Achieving Large Multiplexing Gain in Distributed Antenna Systems via Cooperation with pCell Technology

3.1 MB·

by  Antonio Forenza

Achieving Large Multiplexing Gain
in Distributed Antenna Systems
via Cooperation with pCell Technology
Antonio Forenza∗, Stephen Perlman∗, Fadi Saibi∗, Mario Di Dio∗, Roger van der Laan∗, Giuseppe Caire†
∗Artemis Networks, LLC
355 Bryant Street, Suite 110
San Francisco, CA, 94107, USA
6 †Technische Universita¨t Berlin
1 Communications and Information Theory Chair
0 Einsteinufer 25, 10587 Berlin, Germany
2

n
a Abstract—In this paper we present pCellTM technology, the A multiuser wireless system with multiple transceiver sta-
J first commercial-grade wireless system that employs cooperation tions in its simplest form consists of N transmit antennas and

2 betweendistributedtransceiverstationstocreateconcurrentdata U single-antenna receivers (users), which in prior literature is
2 linkstomultipleusersinthesamespectrum.Firstweanalyzethe referred to as multiuser multiple-input multiple-output (MU-
per-usersignal-to-interference-plus-noiseratio(SINR)employing
] a geometrical spatial channel model to define volumes in space MIMO). The information theoretic model underlying MU-
T of coherent signal around user antennas (or personal cells, i.e., MIMOistheso-calledGaussianvectorbroadcastchannel,and
I pCells). Then we describe the system architecture consisting of has been the subject of intense investigation started with the
. a general-purpose-processor (GPP) based software-defined radio
s work of Caire and Shamai [3], that found the sum capacity
c (SDR) wireless platform implementing a real-time LTE protocol for the case U = 2 and proposed linear beamforming with
[ stack to communicate with off-the-shelf LTE devices. Finally we
present experimental results demonstrating up to 16 concurrent interference pre-cancellation, known as dirty-paper coding

1 spatial channels for an aggregate average spectral efficiency of [4], as a general achievability strategy. Successively, the sum
v 59.3 bps/Hz in the downlink and 27.5 bps/Hz in the uplink, capacity for general U ≥ 2 was found almost simultaneously
9 providing data rates of 200 Mbps downlink and 25 Mbps uplink
in [5–7]. The full characterization of the capacity region (with
0 in 5 MHz of TDD spectrum.1
no common message) was eventually given in [8], where
2
6 I. INTRODUCTION the optimality of beamforming and dirty-paper coding was
0 shownforageneralconvexinput covarianceconstraint.These
The increasing popularity of smartphones and tablets, and
. resultsassumethatthechannelstateinformation(CSI)isfixed
1 the growing demand for data-hungry applications like HD
and fully known to the transmitter and to the receivers. The
0
video streaming has resulted in skyrocketing mobile data
6 extension of the above results to the case where the CSI is
traffic.ArecentreportbytheCTIAtotheFCCshowedmobile
1 a random fading matrix, also known to all instantaneously,
datatrafficwillcontinuetogrowthroughoutthenextfouryears
: are almost immediate, especially for the case of ergodic rates,
v at an annual rate of about 40% [1], and states that cellular
i i.e., when the CSI evolves over time according to a matrix
X densification (through small-cells in 4G LTE networks) will
stationary and ergodic process.
be unable to keep pace with this growing demand for more
r Moving from theoretical models to real-world systems,
a data within current spectrum. New spectrum allocation may
several practical limitations arise. Given the low spatial di-
be a short-term fix, but mobile spectrum is finite while data
versity yielded by centralized antenna structures, performance
demand will grow indefinitely. One solution is to radically
of MIMO (or MU-MIMO) systems mostly relies upon the
improve the spectral efficiency (SE) of wireless networks.
limited multi-paths available in propagation channels [9,10],
In this paper we present pCell, a new wireless technology
andinpracticeatmost4xspatialmultiplexinggainisachieved
capable of achieving SE over an order of magnitude higher
[11,12]. One solution is to utilize far more antennas than the
than any current technology while remaining compatible with
number of users to increase spatial diversity, as in massive
existing 4G LTE devices [2]. pCell achieves these gains by
MIMO systems [13,14], and create independent spatial chan-
forgoing cellularization and exploiting interference in wireless
nelstomultipleconcurrentusersviabeamformingtechniques.
networks through large-scale cooperation between distributed
Massive MIMO, however, relies on highly complex base
transceivers, and by enabling high spatial multiplexing gain
station designs with many tightly-packed RF chains and a
via multiuser transmissions.
centralized antenna architecture which still limits the degrees
of freedom in wireless channels.
1TheworkoftheArtemisteamwassupportedbyRearden,LLC.
(cid:13)c2015 IEEE. Personal use of this material is permitted. Permission from The benefit of a de-centralized cellular architecture was
IEEE must be obtained for all other uses, in any current or future media, studied in [15] showing that spectral efficiency comparable
includingreprinting/republishingthismaterialforadvertisingorpromotional
to massive MIMO systems can be achieved with one order
purposes,creatingnewcollectiveworks,forresaleorredistributiontoservers
orlists,orreuseofanycopyrightedcomponentofthisworkinotherworks. of magnitude fewer antennas via network MIMO, by enabling
cooperation between base stationsin adjacent cells to mitigate where y ∈ CU×1 is the receive signal vector, x ∈ CN×1
inter-cell interference for enhancing cell-edge performance is the transmit signal vector subject to the power constraint
of cellular systems [16–22]. From the theoretical viewpoint, E{||x||2} = U, n ∈ CU×1 is the zero-mean additive white
network MIMO is equivalent to the Gaussian vector broadcast Gaussian noise vector with covariance matrix E{nn†} =
channel reviewed before [23,24], unless one takes explicitly N I ,andH∈CU×N isthechannelmatrixwithN ≥U,de-
o U
into account the constraints imposed by the underlying wired scribingthepropagationfromtheN transmitantennastotheU
network that connects the remote radio heads (RRHs) to the receive antennas. The channel vector h =[h ,...,h ]† ∈
u u1 uN
central processor. This network is referred to as backhaul, CN×1 is associated with each user u such that the channel
when the RRHs are seen as individual base stations that matrix is given by H=[h ,...,h ]† ∈CU×N.
