240] 0.0466
23 [6.862e+02,1.241e+04] 1.64e-04
25 [7.469e+02,1.318e+04] 1.20e-04
NOTE: Calculations in Linear scale
show a gradual Decrease in MCS value
calibrated
3 [0.2051,17.348] 0.9835
15 [0.7490,0.6111] 0.2887
15 _n = 1000, N
[0.7903,0.7440] 0.3339
=1000
20 (erroneous) [0.6041,0.7456] 0.0430
23 [0.8813,0.7282] 0.0567
25 [0.8398,0.8028] 0.0645
Note: The MSE of EESM is lower than the MSE of MIESM

Note: The MSE of EESM is lower than the MSE of MIESM
Reason? High values of Beta using EESM?
calibrated
3 [0.0334,0.6226] 0.7683
15 [3.975e+02,4.7833e+03] 0.0037
15 _n = 1000, N
[3.991e+02,5.581e+03] 0.0041
=1000
20 (erroneous) [41.3997,58.1240] 0.0466
23 [6.862e+02,1.241e+04] 1.64e-04
25 [7.469e+02,1.318e+04] 1.20e-04
calibrated
3 [0.2051,17.348] 0.9835
15 [0.7490,0.6111] 0.2887
15 _n = 1000, N
[0.7903,0.7440] 0.3339
=1000
20 (erroneous) [0.6041,0.7456] 0.0430
23 [0.8813,0.7282] 0.0567
25 [0.8398,0.8028] 0.0645

Issues and Future Work
The calibration factors are a bit high for some MCS values for EESM!
WHY!?
Is that the only reason why we see the performance of EESM is better than MIESM??

LTE
OFDM
OFDMA
Cyclic Prefix
ISI
RE
RB

OAI
Eurecom
Physical layer stimulations

Resource Elements Allocation
•N_PILOTS = 6*N_RB*TM
•N_RB - by default set to 25
•N_RE = (OFDM symbols – Prefix length) * (N_RB*sub-carriers per block) - N_PILOTS
•Example: -x1 –y1 –z1 ; Normal cyclic prefix
•N_RE= (14-1)*(25*12) – (6*25*1) = 3750

Map CQI --> MCS
•CQI – feedback
•MCS – chosen
CQI (1-15) MCS(1-28)
3 3
8 15
10 20
13 25 (with extended prefix)

AWGN reference curves
•BLER vs SNR plots
•Monte Carlo stimulations
•Step size
•SNR range
•Interpret .csv
•Target SNR

Plots
•Target SNR vs CQI
•Target SNR vs MCS
•Target SNR vs Code rate
•Observation

Extrapolation of curves
•ΔSNR (db) = f -1(r2) – f-1(r1)
•Normalized capacity is the
effective code rate
•Code rate/ bits per symbol

Extrapolation method
•Choose MCS values belonging to same constellation.
•Stimulate for those MCS values and get the Target SNR value. Target SNR is the SNR value for log
BLER= -1
•ΔSNR value of two MCS schemes from stimulation
•We note down the effective code rate for the MCS used.
•We use the reference curves to get the values of SNR using the appropriate curve (taking into
consideration the Modulation scheme used for that MCS)
•ΔSNR values found from the reference curves by extrapolation

Conclusions
•Extrapolation important
•Needs to be improved

Phy Abstraction for LTE

More Related Content

What's hot (18)

Viewers also liked (11)

Similar to Phy Abstraction for LTE (20)

Recently uploaded (20)

Phy Abstraction for LTE