Vous êtes sur la page 1sur 6

MAC v.

PHY: How to Relay in Cellular Networks


Zinan Lin, Mohammed Sammour, Sana Sfar, Gregg Charlton, Prabhakar Chitrapu, Alex Reznik
InterDigital
King of Prussia, PA 19406
USA
{Zinan.Lin, Mohammed.Sammour, Sana.Sfar, Gregg.Charlton, Prabhakar.Chitrapu, Alex.Reznik}
@InterDigital.com

Abstract—Two cooperative relaying protocols are presented transmissions from the source and relay. While this may sound
and analyzed. The first emphasizes PHY cooperation via syn- far-fetched, an MMSE-SIC receiver is formally a candidate
chronized transmissions and distributed space-time coding from receiver for 3GPP’s Long Term Evolution (LTE) technology
the Base Station (BS) and Relay Station (RS) to the User
Equipment (UE) during the protocol’s cooperative phase. The [4] for separating between spatial streams from the same
second protocol emphasizes MAC cooperation using different transmitter. Thus, placing the source’s and relay’s transmis-
MAC flows from the BS and RS to the UE during the cooperative sions into separate transmission streams and using a SIC to
phase and utilizing successive interference cancellation (SIC) at receive these is in fact no longer beyond the scope of current
the UE. This paper discusses and compares the two protocols’ technology. In fact, at least for OFDM MIMO technologies,
flexibility and performance aspects via analytical and simulation
results. it may not even require additional receiver structures.
The latter fact becomes the starting point for the results
I. I NTRODUCTION presented in this paper. Specifically, we take advantage of the
Utilization of relay nodes in cellular networks is currently apparent practicability of the SIC receiver and demonstrate
the subject of intense study and development by both industry that once such a receiver is introduced into a communication
and the research community [1], [2], [3], [5], [6]. The fact system much of the advantage of cooperative diversity may
is not surprising. Relays are ideally suited to addressing such be relegated to the MAC layer. Instead of collaborative trans-
issues as coverage holes and cellular range extension. More mission and coding, we demonstrate that a well scheduled
interesting still are recent results [9], [10] which demonstrate combination of direct transmission and simple multi-hop can
that utilizing relays appropriately can improve quality of achieve the benefits of cooperative and, in some cases, even
service and capacity of a cellular network even in a purely exceed what can be delivered by a well designed PHY-layer
interference limited deployment. scheme. We compare our PHY and MAC oriented cooperation
This is not surprising given the extent of theoretical results schemes in the cellular system and use a system-level simula-
on the link budget improvements that relays can deliver. These tion, which evaluates the impact of a transmission scheme on
date back to the pioneering analysis by Cover and El Gamal the system as a whole, instead of just under particular channel
[8] and were the subject of extensive investigation in the conditions. We note that our approach to obtaining cooperative
last decade. Notable among the recent result are the work of advantage at the MAC layer is not new, previous work [15],
Kramer, Gastpar and Kumar [11], Laneman et. al [12], [14] as [16], [19], [20] has examined the possibility as well. However,
well as Sendonaris, Erkip and Azhang [15], [16]. All of these we do believe that the observations that this can be done with
results point to the fact that link performance can be improved little loss and even with a gain over PHY-layer schemes are
significantly through intelligent use of relays; however simple new.
multi-hop relaying (i.e. one where the relay just forwards the The rest of this paper is structured as follows. In Section
same data that it receives) is not likely to result in significant 2 we define our cooperative protocols and present link-level
gains. Instead, more sophisticated cooperative techniques must performance analysis for both. In Section 3 we describe the
be employed. system-level simulation methodology and present simulation
A key consequence of these results is that techniques which results. We conclude the paper with some final remarks in
are physical layer in nature must be employed to gain the Section 4.
complete benefit of cooperative utilization of relays. Among
these are cooperative coding schemes, such as e.g. [13], II. C OOPERATIVE P ROTOCOLS
[18] and distributed beamforming techniques (e.g. [17]). It is We now turn our attention to the cooperative protocols
also possible to envisage an approach aligned with the more considered in this paper. As the primary focus of our analysis
theoretical work of [8], [11], where a multi-user detector (more is the cellular setting, we explicitly use the cellular terminol-
precisely a successive interference canceler) is utilized in the ogy of base station, relay station and user equipment (BS,
receiver to optimize the performance of joint reception of RS, UE) and concentrate on the downlink. We shall discuss
paper we shall avoid the radio resource management
(RRM) issues associated with multi-code scheduling.
• We assume a single transmit antenna at all terminals.
Our results extend to multiple antennae, but considering
a single antenna only reduces the number of degrees of
freedom and makes it easier to present a complete picture
within the space constraint.
We are now prepared to describe our protocols. Protocol 1
(P1) is defined as follows:
• Given a message of m bits, the BS encodes the m bits at
a rate R1,BS,RS and transmits these in Phase 1. We note
that since the RS must successfully decode all data, we
must have m ≤ R1,BS,RS T1 .
• In Phase 2, the BS and RS utilize a distributed space-time
code layered with an incremental redundancy encoding of
Fig. 1. Protocol Structure
the data to transmit the data to the UE.
• The UE uses its (optimal) space time decoder and then
combines the two incrementally redundant transmissions
applicability to the uplink in the concluding remarks. Both to fully decode the data at the end of Phase 2. Note
protocols consist of two basic phases, as shown in Figure 1. that the UE combines data from 2 transmissions to
In Phase 1, the BS transmits. The “primary” purpose of successfully decode. Let R1,BS,U E be the maximal rate
