IMTAphy: An IMT-Advanced Spatial Channel Model Implementation and PHY-Abstraction for System-Level Simulations
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Technische Universität München
Lehrstuhl für Kommunikationsnetze
Prof. Dr.-Ing. J. Eberspächer
Simulation and Prototyping Workshop
RWTH Aachen, July 14, 2011
IMTAphy: An IMT-Advanced Spatial Channel
Model Implementation and PHY-Abstraction for
System-Level Simulations
Jan Ellenbeck
jan.ellenbeck@tum.de
Motivation 292 3G Evolution: HSPA and LTE for Mobile Broadband
Time-frequency fading,
•
user # 1
LTE, WiMAX, and future IMT-A cellular systems exploit Time-frequency fading,
user # 2
– time/frequency/spatial selectivity of the channel by
fast scheduling with link adaptation
need multi-path / fast-fading MIMO channel model
Source: Dahlman et al.
• Our general research topic: interference management (IM)
– IM restricts scheduler‘s degrees of freedom User # 1 scheduled
User # 2 scheduled
(frequency reuse, usage patterns in time, available beamforming / precoding) 1m
s
to judge tradeoff of IM schemes: fast-fading spatial channel model needed
Tim
e 180 kHz
ncy
Freque
•
Figure 14.1 Downlink channel-dependent scheduling in time and frequency domains.
ITU-R M.2135 mandates system-level simulations to evaluate mean cell / cell-edge
spectral efficiency of IMT-A candidate systems 14.2.1 Downlink scheduling
To support downlink scheduling, a terminal may provide the network with
– need an efficient implementation channel-status reports indicating the instantaneous downlink channel quality in
both the time and frequency domain. The channel status can, for example, be
– that gives accurate / reliable results obtained by measuring on a reference signal transmitted on the downlink and
used also for demodulation purposes. Based on the channel-status report, the
downlink scheduler can assign resources for downlink transmission to different
mobile terminals, taking the channel quality into account in the scheduling deci-
C++ implementation of the M.2135 channel model available under GPLv3 license at sion. In principle, a scheduled terminal can be assigned an arbitrary combination
of 180 kHz wide resource blocks in each 1 ms scheduling interval.
https://launchpad.net/imtaphy
14.2.2 Uplink scheduling
PHY abstraction (MIMO receivers, eff. SINR, BLERs) and simple LTE implementation
The LTE uplink is based on orthogonal separation of different uplink transmis-
sions and it is the task of the uplink scheduler to assign resources in both time
will be available shortly and frequency domain (combined TDMA/FDMA) to different mobile terminals.
Scheduling decisions, taken once per 1 ms, control what set of mobile terminals
Jan Ellenbeck Simulation and Prototyping Workshop 2
Technische Universität München RWTH Aachen, July 14, 2011WINNER II D1.1.2 V1.2
ITU-R IMT-A Channel Model (Primary Module)
• Adapted from WINNER+ channel model
• Link-level and system-level simulations possible MS
BS
link
segments FRS
•
v
Large Scale effects: MS
"1
"2
$"
$!
#1
!1
MS
#2 !2
– pathloss
BS
– outdoor shadowing Figure 3-3. System level approach, several drops.
– penetration shadowing A single link model is shown in Figure 3-4. The parameters used in the models are also shown in the
Outdoor-to-Vehicle (O2V), Outdoor-to-Indoor (O2I)
figure. Each circle with several dots represents scattering region causing one cluster. The number of
clusters varies from scenario to another.
– antenna patterns
– handover margin N
v
– feeder loss
$!
• Small Scale effects (due to multi-path) : N
#1 % MS
$"
– geometry based stochastic model "1
– channel impulse response (CIR) with proper "2 !1
MS
#2 !2
• angular power distribution among paths BS
• power delay profile
• phase relationships between antenna array elements
– correlations between: Figure 3-4. Single link.
• antenna elements Source: IMT-Advanced Evaluation
In spatial channel model the performance of the single link is defined by small-scale parameters of all
Guidelines
MPCs between two spatial positions ITU-R
of radio-stations. M.2135
According to this, if only one station is mobile
• mobiles associated to same base station (MS), its position in space-time is defining a single link. The more complex network topology also
includes multihop links [Bap04] and cooperative relaying [Lan02], however more complex peer-to-peer
connections could be easily described as collections of direct radio-links.