1 U
somehow cooperate (in the so-called CoMP schemes) or, We use a channel model that accounts for the spatial
more modernly, as fronthaul, when the RRHs are simple dependency of the electromagnetic field between the transmit
and relatively dumb devices that form a whole distributed and receive antennas. Through this model we define the
base station together with the central processor (as in a notion of volumes in space of coherent signal and investigate
C-RAN architecture). Depending on the constraints on the how system configuration parameters affect its geometry. For
backhaul/fronthaul (e.g., topology [25], link rates [26]) and the sake of simplicity, we assume point antennas that are
typeofcooperation(e.g.,fulljointprocessing[23],coordinated unpolarized and isotropic radiators in far-field. We use the
beamforming [27], interference avoidance [28]), a very large modelofsphericalwavesinscatteringenvironmentsasin[29].
number of information theoretic problems and corresponding For ease of exposition, we also make the assumption that the
schemes have been investigated in the literature. distance between users is much smaller than their distance
There are fundamental capacity limits in cooperative net- to transmitters and scatterers as in [30] so that the complex
worksoperatingwithinthecellularframework[22],wherethe channelcoefficientbetweentransmitantennan=1,...,N and
spectral efficiency reaches an upper limit due to out-of-cluster receive antenna u=1,...,U is modeled as a superposition of
interference overwhelming the in-cluster signals, as transmit plane waves
(cid:88)
power increases. In this paper we propose a different net- h = a e−ikvˆp·(cid:126)ru (2)
un p
work architecture with transceivers distributed serendipitously
p∈Sun
without any concept of a cell, exploiting high densification
whereS isthesetofallpaths,includingline-of-sight(LOS)
un
with fixed transmit power to increase spatial multiplexing and non-line-of-sight (NLOS) components3, from propagation
gain. The transceivers are connected through a fronthaul and
and cluster scattering from transmit antenna n to receive
cooperateonalargescaletocreateconcurrentspatialchannels
antenna u, k =2π/λ is the wavenumber, λ is the wavelength,
to multiple users via precoding.
(cid:126)r is the location vector of user u relative to an origin O as
u
We begin by showing the benefits of a distributed ar- shown in Fig. 1, vˆ is the unit vector in the direction of the
p
chitecture over conventional cellular networks with multi- incident path p pointing out from the location of user u, and
ple centralized antennas, through a geometrical propagation a is a complex coefficient modeling pathloss, shadowing and
p
model in Section II. We use this model to analyze the phase terms independent of (cid:126)r for path p.
u
signal-to-interference-plus-noise ratio (SINR) as a function
of system parameters and demonstrate that by distributing
the transceivers randomly in space it is possible to achieve
volumes of coherent signal with high SINR around every
user antenna, which is impractical in centralized antenna
systems. Section III describes the GPP-based SDR wireless
platform implementing in real-time the pCell processing and
theentireLTEprotocolstack.Finally,SectionIVdemonstrates
experimentallyhowthevolumesofcoherentsignalenablehigh
multiplexing gain in a practical propagation environment.
II. SYSTEMMODELANDANALYSIS
A. System and Channel Model
The downlink of a multiuser system is modeled for a single
channel use on the time-frequency plane as2
Fig.1. ModelparametersforLOSchannels:angleγp∈[0,π]betweenpath
y=Hx+n (1) directionvˆp anddisplacementdirectionrˆwithN transmitantennas.
2We use ∗ to denote conjugation, T to denote transposition, † to denote B. SINR Performance via Multipole Expansion
conjugationandtransposition,|·|todenotetheabsolutevalue,||·||todenote The system performs transmit precoding to create multiple
the2-norm,IU todenotetheidentitymatrixofsizeU×U,CU×N todenote independentdownlinkdatastreamstotheusers.Ingeneralthe
complexmatrixofsizeU×N,(cid:104)·,·(cid:105)todenotethecomplexvectorspaceinner-
product, (cid:12) to denote element-by-element vector multiplication,(cid:126)· to indicate
avectorinthe3-dimensionalphysicalspace,ˆ·todenotesuchvectorofunit- 3Suchthatapincludestheterm(cid:112)K/(K+1)(whereK istheRicianK-
(cid:112)
norm,andiistheimaginaryunit. factor)fortheLOScomponentand 1/(K+1)fortheNLOScomponents.
Copyright2015SS&C.SubmittedtotheIEEEAsilomarConferenceonSignals,Systems,andComputers,Nov.8-11th2015,PacificGrove,CA,USA
transmit precoding is adaptively adjusted based on CSI of the location. For example, in the special case of all incident path
users. When the transmit precoding is fixed and the user at directions located on a cone with axis being the displacement
the location (cid:126)r is displaced by (cid:126)r, the received SINR changes directionrˆ(thatistosaythetermscosγ in(6)areallequalto
u p
fromSINR at(cid:126)r =(cid:126)otoSINR ((cid:126)r)astheuser’schannelvaries afixedcosγ),thenb(cid:96)(rˆ)=P (cosγ)h forany(cid:96).Therefore,
u u u (cid:96) u
from h to h ((cid:126)r). We define the volume of coherent signal for all v (cid:54)= u, the terms at the denominator in (7) satisfy
u u
for user u as the space where SINR((cid:126)r) at the displaced user’s the condition (cid:10)b(cid:96)(rˆ),w (cid:11) = 0, which entails there is no
u v
location exceeds a threshold SINR interference in that displacement direction rˆ.
o
V (SINR )={(cid:126)r +(cid:126)r; SINR ((cid:126)r)≥SINR } (3) C. Volumes of Coherent Signal
u o u u o
Next,wemakeafewassumptionstoclarifythedefinitionof
where SINR is chosen to meet a predefined error rate or
o
capacityperformance.ToderiveanexpressionoftheSINR((cid:126)r) volumeofcoherentsignalinthespecialcaseofLOSchannels.