this transmission is to get data to the relay, however the at which the reliable transmission from BS to UE is
UE behaves opportunistically and receives this transmission possible. Let R2,coop be the maximal rate at which
as well. In a forthcoming publication we shall demonstrate reliable transmission to the UE is possible by cooper-
that such an operation carries significant benefits and therefore ation of the RS and BS in Phase 2. Assuming ideal
this option was selected to maximize throughput in all cases incremental redundancy combining, the UE “possesses”
considered. In fact, our conclusions hold even if the UE does R1,BS,U E T1 bits of useful information about the message
not receive in Phase 1, however space limitations do not permit from the first transmission and R2,coop T2 bits of useful
us to present both sets of results. information about the message from the second trans-
In Phase 2 the RS transmits the data which it received in mission. To successfully decode we must therefore have
Phase 1. The behavior of the BS and the UE in Phase 2 m ≤ R1,BS,U E T1 + R2,coop T2 .2
depends on the protocol used. We note that this protocol is similar to that described in
As the central goal of this paper is to compare two schemes: [18] and is, in fact closely related to the standard cooperative
one that emphasizes PHY cooperation and one that emphasizes relaying protocol as enabled by, for example, 802.16j [1]. The
MAC, we shall make a number of simplifying assumptions: main difference between P1 and what is enabled by 802.16j is
• Our transmission occurs in fixed transmission time inter- the presence of opportunistic reception by the UE in Phase 1
vals (TTIs)1 of length T , as shown in Figure 1. These (which would be possible, but difficult to implement within
are partitioned flexibly into T1 for Phase 1 and T2 for the scope of 802.16j). The main difference between P1 and
Phase 2. How the partitioning is done depends on channel the protocol analyzed in [18] is the block-wise definition of
conditions and is discussed below. the protocol and the use of HARQ to incrementally improve
• We shall generally assume that the channel state is not UEs performance over what it can achieve during Phase 1.
known to the transmitter. The maximum amount of data that can be transmitted during
• We shall assume that near-optimal space-time codes are the TTI (time T ) is then given by
available as are near-optimal multi-user receivers (e.g.
m∗ = max min(R1,BS,RS T1 , R1,BS,U E T1 + R2,coop T2 )
SICs). T1 ,T2 :
T1 +T2 =T
• We consider TDM scheduling - i.e. only one UE sched- (1)
uled in any given TTI. The analysis (i.e. comparison) Note that R2,coop > R1,BS,U E (since R2,coop is the BS/RS
clearly extend to OFDMA schemes by simply scheduling cooperation rate). Thus, we maximize (1) by setting
one UE per TTI per sub-carrier (or a group of these). The
case of CDMA is more complex because of inter-code R1,BS,RS T1 = R2,coop T2 + R1,BS,U E T1 (2)
interference. The simulation results we present partially and this rate-balancing equation allows us to determine both
address that issue through the use of an actual (and the split of the TTI into Phase 1 and Phase 2 and the
practical) CDMA channel equalizer. However, in this
2 A formal justification of this argument can be made by recourse to the
1 Thisterminology should not be confused with the definition of TTI in mutual information addition inequality and optimal incremental redundancy
3GPP standards. or rateless coding, see e.g. [22], [21] and references therein
maximal achievable transmission rate. After some algebraic (at the end of the paper). The key difference is the management
manipulation we obtain the maximal achievable rate of cooperation. In Protocol 1, a single flow in transmitted
m∗ R1,BS,RS R2,coop by the MAC during (T1 + T2 ), while Protocol 2 creates and
RP 1 = = (3) transmits 2 MAC flows. Although Protocol 1 might appear
T R1,BS,RS + R2,coop − R1,BS,U E
simpler at first, it does present several challenges. Specifically,
1
= ³ ´ in order to schedule the data, the MAC needs to be aware
1 1 R1,BS,U E
R1,BS,RS + R2,coop 1 − R1,BS,RS of the quality of a “compound” link which consists of the 3
constituent PHY links (BS-to-RS, RS-to-UE and BS-to-UE).
Moving on, we define Protocol 2 (P2) as follows: Moreover, to ensure cooperation between BS and RS in Phase
• The BS creates two messages of m1 and m2 bits. 2, the RS must be centrally scheduled by the BS and the PHY
• In Phase 1, BS transmits the first message (m1 bits) to at the BS and RS must be tightly synchronized (to the channel
the RS at a rate R1,BS,RS , thus m1 ≤ R1,BS,RS T1 . As symbol level).
in P1, the UE listens to this transmission. Protocol 2 is more flexible as it manages the transmission
• In Phase 2 the relay forwards the information it received of the two flows almost independently and without tight PHY-
in Phase 1 to the UE. This is done at a rate R2,RS,U E layer synchronization. The only constraint on the two flows
Simultaneously the BS sends the second message (m2 is that the sum rate at the UE does not exceed its sum-rate
bits) to the UE. This is done at a rate R2,BS,U E . constraint (5). Provided this constraint is satisfied, the BS
• The UE uses an optimal multi-user detector (i.e. SIC) in MAC needs to manage the RS transmission only in a limited
Phase 2 and also optimal incremental redundancy for the fashion. In particular
first message to receive the data. • The BS MAC needs to schedule data to the RS (based
To analyze performance of this protocol, we begin by stating only on the BS-to-RS link quality) to make sure that the
the various constraints which exist. First, as for P1, we wish RS buffer does not become empty.
to transmit the first message efficiently and therefore have the • The BS and RS MACs need to agree how the rates are
following rate-balancing equation partitioned in Phase 2 so that the combined rate to the
R1,BS,RS T1 = R2,RS,U E T2 + R1,BS,U E T1 (4) UE does not violate (5).
• However, the BS MAC does not need to specify to the RS
The rates R2,RS,U E and R2,BS,U E are, however, dependent on MAC which particular packet is to be scheduled for the
each other as well. In addition to satisfying individual per-link transmission. Once the RS indicates reception of a packet,
capacity constraints, these must also satisfy the MAC capacity HARQ management for that packet can be relinquished
constraint to the RS.