Large-Scale Parameters (LSP) are used as control parameters , when generating the small-scale channel
Jan Ellenbeck Simulation and Prototyping Workshop
parameters. 3 of the single
If we are analyzing multiple positions of MS (many MSs or multiple positions
MS) we have a multiple-link model for system level simulations. It can be noted that different MSs being
Technische Universität München RWTH Aachen, Julyat 14, 2011
the same spatial position will experience same LSP parameters.
For multi-link simulations some reference coordinate system has to be established in which positions and
movement of radio-stations can be described. A term network layout is designating complete descriptionWraparound Simulation for Hexagonal Scenarios
!!" !&" +" *" )" ("
%"
Problem: #" !" !!" !&" +" *"
!#" '" %"
• M2135: explicit interference modeling
!*" !$" #&" !%" #" !"
• Same (full) protocol stack in all nodes $" !#" '"
for system-level simulation )" (" !(" !'" !*" !$" #&" !%"
!+" #!" $"
(complexity!)
!!" !&" +" *" )" (" !(" !'"
• But realistic interference situation only %" !)" !+" #!"
in inner cells #" !" !!" !&" +" *" )" ("
!#" '" %" !)"
!*" !$" #&" !%" #" !" !!" !&" +" *"
$" !#" '" %"
!(" !'" !*" !$" #&" !%" #" !"
Solution: wraparound !+" #!" $" !#" '"
• makes cells at opposing ends )" (" !(" !'" !*" !$" #&" !%"
!)" !+" #!" $"
neighbors (like on a torus) !!" !&" +" *" )" (" !(" !'"
%" !)" !+" #!"
• here: per link shift mobile to all
possible wraparound positions and #" !" !!" !&" +" *"
!#" '" %" !)"
choose closest !*" !$" #&" !%" #" !"
$" !#" '"
!(" !'" !*" !$" #&" !%"
All nodes give realistic statistics !+" #!" $"
!(" !'"
!)" !+" #!"
Jan Ellenbeck Simulation and Prototyping Workshop 4
Technische Universität München RWTH Aachen, July 14, 2011! !"#$%%&'()!%*$+,-.! --!
Channel Coefficient"#$%&'!(!
Generation Procedure
/0122"3%45"66747"28%9"2":18752%#:54";MS
BS
link
segments
Step 10: Generate Channel Coefficients
FRS v
�T � ��
$!
M �
MS $"
#1
√ �
� �
"1 !1
vv
κ−1exp(jΦvh
"2
Frx,u,V (ϕn,m ) √ exp(jΦn,m ) n,m ) Ftx,s,V (φn,m )
#2 !2 MS
BS
hu,s,n (t) = Pn
Frx,u,H (ϕn,m )
m=1
κ−1 exp(jΦhv
n,m ) exp(jΦhh
n,m )
Ftx,s,H (φn,m )
Figure 3-3. System level approach, several drops.
exp(jds 2πλ−1 −1
o sin(φn,m )) exp(jdu 2πλo sin(ϕn,m ))
A single link model is shown in Figure 3-4. The parameters used in the models are also shown in the
figure. Each circle with several dots represents scattering region causing one cluster. The number of exp(j2πvn,m t)
clusters varies from scenario to another.
Hu,s,n (t) channel coefficient
u number of Rx antennas
s number of Tx antennas
N
v n = 1 . . . 24 cluster (path) index
m = 1 . . . 20 ray index
N
$! τn delay of path n
#1 % MS Pn power of path n
$"
Frx,u,V (ϕn,m ) electrical field pattern
"1 Frx,u,H (ϕn,m ) for Rx/Tx antennas with
"2 !1 Ftx,s,V (φn,m ) vertical / horizontal
MS
#2 !2 Ftx,s,H (φn,m polarization
BS Φvv vh
n,m , Φn,m random phases
Φhv hh
n,m , Φn,m
κ cross polarization power ratio
ds , effective distance of Tx / Rx
Figure 3-4. Single link. du elements s/u to reference element
In spatial channel model the performance of the single link is defined by small-scale parameters of all φn,m angle of departure
MPCs between two spatial positions of radio-stations. According to this, if only one station is mobile ϕn,m angle of arrival
(MS), its position in space-time is defining a single link. The more complex network topology also vn,m Doppler frequency component
Source:
includes Rep.
multihop linksITU-R
[Bap04] M.2135
and cooperative relaying [Lan02], however more complex peer-to-peer
connections could be easily described as collections of direct radio-links.
vn,m = ||v|| cos(ϕn,m − θv )/λ0
Large-Scale Parameters (LSP) are used as control parameters , when generating the small-scale channel
parameters.
