Escolar Documentos
Profissional Documentos
Cultura Documentos
Abstract—To meet the capacity requirements of the ex- The physical location of BBUs can be in a central office
ponential increase in mobile traffic and to continue to drive
down the per-bit cost for mobile service providers, the (CO) that stretches tens of kilometers from the RRUs, but
cloud radio access network (C-RAN) has become a crucial is connected through a reliable optical network, namely,
step toward 5G. However, the current C-RAN architecture the fronthaul. Although pooling BBUs in a centralized
has some major drawbacks in terms of scalability, cost, and location lowers maintenance cost and provides the benefit
efficiency. In this paper, we propose and numerically dem-
onstrate a novel mobile fronthaul architecture based on of shared resource pooling, deploying a fronthaul network
functional split and time-division multiplexed (TDM) pas- is very costly because sending quantized samples of analog
sive optical networks (PONs) with a unified mobile and wireless signals from a RRU to a BBU requires not only
PON scheduler known as Mobile-PON. The optimal func- high bandwidth but also very low latency.
tional split distributes lower physical layer hardware to-
ward remote radio sites. The new interface that divides Current CPRI application on mobile fronthaul is band-
the remote radio processing, and centralized baseband width inefficient, e.g., often requiring 30 bits to transmit a
processing requires less bandwidth, also opens the pos-
sample [4]. The transmission latency should also not ex-
sibility of sharing and multiplexing the bandwidth with
multiple remote sites. Our combined mobile and PON ceed 250 μs as recommended by the Next Generation
scheduler is mainly based on the more complex wireless Mobile Networks Alliance (NGMA) [5]. Because of such
scheduling that translates its results into the TDM-PON high requirements, only expensive solutions such as WDM-
system through LTE resource block mapping, eliminating PON can fully support CPRI-based C-RAN fronthaul.
additional scheduling delay at the PON. Without the addi-
tional scheduling delay, the cost-effective TDM-PON To increase the bandwidth efficiency of mobile fronthaul
becomes applicable for mobile fronthaul, while the optimal and to better support future generations of mobile systems
fronthaul interface increases bandwidth efficiency by ∼10×
over CPRI. such as 5G, the Next-Generation Fronthaul Interface
(NGFI) [6] has been introduced. In NGFI, various func-
Index Terms—CPRI; C-RAN; DBA; Fronthaul; Functional tional splits are being defined to provide different trade-
split; Mobile scheduling; TDM-PON. offs among RRU complexity, system performance, and
fronthaul bandwidth efficiency. Normally higher level split
interfaces, which distribute more functionality toward
I. INTRODUCTION RRUs, would increase bandwidth efficiency. However, that
would also kill the advantages of centralized baseband ra-
dio processing. A carefully engineered fronthaul interface
T he cloud radio access network (C-RAN) has become a
crucial mobile access architecture, enabling new wire-
less technologies such as coordinated multi-point (CoMP)
is required to maximize the benefit of centralization while
constraining the fronthaul networking requirement.
advanced intercell interference coordination (ICIC), mas- In addition to fronthaul functional split, time-division
sive multiple-input multiple-output (MIMO), etc. [1–3]. multiplexed (TDM) passive optical networks (PONs) have
A typical C-RAN divides the entire function of a traditional been proposed to reduce the implementation cost of CPRI-
base station (BS) into two parts: remote radio units (RRUs) based mobile fronthaul via the sharing of the same optical
and baseband units (BBUs). RRUs are distributed across distribution network (ODN) by multiple RRU sites [7,8].
cell sites handling all wireless signal transmission, and pass The main challenge here is to find a solution to overcome
quantized samples of analog wireless signals via the the high medium access latency of TDM-PON.
Common Public Radio Interface (CPRI) to the BBUs for fur- In this paper, we propose a high-bandwidth-efficiency
ther processing. The hardware function of RRUs therefore and low-processing-latency mobile fronthaul architecture
becomes very simple, while BBUs take care of the rest of based on physical layer functional split and TDM-PON
the physical layer processing and higher layer functions. with a unified mobile scheduler, which we refer to as
Mobile-PON. The use of a unified mobile scheduler elimi-
Manuscript received July 3, 2017; revised September 5, 2017; accepted nates the need for conventional PON scheduling. Through
September 12, 2017; published October 20, 2017 (Doc. ID 301360). simulations, we show that fronthaul bandwidth efficiency
S. Zhou (e-mail: sz1596@nyu.edu) and J. Chao are with New York can be increased 10-fold with no additional processing
University, Tandon School of Engineering, Brooklyn, New York 11201, USA.
X. Liu and F. Effenberger are with Huawei Technologies, Bridgewater,
latency for PON scheduling. We focus our discussion in this
New Jersey 08807, USA. paper on uplink, since most of the challenging issues are
https://doi.org/10.1364/JOCN.10.000A20 only for uplink but do not exist for downlink.