R2,RS,U E + R2,BS,U E ≤ R2,coop (5) Therefore, the RS MAC scheduler can act independently from
the BS MAC scheduler with BS control of RS taking place
where we assumed that the rate R2,coop as defined for P1 at a much slower rate. Finally, we note that the PHY layer
is indeed the optimal transmitter cooperation rate. Although operations of Protocol 2 require no coordination since the BS
cooperation at the PHY layer is not part of P2 (5) illustrates and RS simply transmit different flows in Phase 2 in a non-
the close relationship between achievable throughput for P1 cooperative fashion.
and P2. Clearly, maximizing the throughput would require (5)
to be satisfied with equality. Taking this, together with (4) and III. S IMULATION M ETHODOLOGY AND R ESULTS
the constraint T = T1 + T2 we get (after some algebra):
The simulation methodology we use is the same as in [9].
m1 + m2 R1,BS,RS R2,coop − R1,BS,U E R2,BS,U E We use a modified form of the RUNE simulator to which
RP 2 = =
T R1,BS,RS + R2,RS,U E − R1,BS,U E HSDPA and relaying capability has been added [7]. RUNE is
(6) a set of Matlab functions that allow the creation of hexagonal
An analytical comparison on (3) and (6) would require an cellular networks with fixed or random placement of terminals.
analysis under several different assumptions on the relation- The simulator determines the SINR for each base station
ships between the various rates - and the results would depend and terminal based on the network topology, shadow fading
on these assumptions. For example, one can demonstrate that profile, and scheduling algorithm. In addition to the SINR
P2 can deliver significant advantages over P1 if the BS-to- with the interference treated as AWGN, the modified RUNE
UE channel improves significantly in Phase 2 as compared to simulator calculates the dominant interference profile (DIP1 )
Phase 1. In this paper, we present a system-simulation based for the strongest interfering signal as this has been shown to
study and demonstrate that in typical interference limited impact performance of HSDPA receivers in a way substantially
cellular deployments P2 provides slightly better performance different from AWGN.
then P1 while both provide a significant improvement over a In [24] the DIP ratios were defined as:
no-relay case.
From protocol architecture point of view, the two ap- Iˆori
proaches are quite different, as illustrated in Figures 8 and 10 DIPi = , (7)
Ioc
PNB
where Ioc = j=1 Ior,j + N0 . The power of the interfering BS) represents high IIor
oc
, and deep blue (deep color close to cell
base station signals is Ior,j where Ior1 > Ior2 > · · · . The boundary) represents low IIor oc
. As can be seen, the greater the
number of interfering base stations is NB . Note that the power distance between the UE and BS, the lower the IIor . Around
from the serving cell, IˆorB , is not included in the DIP 20% of a cell, mostly at the cell edge, has IIor
oc
below 0 dB,
calculations. Accordingly, the system simulator uses a two- oc