JanIfEllenbeck
we are analyzing multiple positions of MS (many MSs or multiple
Simulation and positions of the single
Prototyping Workshop 6
MS) we have a multiple-link model for system level simulations. It can be noted that different MSs being
Technische Universität München
at the same spatial position will experience same LSP parameters.
RWTH Aachen, July 14, 2011
For multi-link simulations some reference coordinate system has to be established in which positions and
movement of radio-stations can be described. A term network layout is designating complete descriptionChallenge: Complexity
Typical scenario:
• 57 cells
• 10 mobiles per cell
57 * 57 * 10 = 32.490 links
• 4 x 4 (Tx x Rx) antenna configuration
16 antenna pairs per link
• Up to n = 1..24 paths (clusters) between an antenna pair
32.490 * 16 * 24 = 12.476.160 channel coefficients to compute (for one t)
• Each path consists of m = 1..20 rays (+ LoS), so summation goes over
249.423.200 rays in total
For each time instant t, 250 million values are computed and summed up!
We need a very efficient implementation
to simulate over a significant time span
Jan Ellenbeck Simulation and Prototyping Workshop 7
Technische Universität München RWTH Aachen, July 14, 2011Key Observation: Most Things are time-invariant!
� �M � �T � √ �� �
Frx,u,V (ϕn,m ) √ exp(jΦvvn,m ) κ−1 exp(jΦvh
n,m ) Ftx,s,V (φn,m )
hu,s,n (t) = Pn −1 exp(jΦhv )
m=1
Frx,u,H (ϕn,m ) κ n,m exp(jΦhh
n,m )
Ftx,s,H (φn,m )
exp(jds 2πλ−1 −1
o sin(φn,m )) exp(jdu 2πλo sin(ϕn,m ))
• Between subsequent time instances, only the red term changes: exp(j2πvn,m t)
• The rest (green part) only has to be computed once
Implementation:
1. generate all input parameters with proper correlations
2. pre-compute green part and the coefficient of the red part (without t)
3. store for the rest of the simulation (size: e.g. 250 million values)
store with single precision in large vectors to exploit vectorization/caches in CPU
To compute CIR H for a new time instant t:
1. multiply coefficient with current t
2. take complex exponential of (1) (specialized library functions, e.g. Intel MKL vzCIS)
3. multiply constant (green part) with result (2) and sum over all rays (dot product)
Jan Ellenbeck Simulation and Prototyping Workshop 8
Technische Universität München RWTH Aachen, July 14, 2011Performance Comparison
Computing the CIR for 32,490 links, 4x4 MIMO, 100 TTIs
Hardware: 8 cores Intel X5460 @ 3.16GHz (from 2008)
Implementation / Total Time for Max. mem.
(real) each
IMTAphy mode runtime additional consumption
TTI
TTA PG707 [1] 31000 310 sec. 23 GB
sec.
WINNER [2] 1962 sec. 18 sec. 27 GB
with C-MEX
IMTAphy 1 thread 1015 sec 7.73 sec
double 8 threads 501 sec 4.40 sec 6.4 GB
IMTAphy 1 thread 728 sec 4.19 sec
single 8 threads 275 sec 2.21 sec 3.2 GB
[1] TTA PG707: C-Code for Report ITU-R M.2135 Channel Model Implementation http://www.itu.int/oth/R0A06000024/en
[2] Finland: Software Implementation of IMT.EVAL Channel Model http://www.itu.int/oth/R0A06000022/en
Jan Ellenbeck Simulation and Prototyping Workshop 9
Technische Universität München RWTH Aachen, July 14, 2011Pathgain Calibration UMa
1.0
Org. 1
Org. 2
0.8 Org. 3
Org. 4
Org. 5
P r(X < abscissa)
0.6 Org. 6
Org. 7
IMTAphy
0.4
0.2
0.0
−140 −120 −100 −80 −60 −40
Pathgain [dB] Reference results WINNER project IMT-A evaluation