The rest of this paper is divided into four sections. In BBU at the peak rate as if the wireless interface is always
Section II we present some existing efforts in redefining func- under full load, even when the antennas are receiving noth-
tional split interfaces to reduce fronthaul bandwidth and our ing, because the RRU functionality is so simple that it can-
own position in this discussion. Section III introduces the not distinguish whether the received signal is noise or data.
concept of a cost-effective Mobile-PON based on an optimal To relax the fronthaul bandwidth requirement, one reason-
functional split scheme and discusses the specific operation able split interface is between MAC and PHY (MAC-PHY
details of the unified scheduling in Mobile-PON, including split). Rather than sending IQ samples, LTE MAC frame data
the static and dynamic mapping schemes. Section IV shows would be communicated through the mobile fronthaul, and,
a proof-of-concept simulation verification of the design. therefore, a MAC-PHY split significantly reduces the optical
Finally, Section V concludes the paper. bandwidth requirement [10]. However, a MAC-PHY split also
eliminates key CoMP functions such as joint reception and
multi-cell centralized MIMO processing, because all PHY
II. NEW FUNCTIONAL SPLIT INTERFACES functions are distributed at the RRU. The LTE MAC frames
are decoded wireless data. Other advanced wireless concepts,
The general fronthaul functional modules for the uplink such as user centric networking (UCN), which erases the tra-
direction are shown in Fig. 1. The eNode user plane protocol ditional cell boundary, also depends on centralized PHY radio
stack can be divided into the following layers: packet data processing. Therefore, a simple MAC-PHY split is unsatisfac-
convergence protocol (PDCP), radio link control (RLC), tory given such requirements.
medium access control (MAC), and physical (PHY) layer. Another promising split, as shown in Fig. 1(b), is within
The PHY layer is further divided since it is responsible the PHY layer between the wireless resource demapping
for most of the radio processing. The conventional CPRI and the modulation decoding, which offers two major ben-
is the most aggressive functional split interface, as shown efits: (1) reduced quantization resolution requirement and
in Fig. 1(a), where almost all PHY layer and higher layer (2) increased transmission efficiency by transmitting only
functions are centralized, and RRU signal processing is kept the populated wireless resource blocks (RBs). If the uplink
at bare minimum. It needs only to sample and quantize is based on the actual wireless RB utilization, we can take
time-domain wireless signal waveforms, and digitized radio advantage of the tidal effect [1] with additional multiplex-
waveforms are communicated through the mobile fronthaul. ing gain by sharing the bandwidth when multiple RRUs
This scheme provides the maximum benefits of centraliza- experience different load at different times. Compared to
tion, shared resource pooling, and best architectural support conventional CPRI, which samples at the time domain with
for CoMP. Since all PHY radio processing is at the central- high PAPR, a PHY split allows us to sample at the fre-
ized BBU, RRUs can be taken as extended antennas that can quency domain with seven to eight fewer quantization bits
jointly receive and transmit to the same user from different [9]. More importantly, a PHY layer functional split at RB
cell sites, bringing cell edge and MIMO performance to a demapping may still preserve the possibility of CoMP.
new level. However, to support such a system, the mobile Wireless channel decoding is the last PHY layer function
fronthaul must bear a huge constant transmission rate that in an uplink, and it uses the log-likelihood ratio (LLR) data
is independent of the actual wireless data volume. Because output from the wireless demodulation function to retrieve
of the high peak-to-average power ratio (PAPR) of orthogo- the received bit data, so, as long as the final decoding func-
nal frequency-division multiple access (OFDMA), 15 bits are tion is centralized, joint reception and multi-cell MIMO
required to quantize each sample, and the fronthaul data could still be possible [10]. Since most PHY layer hardware
transmission can be as much as 30 times more than the is difficult to share, distributing some PHY hardware func-
actual data throughput on the wireless interface [9]. tions while keeping higher layer functions centralized does
Furthermore, a RRU also needs to always upload to a not decrease the benefit of hardware resource pooling.
In short, redefining the split interface at RB demapping
still preserves the major benefits of C-RAN, while it signifi-
cantly alleviates fronthaul bandwidth requirements. With
the functional split interface shown in Fig. 1(b) as a foun-
dation, we developed the Mobile-PON, a low-latency mobile
fronthaul based on-cost effective TDM-PON.
III. MOBILE-PON
the relative transparency is only on the data plane. Basic support all cells operating at full load. However, the actual
PON system setup is still required for mobile-PON. The op- cellular load varies significantly with respect to the location
tical propagation delay from different remote sites to the and time, and peak load happens at rare times of a day.
centralized side may be different, so each remote unit must Therefore, the multiplexing gain would be significant if
be synchronized such that all TBs from different sites reach residual cellular bandwidth can be shared with non-
the central OLT interface at the proper time [15]. fronthaul data transfers, such as fiber-to-home or edge cloud
Fortunately, current TDM-PON technology can obtain communications. Dynamic mapping opens the possibility of
precise synchronization timing of only 125 ns. sharing those unmapped TBs for other data transfer usage.