Forsmalldisplacementdistances,(7)yieldstheapproximation
that allows insightful review through a tractable analytical
form, we assume x = Ws in (1), where s ∈ CU×1 is the |(cid:104)h ,w (cid:105)|2
transmitsignalvectorwithpowerconstraintE{||s||2}=U and SINR ((cid:126)r)≈ u u (8)
W=[w1,...,wU]∈CN×U isalinearzero-forcingprecoder u No+(kr)2 (cid:80) |(cid:104)b1u(rˆ),wv(cid:105)|2
v(cid:54)=u
as described in [31,32], although other precoding techniques
may be applied in practical deployments. which stems from j (ρ) = ρ(cid:96)/(2(cid:96) + 1)!!(1 + O(ρ2)) as
(cid:96)
Assuming the channel matrix H is full-rank and all users ρ → 0 [33]. In the case where the only significant path
are allocated equal power, for a given user u displaced by (cid:126)r is the LOS path as illustrated in Fig. 1, then b1(rˆ) =
u
relative to its original location (cid:126)r the SINR is given by h (cid:12) [cosγ ,...,cosγ ]T, where we used the fact that
u u u1 uN
P (cosγ) = cosγ, and γ is the relative angle between
|(cid:104)h ((cid:126)r),w (cid:105)|2 1 un
SINRu((cid:126)r)= No+ (cid:80)u |(cid:104)hu((cid:126)ru),wv(cid:105)|2 (4) tthhee ldoicsaptliaocnemofenutsedriruecttioontherˆlaoncdatitohneoufntirtanvsemctiotravnˆutennnfaronm.
v(cid:54)=u Therefore, the magnitude of the directional component of the
where h ((cid:126)r) = [h ((cid:126)r),...,h ((cid:126)r)]† ∈ CN×1 and the interferencetermexplicitlydependsontherelativeanglesγun
u u1 uN
for a given user location.
entries of the vector are defined from (2) as h ((cid:126)r) =
un
(cid:80)p∈Sunape−ikvˆp·((cid:126)ru+(cid:126)r). Applying the multipole expansion ortBhoygofunratlh,etrheanssuthmeinzgeroth-afotrcthinegUprecchoadnenrelbevceocmtoerss hsoutharaet
for plane waves [30] to the phasor term e−ikvˆp·(cid:126)r yields
w = h /||h || for u = 1,...,U. If in addition N = U then
u u u
+∞ (8) simplifies to5
(cid:88)
h ((cid:126)r)= i(cid:96)(2(cid:96)+1)j (kr)b(cid:96)(rˆ) (5)
u (cid:96) u SNR
(cid:96)=0 SINRu((cid:126)r)≈ 1+(kr)2SNR (cid:80) ξ ξu (cosγ −cosγ )2
where the displacement vector (cid:126)r is decomposed into its norm u us ut us ut
1≤s<t≤N
r and unit direction vector rˆ = (cid:126)r/r, j(cid:96)(·) is the spherical (9)
Bessel function of the first kind and order (cid:96) and b(cid:96)(rˆ) = where SNR = SINR ((cid:126)o) = ||h ||2/N is the SINR for
u u u u o
(cid:2)b(cid:96)u1(rˆ),...,b(cid:96)uN(rˆ)(cid:3)† ∈CN×1 is such that user u at its original location, ξun = |hun|2/||hu||2 is the
fraction of the total channel power gain from antenna n and
(cid:88)
b(cid:96)un(rˆ)= ape−ikvˆp·(cid:126)ruP(cid:96)(cosγp) (6) cosγun = rˆ·vˆun. For a fixed displacement direction rˆ, the
p∈Sun SINR approximation in (9) is a Lorentzian function of the
displacement distance r with maximum value SNR at r =0.
where P (·) is the Legendre polynomial of degree (cid:96) and u
(cid:96)
Based on the definition of volume of coherent signal in
cosγ =rˆ·vˆ . Substituting (5) into (4) we obtain
p p
(3), we consider the surface boundary for V (SINR ) where
u o
(cid:12)(cid:12)(cid:12)(cid:12)+(cid:80)∞(−i)(cid:96)(2(cid:96)+1)j(cid:96)(kr)(cid:10)b(cid:96)u(rˆ),wu(cid:11)(cid:12)(cid:12)(cid:12)(cid:12)2 tShIeNRSIuN((cid:126)rR) i=s eSqIuNaRl oto. Tahpernedweefinuesdetthhreesahpopldroxviamluaetiosunchinth(9a)t
SINRu((cid:126)r)= No+(cid:96)v=(cid:80)(cid:54)=0u(cid:12)(cid:12)(cid:12)(cid:12)+(cid:96)(cid:80)=∞1(−i)(cid:96)(2(cid:96)+1)j(cid:96)(kr)(cid:104)b(cid:96)u(rˆ),wv(cid:105)(cid:12)(cid:12)(cid:12)(cid:12)2 dtcooirhedecertreiiovnnet sraˆigcanlsoa6slefdo-rfoursmerexupraesssaiofnunocftitohneorafdtihues odfisvpolalucemmeeonft
(7) (cid:114)
λ 1 1
where the term of order 0 in the series at the denominator R (rˆ)≈ − .
u (cid:114)
vanishes since b0u(rˆ) = hu and we assumed a zero-forcing 2π (cid:80)ξusξut(cosγus−cosγut)2 SINRo SNRu
precoder4. s<t
The expansion in (5) decouples the effect of displacement (10)
distance r and displacement direction rˆ, thereby showing how
5IfN >U thentherighthandsideof(9)becomesanapproximatelower
theSINRin(7)variesastheuserisdisplacedfromitsoriginal
boundforSINRu((cid:126)r).