Jan Ellenbeck Simulation and Prototyping Workshop 10
Technische Universität München RWTH Aachen, July 14, 2011SINR Calibration UMa
1.0
Org. 1
Org. 2
0.8 Org. 3
Org. 4
Org. 5
P r(X < abscissa)
0.6 Org. 6
Org. 7
IMTAphy
0.4
0.2
0.0
−10 −5 0 5 10 15 20 25
Wideband SINR [dB] Reference results WINNER project IMT-A evaluation
Jan Ellenbeck Simulation and Prototyping Workshop 11
Technische Universität München RWTH Aachen, July 14, 2011CDF of Angular Spread at BS (AoD, UMa)
1.0
LoS
0.8
NLoS
P(X ≤ abscissa)
0.6
0.4 WINNER Org. 1
WINNER Org. 2
WINNER Org. 3
0.2 CATR
CMCC
IMTAphy
0.0
0 20 40 60 80 100
AoD (degrees)
Reference results WINNER project IMT-A evaluation
Jan Ellenbeck Simulation and Prototyping Workshop 12
Technische Universität München RWTH Aachen, July 14, 2011CDF of Delay Spread UMa scenario
1.0
WINNER Org. 1
WINNER Org. 2
0.8 WINNER Org. 3
CATR
CMCC
P(X ≤ abscissa)
0.6 IMTAphy
0.4
LoS NLoS
0.2
0.0 −3
10 10−2 10−1 100 101
Delay µsec. Reference results WINNER project IMT-A evaluation
Jan Ellenbeck Simulation and Prototyping Workshop 13
Technische Universität München RWTH Aachen, July 14, 2011Temporal Autocorrelation (UMa scenario)
1.0
IMTAphy
0.8 WINNER ITU
SCM
0.6 Classical Doppler
Autocorrelation
0.4
0.2
0.0
−0.2
−0.4
0 1 2 3 4 5 6 7 8
Distance / Wavelength (d/λ)
Jan Ellenbeck Simulation and Prototyping Workshop 14
Technische Universität München RWTH Aachen, July 14, 20114x4 MIMO Capacity at 14 dB SINR (UMa)
� ρ �
H
C = log2 det I + HH
S
1.0
Gaussian i.i.d.
IMTAphy LoS
0.8 IMTAphy NLoS
WINNER LoS
WINNER NLoS
P(X > abscissa)
0.6 SCM LoS
Reference values*
0.4
0.2
0.0
6 8 10 12 14 16 18 20
Capacity [bits/s/Hz]
* Chong et al. Evolution trends of wireless MIMO channel modeling towards IMT-Advanced, IEICE Trans. on Comm., vol. 92, no. 9, 2009.
Jan Ellenbeck Simulation and Prototyping Workshop 15
Technische Universität München RWTH Aachen, July 14, 2011Eigenvalue Distribution (UMa NLoS)
CDF of Eigen Values UMa (NLos)
1
!1 SCM
! SCM
0.9 2
!3 SCM
! SCM
0.8 4
!1 Winner
!2 Winner
0.7
!3 Winner
Pr (Power < abscissa)
!4 Winner
0.6
!1 IMTA Phy
!2 IMTA Phy
0.5
!3 IMTA Phy
! IMTA Phy
0.4 4
0.3
0.2
0.1
0
−80 −70 −60 −50 −40 −30 −20 −10 0 10 20
Power [dB]
Jan Ellenbeck Simulation and Prototyping Workshop 16
Technische Universität München RWTH Aachen, July 14, 2011Summary M.2135 Spatial Channel Model
• ITU’s IMT-Advanced channel model is complex to implement
• Efficient implementations needed for system-level simulation
• Simulators need to be calibrated
• Complete C++ source code together with simulation scenarios and
MATLAB evaluation scripts available online under GPL license at:
https://launchpad.net/imtaphy
Reproducible simulation results
Jan Ellenbeck Simulation and Prototyping Workshop 17
Technische Universität München RWTH Aachen, July 14, 2011Current LTE Simulator implementation (DL only)
• Phy Abstraction:
– MIMO (MMSE, MRC) receiver performance model
– MI-effective SINR computation model (802.16m EMD)
– Block Error Rate modeling (link level simulations using TU Vienna simulator)
• LTE (Release 8) feedback computation:
– Derive CQI values from estimated SINRs taking out-of-cell interfering
transmissions into account
– Exhaustive search over closed-loop spatial multiplexing codebook to determine
PMI and Rank indicators