ˆ and the IIor can be as low as -4 dB.


dimensional lookup table (for IIoroc
and DIP1 ) to determine the oc

instantaneous throughput for each UE location. These through-


put values are averaged over a large number of fast fading
400
6

realizations to determine the typical throughput for a particular 300


5.5

scenario. The number of UEs within a cell can be varied to take 200 5

into account the effects of intelligent scheduling. The link level 100 4.5

throughput used is typical for a receiver diversity equipped 0


4

chip-level equalizer that meets the applicable requirements in −100


3.5

[23] with appropriate implementation margins. 3

−200 2.5
The system model is shown in Fig. 2. Here red circles 2
denote base stations (BSs), and green squares denote relay −300
1.5

stations (RSs). A typical set of parameters for a cellular system −400


0 100 200 300 400 500 600 700

Ior
Fig. 3. Ioc
Contour for an HSDPA System
1000

UEs are assigned to cells on a long-term basis (for the


500
duration of the simulation). However, within a cell, UEs can be
0
assigned to either use a relay or to forgo a relay dynamically.
The latter process, called intra-cell assignment, operates as
−500
follows. A single UE is served per time slot. For a specific UE
−1000
in a system with L RSs, data can reach the UE via L+1 paths
(see Figure 4), BS-UE, BS-RS1-UE, · · · , and BS-RSL-UE,
−1500 −1000 −500 0 500 1000 1500 where L is the number of RSs. Note that, although diversity
combining techniques can be exploited in the UE, we restrict
Fig. 2. HSDPA System with RSs
ourselves in this paper to selecting only a single RS.
simulation (and one we use here) are listed in Table I [24]3 ,
where R is the distance between BS and UE; θ is defined as 400

300
Parameter Assumptions RS5
Cell Radius 333.33 meters 200

Propagation Model 128.1 + 37.6Log10(R) RS4

Shadowing std. dev. 0 dB 100

MCL 70 dB UE

0 BS RS3
BS antenna gain 14 dB
1

· ³ ´2 ¸
BS antenna pattern A(θ) = − min 12 θ θ , Am −100
3dB RS2
BS total TX power 43 dBm −200