6Note that this approximation relies on the orthogonality assumption and
4Notethatattheuserantennaoriginallocation(i.e.,(cid:126)r=(cid:126)o)alltermswith is only quantitatively valid if Ru(rˆ) (cid:28) λ/2π but is useful for drawing
(cid:96) ≥ 1 are zero, j0(0) = 1 and the interference term at the denominator qualitative conclusions nonetheless. An expression that does not use that
vanishessuchthatSINRu((cid:126)o)=SNRu=|(cid:104)hu,wu(cid:105)|2/No. assumptioncanbederivedfrom(8).
Copyright2015SS&C.SubmittedtotheIEEEAsilomarConferenceonSignals,Systems,andComputers,Nov.8-11th2015,PacificGrove,CA,USA
From (10) we derive the following important observations
about R (rˆ):
u
• it is proportional to the wavelength λ;
• it decreases as SINRo increases;
• it depends only on SINRo when SNRu (cid:29)SINRo;
• it depends, for a given displacement direction rˆ, on the
layout of the transmit antennas through the angles γ ;
un
• it becomes large when the angles γun become close to
each other as in centralized antenna arrays.
Equation (10) provides insights on how system parameters
affect the geometry of the volumes of coherent signal. In
particular it indicates that a system with distributed transmit
antennas can create a volume of coherent signal with very
small radius R (rˆ) in all directions. Whereas if the transmit
u
antennas are centralized, the radius becomes larger in all
directionsand,morespecifically,muchlargerinthedirections
parallel to the line joining the user location to the center of Fig.3. Distributedtransmitantennas:envelopeofSINRu((cid:126)r)onthehorizontal
the distant group of transmit antennas. planefor8usersregularlyspaced4λapartontheaxis{y=0,z=0}and
10transmitantennasrandomlydistributedabovetheuserantennaswithinthe
Indeed for centralized transmit antennas distant from the region{−50λ≤x≤+50λ,−50λ≤y≤+50λ,+50λ≤z≤+200λ}.
user locations the terms (cosγ −cosγ )2 in (10) are all
us ut
smallresultinginlargedimensionsforthevolumeofcoherent
signal(ifitexistsatall7).Furthermorethevolumeofcoherent By contrast, when the transmit antennas are distributed and
signal takes on an elongated beam shape. For example, Fig. 2 randomly placed, most of the terms (cosγ − cosγ )2 in
us ut
showstheenvelopeoftheSINRoveratwo-dimensionalcross- (10)arenotnegligible,therebythevolumesofcoherentsignal
section for all users computed through the exact expression have smaller dimensions in all directions. For example, Fig. 3
in (4) accounting for spherical wave propagation for U = 8 shows the envelope of the SINR for the same conditions as in
users uniformly spaced 4λ apart and placed parallel at the Fig. 2 but for distributed transmit antennas instead.
broadside of a λ/2 uniform linear array (ULA) of N = 10 Fig. 4 shows the volume of coherent signal in (3) around
transmit antennas located 50λ away. The channel model is each user with SINR = 5dB (e.g., corresponding to CQI
o
urban microcell LOS with pathloss and shadowing computed 7 or 16-QAM with spectral efficiency of 1.48 bps/Hz at a
according to the 3GPP model [9]. Note that for all users the block error rate of 10% as per the LTE standard [35,36]).
radius of the volumes of coherent signal is much larger in the The SINR is computed using (4) for a system with U = 10
direction pointing to the center of the ULA. users randomly located in a cube of side dimension 2λ and
N =16 transmit antennas distributed above the UE locations
with a maximum distance of 300λ (i.e., realistic dimensions
for a deployment at 1.9 GHz carrier frequency).
Fig.2. Centralizedtransmitantennas:envelopeofSINRu((cid:126)r)onthehorizon-
tal planefor 8users regularlyspaced 4λ apart onthe axis{y =0,z =0}
and10transmitantennasspacedλ/2apartontheaxis{y=50λ,z=0}.
7Note that for centralized antennas the initial assumption of a full-rank Fig. 4. Volumes of coherent signal for U = 10 users randomly placed
channelmatrixandevenmoresotheassumptionoforthogonalchannelvectors in a cube of side dimension 2λ for N = 16 transmit antennas distributed
isaveryhardonetomeetinaLOSchannel.Itrequirespurposefulplacement above the user locations in {−300λ ≤ x ≤ +300λ,−300λ ≤ y ≤
oftheuserantennas.Atwo-dimensionalexampleisprovidedin[34]. +300λ,+200λ≤z≤+300λ}.Differentcolorsrefertodifferentusers.
Copyright2015SS&C.SubmittedtotheIEEEAsilomarConferenceonSignals,Systems,andComputers,Nov.8-11th2015,PacificGrove,CA,USA
We obtained experimental evidence of the volumes of co-
herent signal by measuring variations of the SINR for fixed
precodingwhiledisplacinguserantennaseversoslightlyfrom
their baseline position. These experiments indicated that the
size of the volume of coherent signal is a fraction of the
wavelength, such that user devices can be densely packed
as in the experimental results in Section IV. Furthermore,
the precoder of the pCell wireless platform is periodically
updatedsuchthatthevolumesofcoherentsignalfollowusers’
motion. Experiments also showed that the volume of coherent
signalactuallydependsondisplacementvariables((cid:126)r,ψ(cid:126))ina6-
dimensional manifold, where in addition to the 3-dimensional Fig.5. pCellhardwarearchitecture.