• Hybrid-ARQ retransmission support
– Timing and feedback support (magic)
– Soft-combining support (currently Chase Combining only)
• Scheduler
– Currently Round-Robin frequency-nonselective for calibration
– MU MIMO / spatial multiplexing also supported by IMTAphy
• RLC functionality taken from http:/launchpad.net/openwns-lte
Jan Ellenbeck Simulation and Prototyping Workshop 18
Technische Universität München RWTH Aachen, July 14, 2011Parallelization ,QWHO0DWK.HUQHO/
• Main simulation remains single threaded
• Performance critical parts are parallelized 3URGXFW%ULHI 7KH)ODJVK
,QWHO0DWK.HUQHO/LEUDU\ /LEUDU\IRU
• Time evolution of channel model: ,QWHO0DWK.HU
H[WHQVLYHO\WKU
– OpenMP #pragma omp parallel for IRUVFLHQFHHQ
– Intel Math Kernel Library / Vector Math Library
0DWKHPDWLFDO
• Evaluation of transmissions is parallelized: 'HQVH/LQHDU$OJ
6SDUVH/LQHDU$OJ
– Each TTI, all transmissions register at single entity 6ROYHU,WHUDWLYH6
– Afterwards they can be evaluated independently in parallel )DVW)RXULHU7UDQ
2SWLPL]HG/,13$
– The results (SINRs) are then serially fed to each Rx node 9HFWRU0DWK/LEU
6WDWLVWLFV)XQFWLR
&OXVWHU6XSSRUW³
´,QWHO0./LVLQGLVSHQVDEOH
Parallel Region Register
Parallel RegionIRUDQ\KLJKSHUIRUPDQFH
Notify
Transmissions FRPSXWHUXVHURQ[
Receivers
SODWIRUPVµ
3URI-DFN'RQJDUUD
Single-Thread Evolve Channel Evaluate Transmissions
,QQRYDWLYH&RPSXWLQJ/DE
8QLYHUVLW\RI7HQQHVVHH.QR[YLOOH
Discrete-Time
´2YHURIWKH7RS
Event-based VXSHUFRPSXWHUVXWLOL]H
Simulation ,QWHOSURFHVVRUV,QWHO0./
LQFOXGHVIHDWXUHVDQGKHOSV
XQORFNWKHSHUIRUPDQFH
Jan Ellenbeck Simulation and Prototyping Workshop 19
LQ,QWHO·VODWHVWJHQHUDWLRQ
Technische Universität München RWTH Aachen, July 14, 2011
RI&38·Vµ
6KDQH6WRU\
,QWHO0./(QJLQHHULQJ0DQDJHU6 Simulator calibration
Initial System-Level Calibration Results
Step 2 results: DL user throughput
! Downlink Calibration
100
1x2
90
80
70
60
C.D.F . [%]
50
40
InH
30 UMi
20 UMa
RMa
10 Case 1 3D
Case 1 2D
0
0 0, 1 0,2 0,3 0,4 0,5
Normalized User Throughput [bps/Hz]
Ref: 3GPP TR36.814 ver 1.5.0 Averaged over 16 sources
©
© 3GPP
3GPP 2009
2009
3GPP TR 36.814 (UMa, 1x2 MRC)
Mobile
Our (current) results (UMa, 1x2 MRC)
44
Cell spectral efficiency: 1.0 Bit/s/Hz ± 8% Cell spectral efficiency: 1.09 Bit/s/Hz
Cell edge user : 0.022 Bit/s/Hz ± 17% Cell edge user : 0.0154 Bit/s/Hz
(5%-tile user throughput CDF) (5%-tile user throughput CDF)
Jan Ellenbeck Simulation and Prototyping Workshop 20
Technische Universität München RWTH Aachen, July 14, 2011Simulator Benchmarking Scenario 1: about 10s (10000 TTIs) simulation time per 1 hour wall clock time • 57 cells, 10 users per cell, explicit interference modeling • 10 MHz (50 PRBs) • 1x1 antenna configuration • Spatial channel model only on 570 serving links • No PMI or Rank feedback computation Scenario 2: about 1s (1000 TTIs) simulation time per 1 hour wall clock time • 57 cells, 10 users per cell, explicit interference modeling • 10 MHz (50 PRBs) • 4x4 antenna configuration • Spatial channel model on all 32490 links • Closed-loop spatial multiplexing PMI, CQI and Rank feedback computation every 2ms for PMI/CQI on every second PRB) Hardware: 8 cores Intel X5460 @ 3.16GHz (from 2008) Jan Ellenbeck Simulation and Prototyping Workshop 21 Technische Universität München RWTH Aachen, July 14, 2011
You can also read