UE antenna gain 0 dBi RS1

UE noise figure 9 dB −300

RS antenna gain 9 dB
−400
RS noise figure 9 dB 0 100 200 300 400 500 600 700

TABLE I Fig. 4. Available Data Paths for a UE


PARAMETERS FOR EVALUATION OF SYSTEM LEVEL
For the results presented here we use 5 RSs evenly spread
the angle between the direction of interest and the boresight of
across the 120-degree sector along the contour line IIor =
the antenna, and −180o ≤ θ ≤ 180o ; θ3dB represents the 3 dB oc
3 dB. For most results we assume free-space propagation
beamwidth in degrees, and is set to 70 degrees; Am denotes
between BS and RS while BS-UE and RS-UE links have the
the maximum attenuation. Front-to-back ratio, Am , is set to
propagation model listed in Table I. This is a fairly good, if
20 dB. Notice here RSs are equipped with omnidirectional
a bit simplistic, model for an above-the-rooftop (ART) relay
antennas. Without loss of generality, we will study a specific
placement. We do note (see Figure 9 and accompanying text)
cell (inside the blue square).
that the benefits we observe are also present in the below-
Fig. 3 shows an IIor contour plot for a conventional HSDPA
oc the-rooftop (BRT) placement of the relay, where the BS-RS
system (without RSs). In Fig. 3, deep red (deep color close to
propagation is also taken according to Table I.
3 Note that cell radius of 333.33 meter means that the distance from BS to We begin with contour plots for the two protocols (with
the farthest cell corner is 666.66 meters ART relay), shown in Figures 5 and 6. We observe that
400 the downlink, the shift in complexity to the receiver makes
6
300 Protocol 2 a more natural choice for the uplink. Indeed,
200
5
in the uplink, it is hard to imagine the UE coordinating
100 4 scheduling with the RS, while having the BS do so can become
0
cumbersome and expensive (in terms of signaling overhead).
3