variable (cid:126)r used in the present model the variable ψ(cid:126) belongs
to the 3-dimensional manifold of Euler angles parameterizing
the rotation of each user antenna. Future work will analyze B. Software Architecture
dependencyonrotationanglesbyextendingthepresentmodel
The pCell system is an SDR platform where all the func-
to account for antenna radiation pattern and polarization.
tional blocks of the LTE protocol stack, from the gateway
III. DESCRIPTIONOFTHESDRWIRELESSPLATFORM down to the physical layer, and all pCell processing have
ThepCellsoftware-definedradio(SDR)wirelessplatformis been implemented from scratch in C++ modules running in
implementedasacloudradioaccessnetwork(C-RAN),where real-time on the GPP platform, without the need for special-
baseband processing is performed on GPP servers in a data ized hardware such as DSPs, co-processors or FPGAs. This
center. The data center provides I/Q waveforms through fiber SDR implementation provides maximum flexibility, interoper-
connections to RRHs called pWaveTM radios, which consist ability and portability. The software architecture is designed
only of analog-to-digital (A/D), digital-to-analog (D/A), and to minimize computational overhead, optimize throughput
RF up/down converters, power amplifier and antenna. This and provide stable and deterministic computational load be-
section describes the hardware and software architectures of havior. The proprietary development framework provides a
the pCell SDR wireless platform as well as aspects of its module-oriented environment with multi-threading/multi-core
operation with existing off-the-shelf LTE devices. programming capability and support for real-time over-the-
network operation.
A. Hardware Architecture
For achieving the level of efficiency required to meet the
As illustrated in Fig. 5 the pCell system is composed of real-time constraints of the LTE protocol and the pCell pro-
two parts: i) a GPP-based data center that implements the cessing, the software modules are categorized in two classes:
LTE protocol and pCell processing; ii) a radio access network hard real-time and soft real-time. The hard real-time modules
(RAN) including data switches and radio transceivers. While implement tasks that must be completed within a fraction
thesoftwarerunninginthedatacenterremainsunchanged,the of the 1 ms LTE subframe (SF) duration. Completing these
pCell RAN comes in different flavors so as to accommodate tasks on time is critical for maintaining the integrity of the
differentdeploymentscenarios.Afiberfronthaul,whichtrans- pCell-synchronizedLTEwaveformandthestabilityofthedata
mitsaduplexofI/Qdigitalsamplestreams,isroutedfromthe connection with the user equipment (UE). These operations
servers in the pCell data center and can be connected to: include all of the pCell processing and most of the LTE
1) pWaves configured with a fiber interface; physical layer functions (e.g., turbo coding/decoding, FFT,
2) LOS radios that connect to pWaves configured for channel estimation/equalization) as well as some of the MAC
1000BASE-T (i.e., copper Gigabit Ethernet); layer features such as the PRACH procedure. The soft real-
3) the uplink port of a 1000BASE-T switch that connects time modules implement functional blocks from the higher
to pWaves configured for 1000BASE-T; layersoftheLTEprotocolstack(e.g,RLC,PDCP,RRC,NAS,
4) an Artemis Hub composed of 32 pWave radios that Gateways) that are subject to time constraints in the order
connect to up to 32 antennas through coaxial cables. of multiples of an SF interval. The software architecture is
In this paper we present experimental results obtained with built around this classification and, by using tools provided
the RAN hardware configuration #4. The pWave radios can by the development framework, the system can balance the
operate at any carrier frequency from 400 MHz to 4.4 GHz computational load and guarantee on-time execution of the
and are synchronized using 10MHz/PPS signals that can be critical hard real-time tasks without missing the deadlines of
retrieved from a GPS reference, or from an in-band timing the soft real-time ones.
signal itself slaved to a GPS reference. For the measurement For each UE a set of computing resources and data struc-
results presented in this paper the GPP-based data center tures, called a virtual radio instance (VRI), is allocated to
utilizes three off-the-shelf dual-processor Intel motherboards, instantiate an entire LTE protocol stack, thus forming the
two of them equipped with Intel Xeon E5-2687W v2 (8 cores functional equivalent of a dedicated LTE eNodeB per UE. A
@3.40GHz) and the third with Intel Xeon E5-2697 v3 (14 VRI is spawned as soon as a user begins the attach procedure
cores @2.60GHz). and remains operational throughout the duration of the user’s
Copyright2015SS&C.SubmittedtotheIEEEAsilomarConferenceonSignals,Systems,andComputers,Nov.8-11th2015,PacificGrove,CA,USA
connection, maintaining its active state. When a user detaches
from the network, the VRI manager saves any relevant state
andtheVRIinstanceisreleased.Throughphysicallayersignal
processing each VRI is associated with a volume of coherent
signal (which was duly defined in Section II-B) around the
antenna of the corresponding user.
AsshowninSectionII-C,suchavolumeofcoherentsignal
isconfinedinspatialdimensionsontheorderoforsmallerthan Fig.7. LTETDDframestructure(TDDconfig.#2andSpcconfig.#7)
the wavelength, resulting in a concurrent spatial multiplexing
unit for each UE. Thus, every VRI has a concurrent full-
bandwidth, independent physical-layer link with its associated
well as timing advance to compensate for round-trip time of
UE, providing each UE with the experience of an unshared
flight. These parameters are broadcasted by the eNodeB for
eNodeBdeliveringthefullLTEchannelbandwidth,regardless
theUEstosetupcompatibleoperationpriortoattachingtoit.
ofnumberofVRI/UElinksconcurrentlyinthesamespectrum.