A notable property of Mobile-PON is that RB to TB map-
B. Dynamic Mapping ping requires extremely short laser bursts, as anticipated in
Ref. [16] for future PON technologies. OFDMA is a two-
Although current optical technology can support Mobile- dimensional multiplexing scheme for both time and fre-
PON, the static mapping scheme has limitations and is in- quency. If all RBs are to be mapped to a one-dimensional
sufficient to handle complex situations. The RB to TB map- time-multiplexed TB within the same time frame, the dura-
ping does not have to be fixed at all times. For example, the tion of a single TB should be much shorter than that of a RB,
total number of TBs must be greater than or equal to the depending on the spectral frequency. For example, at 20 MHz
total number of possible RBs for static mapping to work. It there are 100 RBs stacked in the 0.5 ms duration of a RB.
also requires that no RRU uses the same uplink RB within Therefore, for a single carrier the short burst should be no
the frequency reuse area. Normal wireless operation meets more than 5 μs. The authors of Ref. [17] have already exper-
this requirement but not advanced CoMP features, such as imentally demonstrated that the 10G PON can achieve
joint reception and inter-cell MIMO, where at the cell edge microsecond level packet bursts with only 25.6 ns guard time.
a UE could be communicating with two RRUs using the However, given the speed of fiber, a mere 5 μs TB may still
same RB. Also, TBs cannot be shared across RBs from dif- have the capacity of several RBs. For example, in 10G PON, a
ferent carriers. To overcome such limitations, we developed 5 μs slot can transmit about 50 kb of data, which is 5–10
a more sophisticated dynamic mapping scheme. times the capacity of a RB. To efficiently use the entire optical
bandwidth, we should pack multiple RBs from the same
Under dynamic mapping, the RB mapping rule can be
RRU to one TB. There are two requirements when packing
changed dynamically. If two RRUs are using the same RB
multiple RBs to one TB. First, the RBs must all belong to the
on the same carrier (e.g., in the case of CoMP) or when a
same RRU; otherwise, signals from different remote sites
BBU makes a new wireless scheduling decision, the BBU re-
would interfere with each other if packed to the same TB.
calculates mapping rules dynamically such that no two
Second, the RBs must all be at the same time slot so that
ONUs use the same TB, as shown in Fig. 5. For instance,
there would be no additional waiting delay in collecting all
RB16 does not have to map to T16 as in static mapping.
RBs for the TB pack. The flexible multiplexing and RB pack-
At time frame 1 it could map to T2, while it could also
ing schemes are only possible in dynamic mapping mode, but
map to other TBs at different time frames. However, dynamic
had not been discussed in the previous work [18].
mapping needs to make sure that none of the remote sites
uses the same TB simultaneously. The process will not incur In a nutshell, dynamic mapping is much more flexible
extra time delay because a new wireless scheduling decision than static mapping. A good mapping scheme can optimize
needs to be updated with UEs anyway. Thus, updating ONUs PHY layer latency and bandwidth. Here we provide an
based on the new mapping information involves no additional elementary mapping that minimizes PHY layer delay
timing overhead. When the total load is small, dynamic map- and bandwidth waste. Given that TBk is the kth PON
ping can also minimize latency by assigning those occupied TB for the timing interval of subframe 1, TBk ∈ Sub1 .
RBs to the earlier TBs, and assigning empty RBs to later TBs. RRUn is the nth RRU served under the BBU. RBn;c;i repre-
sents the ith RB of carrier c served by RRUn. Let Cn be the
Dynamic RB to TB mapping also provides higher flexibil-
total number of carriers on RRUn . At the peak wireless
ity for bandwidth multiplexing. Normally the hardware in-
data rate, the throughput on mobile-PON would be
frastructure is initially provisioned such that it can
C1 X
X bw C2 X
X bw Cn X
X bw
Peak Rate D D D; (1)
c
JRj Rbn;c;i ; Rb n;c;i ; (2)
Rbn;c;i → Tbk ; c
Rbn;c;i → Tbk1 : (3)
TABLE I
LTE SIMULATION PARAMETERS FOR EVALUATING MOBILE-PON
Uplink bandwidth 10 MHz
Bandwidth/channel (RB) 180 kHz
Subcarriers/channel (RB) 12 subcarriers
Resource element 7 per subcarrier
Frame duration 10 ms
Sub-frame cycle 1 ms (2RBs/frame)
RB per time slot 50 RB/0.5 ms
RB per frame 50 × 20 1000
Fig. 7. Mobile-PON TBs scheduled by (a) static mapping and
Duplex FDD
(b) dynamic mapping.
Zhou et al. VOL. 10, NO. 1/JANUARY 2018/J. OPT. COMMUN. NETW. A25
Fig. 8. Average serialization delay per RB versus average load Fig. 9. Average latency per RB with respect to the number of
(assuming 10G PON). remote units.