−100
2
R EFERENCES
−200

1
[1] IEEE Lan/Man Standards Committee. Multihop Relay Specification:
−300
Draft Amendment to IEEE Standard for Local and Metropolitan Area
−400 0 Networks: Part 16: Air Interface for Fixed and Mobile Broadband
0 100 200 300 400 500 600 700
Wireless Access System. IEEE P802.16j/D5, May 2008.
[2] IEEE 802.16 Broadband Wireless Access Working Group. The Draft
Ior IEEE 802.16m System Description Document, Jul. 2008.
Fig. 5. Ioc
Contour for an HSDPA System with Protocol 1 and ART Relay [3] 3GPP RAN1, Tdoc R1-083410, ”Text Proposal for RAN1 TR on LTE-
Advanced,” August, 2008.
400 [4] 3GGP RAN1, Tdoc R1-050617, ”Throughput Evaluations Using MIMO
300
6 Multiplexing in Evolved UTRA Downlink,” NTT DoCoMo, 2005.
[5] R. Tafazolli, editor, Technologies for the Wireless Future: Wireless World
200 5 Research Forum (WWRF), Vol. 2, Wiley, New York, 2006.
100
[6] F. H. P. Fitzek and M. D. Katz, editors. Cooperation in Wireless
4
Networks: Principles and Applications. Real Egotistic Behavior is to
0 Cooperate! Springer, Dordrecht, The Netherlands, 2006.
3
[7] J. Zander and S.-L. Kim. Radio Resource Management for Wireless
−100
2
Networks. Artech House. Bostom, MA. 2001.
−200 [8] T. M. Cover and A. A. El Gamal. Capacity theorems for the relay
1
channel. IEEE Tr. Inf. Theory, IT-25:572–584, Sep. 1979.
−300
[9] R. Hu, S. Sfar, G. Charlton, and A. Reznik. Protocols and system
−400 capacity of relay-enhanced hsdpa systems. In Proc. Conference on Inf.
0 100 200 300 400 500 600 700
Sciences and Systems (CISS), pages 179–184, 2008.
[10] E. Zeira and A. Reznik. ”Relaying Strategies for 802.16m Multi-Hop
Ior Relays,” IEEE 802.16 Broadband Wireless Access Working Group,
Fig. 6. Ioc
Contour for an HSDPA System with Protocol 2 and ART Relay C802.16m-Relay-08/026, July 2008.
[11] G. Kramer, M. Gastpar, and P. Gupta. Cooperative strategies and
capacity theorems for relay networks. IEEE Tr. Inf. Theory, IT-51:3037–
both protocols appear to deliver significant gains, including at 3063, Sep. 2005.
cell edge, and the performance appears to be comparable. To [12] J. N. Laneman. Coperative Diversity in Wireless Networks: Algorithms
better compare the two, we show, in Figure 7 the CDF of the and Architectures. PhD thesis, Massachusetts Institute of Technology,
Cambridge, MA, 2002.
achieved throughput over the cell (with ART). We observe that [13] J. N. Laneman and G. W. Wornell. Distributed Space-Time Coded
both protocols outperform significantly the no-relay baseline Protocols for Exploiting Cooperative Diversity in Wireless Networks.
with Protocol 2 delivering better overall performance. The IEEE Tr. Inf. Theory, IT-49:2415-2425, Oct. 2003.
[14] J. N. Laneman, D. N. C. Tse, and G. W. Wornell. Coperative diversity
same holds also in the BRT case (Figure 9), although the in wireless networks: Efficient protocols and outage behavior. IEEE Tr.
overall gains are reduced. Inf. Theory, IT-50:3062–3080, Dec. 2004.
[15] A. Sendonaris, E. Erkip, and B. Aazhang. User cooperation diversity -
IV. C ONCLUDING R EMARKS part I: System description. IEEE Tr. Comm., 51:1927–1938, Nov. 2003.
[16] A. Sendonaris, E. Erkip, and B. Aazhang. User cooperation diversity
In this paper we have argued that effective cooperative - part II: Implementation aspects and performance analysis. IEEE Tr.
Comm., 51:1939–1948, Nov. 2003.
protocols can be implemented with cooperation managed at [17] H. Ochiai, P. Mitran, H. V. Poor and V. Tarokh. Collaborative Beam-
either the PHY or the MAC layer. To that end we demonstrated forming for Distributed Wireless Ad Hoc Sensor Networks. IEEE Tr.
two highly effective relaying protocols - one with cooperation Sig. Proc., 53:4110–4124, Nov. 2005.
[18] R. U. Nabar, H. Bolcskei and F. W. Kneubuhler. Fading Relaying
at the PHY and one at the MAC. We showed that moving Channels: Performance Limits and Space-Time Signal Design. IEEE
the management of cooperation to the MAC, provided it is Jour. Sel. Areas Comm., 22:1099–1109, Aug. 2004.
done “correctly,” does not degrade performance - and may [19] Z. Lin, E. Erkip and M. Ghosh. Rate adaptation for coded cooperative
systems. in Proc. GLOBECOM, 2006.
even improve it (as it does in our case). Further results and a [20] P. Liu, Z. Tao, Z. Lin, E. Erkip and S. Panwar. Cooperative wireless
more detailed analysis of cooperation schemes will be made communications: A cross-layer approach. IEEE Wireless Comm. Mag.,
available in a forthcoming publication. 13:84–92, Aug. 2006.
[21] S. Sfar, et. al. Is Collaborative Communications Worth it? Proc. 42nd.,
In comparing the two protocols introduced here, we further Asilomar Conf. on Signals, Systems and Computers, 2008.
note that in going from Protocol 1 to Protocol 2 we have not [22] U. Erez, M. D. Trott and G. W. Wornell. Rateless Coding for Gaussian
only shifted the emphasis from the PHY to the MAC - but have Channels Submitted to IEEE Tr. Inf. Theory, Aug. 2007.
[23] 3GPP TS 25.101 v7.8.0, “3rd Generation Partnership Project; Technical
also shifted complexity from the transmitter to the receiver. Specification Group Radio Access Network; UE Radio Transmission
Specifically, Protocol 1 requires transmitter cooperation, tight and Reception (FDD) (Release 7),” June 2007.
scheduling and distributed space-time coding. In contrast, [24] 3GPP RAN4, Tdoc R4-060364, “Minutes of Ad Hoc on Further
Improved Performance Requirements for UMTS/HSDPA UE (FDD),”,
Protocol 2 requires a potentially more complex receiver, but Nokia, 2006.
little transmitter complexity beyond what would normally exist
in a non-cooperative setting. While both are viable options for
100 100

90 90

80 80

70 70

60 60
Protocol 1

CDF (%)
CDF (%)

50 50 Protocol 1
Protocol 2

40 40 No relay
No relay Protocol 2

30 30

20 20

10
10

0
0 1 2 3 4 5 6 7
1 2 3 4 5 6 7
Throughput in Mbps Throughput (Mbps)

Fig. 7. Throughput CDF for ART Relay Fig. 9. Throughput CDF for BRT Relay

Fig. 8. Protocol 1 Architecture Fig. 10. Protocol 2 Architecture

Vous aimerez peut-être aussi