Within the LTE protocol, the DwPTS is used to transmit
In Fig. 6 the spatial processing functions are denoted pCell
DLdataandtheUpPTStosendULsoundingreferencesignal
processing and the volumes of coherent signals are conceptu-
(SRS) from the UEs, consisting of Zadoff-Chu sequences
ally depicted as dashed circles with numeric labels matching
multiplexed over the entire system bandwidth. The SRS is
the labels of their associated VRIs. Consequently, different
used in current LTE networks to perform channel condition
VRIs can implement different protocols concurrently in the
measurements for scheduling purposes. pCell uses the SRS to
samespectrum.SomeVRIscanimplementLTEeNodeBwhile
derive precise UL CSI and Doppler information from all UEs,
other VRIs can implement proprietary or standard protocols,
while avoiding additional overhead compared to existing LTE
for example, ones better suited for low-power “Internet of
networks. In 5 MHz bandwidth and with the TDD frame in
Things” devices. Both LTE and non-LTE devices will concur-
Fig. 7, every 5 ms up to 32 concurrent UEs transmit the SRS
rently and independently operate in the same spectrum, with
tothepWaves.Inpracticaldeploymentsfordenselypopulated
each device experiencing only the protocol from its own VRI.
areas (e.g., stadium) users are typically stationary or nomadic,
and different periodicities of SRS transmissions can be set
depending on their speed as allowed by the LTE standard.
For example, if the periodicity is set to 20 ms for all UEs,
up to 128 concurrent users can be supported in the same
bandwidth.Further,typicalLTEdeploymentsallocate20MHz
blocks of spectrum which can be subdivided in four channels
of 5 MHz each, thereby allowing up to 512 concurrent users
with orthogonal SRS in the same coverage areas with 20
ms periodicity. In scenarios requiring even larger numbers
of concurrent users, pCell employs spatial reuse schemes to
increase further the number of orthogonal SRS transmissions.
The pWave radios synchronously receive the SRS signals,
convert from RF domain to I/Q samples, which are sent to the
Fig.6. pCellsoftwarearchitecture. pCell data center. The samples are processed to produce accu-
rate UL CSI estimates, which are used to derive accurate DL
CSI by exploiting TDD reciprocity through proprietary signal
C. LTE Protocol and Frame Structure
processingalgorithms.TheUL/DLCSIisfurtherprocessedto
The pCell system is compatible with off-the-shelf LTE deriveparametersforUL/DLspatialprocessing.Followingthe
devices. In LTE systems, the downlink (DL) uses orthogonal pCell processing operations on UL I/Q samples, conventional
frequency division multiple access (OFDMA) and the uplink LTE physical (PHY) layer processing is applied to each user
(UL)usessingle-carrierorthogonalfrequencydivisionmultiple UL stream including equalization and turbo decoding to form
access (SC-FDMA), with quadrature amplitude modulation a protocol data unit (PDU) passed to the MAC layer. The UL
(QAM) of different orders. The system bandwidth can range datathenproceedsuptheLTEprotocolstackwithineachVRI
from 1.4 to 40 MHz. The frame structure changes depending associated with a particular UE to eventually be sent to the
on whether the frequency-division duplex (FDD) or time- Internet. The VRI also provides DL data through the protocol
divisionduplex(TDD)modeisused.TheLTEframeisformed stackintheformofMACPDUsscheduledfortransmissionin
of 10 subframes (SF) of duration 1 ms each and sub-divided the DL SFs and DwPTS, processed according to conventional
into 14 OFDM symbols. In TDD mode, the SF is either of LTE PHY operations. The pCell processing then converts the
type DL, UL or special (Spc). The Spc SF is placed between U streams of user DL samples into N streams of pWave I/Q
every DL and UL SF as in Fig. 7, and consists of a downlink samples, which are finally transported to the pWave radios.
pilot time slot (DwPTS), a guard period (GP) and an uplink The pWaves convert the I/Q samples to the RF domain and
pilot time slot (UpPTS). The GP allows for RF switching as synchronously transmit the waveforms.
Copyright2015SS&C.SubmittedtotheIEEEAsilomarConferenceonSignals,Systems,andComputers,Nov.8-11th2015,PacificGrove,CA,USA
ThepCellspectralefficiency(SE)isdeterminedbymeasur-
ingtheaggregateSEofagroupofusers(unmodifiediPhone6
Plusdevicesareusedforthesemeasurements)allconcurrently
transmitting and receiving data from the pCell antennas in
the coverage area. The iPhone 6 Plus devices are placed in
a uniform pattern on a 1m2 plexiglass table as in Fig. 8
movedthroughoutthecoverageareain75cmincrements.The
LTE MAC layer PDU throughputs are used to calculate the
aggregate DL and UL SE at every location. The calculation
of the LTE SE numbers presented in this paper are calibrated
against the SE tables reported in [37].
Heat maps of the aggregate DL and UL SE for the 16
iPhone6Plusdevicesthroughoutthecoverageareaareshown
in Fig. 9 and Fig. 10, respectively, along with the layout of
32 antennas (white squares with the blue Artemis logo). The
Fig.8. 4GLTEdevicesdenselypackedin1m2.
average aggregate DL SE across all locations is 59.3 bps/Hz,
with peak of 59.8 bps/Hz (corresponding to an aggregate DL
throughput of approximately 200 Mbps) and 5% outage [38]
IV. EXPERIMENTALRESULTS
of 58.1 bps/Hz mostly due to locations in the upper right
pCellhasbeentestedbothindoorandoutdoor,withdetailed
corner(obstructedbyseveralwalls).AtitsDLSEpeak,every
indoortestingcompletedthusfar.Theresultspresentedinthis
UE receives data using LTE modulation and coding scheme
paper have been obtained using the RAN configuration #4
(MCS)28whichcorrespondstoa64-QAMOFDMmodulation
described in Section III-A. Although pCell supports arbitrary
with FEC coding rate of 0.9. Fig. 10 shows the aggregate
antenna placement, for the purpose of the SE testing, 32
UL SE is consistently at peak of 27.5 bps/Hz (corresponding
antennas are placed in a regular grid with roughly 2.5 meter
to an aggregate UL throughput of 25 Mbps) throughout the
spacing, with aligned polarization, at a uniform height (except
entire coverage area. Every UE transmits data using MCS 20
in low-ceiling corridors) and pointing downward, resulting in
(i.e., 16-QAM SC-FDMA modulation with FEC coding rate
amixofLOSandNLOSpaths,somethroughwallsandsome
of 0.7), which is the maximum UL MCS supported by LTE
through free space. The antennas are 2”x2” patch antennas
UE category 4 chipsets [39] used by the iPhone 6 Plus.
with 8 dBi and HPBW = 75◦. Every antenna transmits a
waveform with LTE-compliant spectral envelope generated in Because pCell processing results in high SINR throughout
the data center with the time-domain frame structure in Fig. 7 the coverage area, in almost all locations the SE is limited by
and the following parameters: the maximum MCS supported by the iPhone 6 Plus (MCS 28
forDLandMCS20forUL).FutureLTEdevicesareexpected
• Carrier frequency: 1917.5 MHz;
to enable higher order modulation schemes (e.g. 256-QAM
• Systembandwidth:5MHz(with300OFDMsubcarriers);
[40]), which pCell will be supporting given the large SINR
• Average transmit power per antenna: 1 mW;
margin it provides.
• TDD configuration #2 for a DL to UL ratio of 3:1;
• Spc SF configuration #7.
Fig.9. DownlinkpCellSE(average=59.3bps/Hz,peak=59.8bps/Hz). Fig.10. UplinkpCellSE(average=27.5bps/Hz,peak=27.5bps/Hz).
Copyright2015SS&C.SubmittedtotheIEEEAsilomarConferenceonSignals,Systems,andComputers,Nov.8-11th2015,PacificGrove,CA,USA
V. CONCLUSION [19] Y.Liang,R.Valenzuela,G.Foschini,D.Chizhik,andA.Goldsmith,“In-
terferencesuppressioninwirelesscellularnetworksthroughpicocells,”
WepresentedpCell,amultiuserwirelesssystemthatenables Proc.ofAsilomarConf.onSign.,Syst.andComputers,pp.1041–1045,
Nov.2007.
cooperation between distributed transceivers to achieve large
[20] S. Venkatesan, A. Lozano, and R. Valenzuela, “Network MIMO:
multiplexing gain while foregoing cellularization. By forming overcoming inter-cell interference in indoor wireless systems,” Proc.
individual volumes of coherent signal with high SINR around ofAsilomarConf.onSign.,Syst.andComputers,pp.83–87,Nov.2007.
[21] S.Venkatesan,H.Huang,A.Lozano,andR.Valenzuela, “AWiMAX-
everyuser’santenna,pCellcreatesindependentspatiallinksto
basedimplementationofnetworkMIMOforindoorwirelesssystems,”
multiple users, thereby providing over an order of magnitude EURASIPJourn.onAdv.inSign.Proc.,Sep.2009.
increase in spectral efficiency compared to existing technolo- [22] A. Lozano, R.W. Heath, and J.G. Andrews, “Fundamental limits of
cooperation,” IEEETrans.Info.Th.,vol.59,no.9,pp.5213–5226,Sep.
gies. We introduced an analytical framework based on a
2013.
geometricalchannelmodeltodescribethevolumesofcoherent [23] O.Somekh,B.M.Zaidel,andS.Shamai(Shitz), “Sumratecharacter-
signal. Then we reviewed hardware and software architecture ization of joint multiple cell-site processing,” IEEE Trans. on Inform.
Theory,vol.53,pp.4473–4497,Dec.2007.
of pCell as implemented within the LTE protocol. Finally
[24] H. Huh, A. M. Tulino, and G. Caire, “Network MIMO with linear
we demonstrated high downlink and uplink spectral efficiency zero-forcing beamforming: large system analysis, impact of channel
achieved with pCell in practical propagation conditions. estimation,andreduced-complexityscheduling,”IEEETrans.onInform.
Theory,vol.58,no.5,pp.2911–2934,May2012.
[25] V.S.Annapureddy,A.ElGamal,andV.V.Veeravalli, “Degreesoffree-
REFERENCES dom of interference channels with CoMP transmission and reception,”
IEEE Trans. on Inform. Theory, vol. 58, no. 9, pp. 5740–5760, Sep.
2012.
[1] T.K.SawanoboriandR.Roche, “Mobiledatademand:growthforecast
[26] S.-N.HongandG.Caire, “Compute-and-forwardstrategiesforcooper-
met,” CTIA,Jun.2015.
ativedistributedantennasystems,” IEEETrans.onInform.Theory,vol.
[2] S. Perlman and A. Forenza, “An introduction to pCell,” Feb. 59,no.9,pp.5227–5243,Sep.2013.
2015, http://www.rearden.com/artemis/An-Introduction-to-pCell-White- [27] H.Huh,H.C.Papadopoulos,andG.Caire,“MultiuserMISOtransmitter
Paper-150224.pdf. optimizationforintercellinterferencemitigation,” IEEETrans.onSig.
[3] G. Caire and S. Shamai (Shitz), “On the achievable throughput of Proc.,vol.58,no.8,pp.4272–4285,Aug.2010.
a multiantenna Gaussian broadcast channel,” IEEE Trans. on Inform. [28] G. Caire, S.A. Ramprashad, H.C. Papadopoulos, C. Pepin, and C.-
Theory,vol.49,pp.1691–1706,Jul.2003. E.W. Sundberg, “Multiuser MIMO downlink with limited inter-cell
[4] M. Costa, “Writing on dirty paper,” IEEE Trans. on Inform. Theory, cooperation:approximateinterferencealignmentintime,frequencyand
vol.29,pp.439–441,May1983. space,” inProc.AllertonConf.onCommun.,Control,andComputing,
[5] P.ViswanathandD.Tse,“SumcapacityofthevectorGaussianbroadcast Urbana-Champaign,IL,Sept.2008.
channelanduplink-downlinkduality,” IEEETrans.onInform.Theory, [29] T. Svantesson, “A study of polarization diversity using an electromag-
vol.49,pp.1912–1921,Aug.2003. neticspatio-temporalchannelmodel,” Proc.IEEEVeh.Tech.Conf.,vol.
[6] S.Vishwanath,N.Jindal,andA.Goldsmith, “Duality,achievablerates, 1,pp.79–86,Sep.2000.
and sum-rate capacity of Gaussian MIMO broadcast channels,” IEEE [30] A.S.Y.Poon,R.W.Brodersen,andD.N.C.Tse, “Degreesoffreedomin
Trans.onInform.Theory,vol.49,pp.2658–2668,Oct.2003. multiple-antennachannels:Asignalspaceapproach,” IEEETrans.Info.
[7] W. Yu and J. M. Cioffi, “Sum capacity of Gaussian vector broadcast Th.,vol.51,pp.523–536,Feb.2005.
channels,”IEEETrans.onInform.Theory,vol.50,pp.1875–1892,Sep. [31] Q.H.Spencer,A.L.Swindlehurst,andM.Haardt,“Zero-forcingmethods
2004. fordownlinkspatialmultiplexinginmultiuserMIMOchannels,” IEEE
[8] H.Weingarten,Y.Steinberg,andS.Shamai(Shitz),“Thecapacityregion Trans.Sig.Proc.,vol.52,pp.461–471,Feb.2004.
oftheGaussianmultiple-inputmultiple-outputbroadcastchannel,”IEEE [32] R.Chen,J.G.Andrews,andR.W.HeathJr.,“Transmitselectiondiversity
Trans.onInform.Theory,vol.52,pp.3936–3964,Sep.2006. formultiuserspatialmultiplexingsystems,” IEEETrans.Sig.Proc.,vol.
[9] 3GPP Technical Report, “Spatial channel model for multiple input 55,pp.1159–1171,Mar.2007.
multiple output (MIMO) simulations (Release 6),” TR 25.996, V6.1.0, [33] M.AbramowitzandI.A.Stegun, HandbookofMathematicalFunctions
Sep.2003. withFormulas,Graphs,andMathematicalTables,UnitedStatesDepart-
[10] V. Erceg et. al., “TGn channel model,” IEEE 802.11 Wireless LANs, mentofCommerce,10thPrinting,1972,(page437).
doc.:IEEE802.11-03/940r4,May2004. [34] P.F.DriessenandG.J.Foschini, “Onthecapacityformulaformultiple
[11] K.Werner,H.Asplund,D.V.P.Figueiredo,N.Jaldn,andB.Halvarsson, input-multiple output wireless channels: A geometric interpretation,”
“LTE-Advanced 8x8 MIMO measurements in an indoor scenario,” IEEETrans.Comm.,vol.47,pp.173–176,Feb.1999.
EricssonAB,Stockholm,Sweden,Apr.2013. [35] 3GPP Technical Specification, “E-UTRA: physical layer procedures,”
[12] A. Goldsmith, “Designing reliable Wi-Fi for HD delivery through- TS36.213,V8.8.0,Oct.2009(Table7.2.3-1).
out the home,” Quantenna communications white paper, Mar. 2009, [36] A. Alexiou, C. Bouras, V. Kokkinos, A. Papazois, and G. Tsichritzis,
http://www.quantenna.com/pdf/DesigningReliableWiFi.pdf. “EfficientMCSselectionforMBSFNtransmissionsoverLTEnetworks,”
[13] T.L.Marzetta,“Noncooperativecellularwirelesswithunlimitednumbers IFIP,pp.1–5,Oct.2010.
ofbasestationantennas,”IEEETrans.onWirelessCommun.,vol.9,no. [37] 3GPPTechnicalReport, “Feasibilitystudyforfurtheradvancementsfor
11,pp.3590–3600,Nov.2010. E-UTRA(LTE-Advanced),” TR36.912,V12.0.0,Sep.2014.
[38] ITU-R, “Requirements related to technical performance for IMT-
[14] E. Larsson, O. Edfors, and T.L. Marzetta, “Massive MIMO for next
Advancedradiointerface(s),” M.2134,2008.
generation wireless systems,” IEEE Comm. Mag., pp. 186–195, Feb.
[39] 3GPP Technical Specification, “User equipment (UE) radio access
2014.
capabilities,” TS36.306,V12.6.0,Oct.2015(Table4.1-2).
[15] H.Huh,G.Caire,H.C.Papadopoulos,andS.Ramprashad, “Achieving
[40] 3GPP Technical Specification, “E-UTRA: physical layer procedures,”
”massive MIMO” spectral efficiency with a not-so-large number of
TS36.213,V12.4.0,Dec.2014(Table7.2.3-2).
antennas,” IEEETrans.onWirelessCommun.,vol.11,no.9,pp.3226–
3239,Sep.2012.
[16] G.J. Foschini, M.K. Karakayali, and R.A. Valenzuela, “Coordinating
multiple antenna cellular networks to achieve enormous spectral effi-
ciency,” IEEProc.Comm.,vol.153,no.4,pp.548–555,Aug.2006.
[17] M.K.Karakayali,G.J.Foschini,R.A.Valenzuela,andR.D.Yates, “On
themaximumcommonrateachievableinacoordinatednetwork,” Proc.
IEEEInt.Conf.onComm.,vol.9,pp.4333–4338,Mar.2006.
[18] M.K.Karakayali,G.J.Foschini,andR.A.Valenzuela, “Networkcoordi-
nationforspectrallyefficientcommunicationsincellularsystems,”IEEE
WirelessComm.Mag.,vol.13,no.4,pp.56–61,Aug.2006.
Copyright2015SS&C.SubmittedtotheIEEEAsilomarConferenceonSignals,Systems,andComputers,Nov.8-11th2015,PacificGrove,CA,USA