Professional Documents
Culture Documents
Abstract—Time division multiplexing passive optical radio-over-fiber link known as front-haul [2] [e.g., Common
network (TDM-PON) technologies are viewed as an attrac-
Public Radio Interface (CPRI)]. To connect RRHs in multi-
tive solution for flexible and cost-efficient mobile front-
haul for cloud radio access network (C-RAN) architecture. ple locations to the BBU pool, a lot of high-capacity fiber-
However, it is a challenge for TDM-PONs to meet the based front-haul connections are required, which increases
strict latency requirement of mobile front-haul in C-RAN the deployment cost of C-RAN. A functional split of mobile
because they use a dynamic bandwidth allocation (DBA) baseband processing [3] provides a way to address the cost
mechanism to manage upstream traffic. The latency issue
of TDM-PON based mobile front-haul has been extensively issue of fiber-based front-haul, as it reduces the capacity re-
investigated in the literature with particular focus on IEEE quired for front-haul, allowing the usage of low cost and
TDM-PON (e.g., 10G-EPON). However, ITU TDM-PONs such shared front-haul transport technologies such as switched
as XG-PON and XGS-PON have not yet even been explored Ethernet [4,5], OTN architecture [6], and time division multi-
in the context of mobile front-haul. To cover this gap, we
plexing passive optical networks (TDM-PONs) [7]. In this
first evaluate the performance of two recently proposed
XG-PON-standard-compliant DBAs, namely, group-assured study we focus on TDM-PONs, as adopting the other two
GIANT (gGIANT) and round-robin DBA (RR-DBA), over technologies in the access network may need a long time
simulated mobile front-haul traffic. We conclude based on to become feasible.
our evaluation that neither RR-DBA nor gGIANT satisfies
the delay required for mobile front-haul. Therefore, we pro- TDM-PONs are a promising technology for realizing
pose an optimized DBA known as optimized round-robin cost-efficient mobile front-haul transport as they allow
(optimized-RR) that support front-hauling over XG-PONs. the sharing of optical fibers and transmission equipment.
The performance evaluation of our optimized DBA against
gGIANT and RR-DBA shows significant improvement in
However, the latency in TDM-PON upstream transmission
upstream delay and utilization as well as lower packet loss due to the dynamic bandwidth allocation (DBA) remains a
and jitter for front-haul uplink traffic transmitted via big challenge, as it is much higher than the front-haul la-
XG-PON in virtualized C-RAN architecture. tency tolerance in C-RAN (i.e., 300 μs [8]). There are many
DBA proposals in the literature for low-latency mobile
Index Terms—C-RAN; Dynamic bandwidth allocation;
Mobile front-haul; Performance evaluation; XG-PON; front-haul over IEEE TDM-PONs (e.g., 10G-EPON).
XGS-PON. However, ITU TDM-PONs (e.g., XG-PON) still remain un-
explored in this context. In this study, we evaluate in terms
of upstream delay the performance of some recently pro-
posed XG-PON-standard-compliant DBAs over simulated
I. INTRODUCTION mobile front-haul traffic. We also propose an optimized low-
latency DBA to accommodate dynamics of mobile front-
haul traffic in C-RAN architecture.
minimize the upstream latency of 10G-EPON based mobile III. XG-PON BASED MOBILE FRONT-HAUL FOR
front-haul has been proposed in Ref. [12]. However, this ap- VIRTUALIZED SMALL-CELL C-RAN
proach requires an additional interface between BBUs and
the OLT to exchange mobile scheduling information.
A. Virtualized Small-Cell C-RAN Employing
Another study in Ref. [13] has introduced the usage of a
simple statistic function along with FBA to accommodate XG-PON Front-Haul Transport
dynamic data-rate front-haul traffic over 10G-EPON.
However, as this approach is based on periodical estima- In conventional small-cell base station (BS) architecture,
tion of RRH traffic, it is difficult for it to accurately capture a transport network known as backhaul [Fig. 1(a)] is
very low scale variation in front-haul traffic at every RRH. used to connect small cells to the mobile core network. By
A recent DBA proposal reported in Ref. [14] has also fo- introducing the concept of small-cell virtualization, the
cused on minimizing the upstream latency of 10G-EPON processing functionalities of a small-cell base station are
based front-haul by reducing the DBA cycle length through decomposed and shared between virtual BBU (vBBU) and
reduction of the DBA grant processing time. However, DBA RRH entities (e.g., MAC-PHY decomposition or the func-
algorithms for IEEE TDM-PONs are not fully compatible tional split between the MAC layer (L2) and the PHY layer
with ITU TDM-PONs due to the fact that IEEE TDM- (L1) of the LTE baseband) [Fig. 1(b)]. As a result, the
PONs are neither synchronous nor have a fixed frame front-haul interface is also introduced to transport mobile
length like ITU TDM-PONs [15]. On the other hand, traffic results from small-cell base station decomposition
ITU PONs have a brief transmission timing scale of 125 [Fig. 1(b)].
μs and support frame fragmentation, which allows them
The typical network topology of virtualized small-cell
to achieve a much lower level of transmission latency than
C-RAN architecture employing XG-PON front-haul trans-
in IEEE PONs [16].
port is illustrated in Fig. 2. In this architecture, RRHs are
All the above-mentioned reasons have motivated us to connected to ONUs, which are placed together with RRHs
investigate the feasibility of front-hauling over XG-PON. at the remote cell site. ONUs are connected using shared
Since there are no other works considering XG-PON, we optical fibers and a passive splitter to the OLT. The OLT is
initially evaluated the performance of two XG-PON-
standard-compliant DBAs over simulated mobile front-
haul traffic using the network simulator NS-3 [17]. The
first DBA is round-robin DBA (RR-DBA) [18], which is
the simplest DBA compliant with the XG-PON-standard
found in the literature, and the second one is group-assured
GIANT (gGIANT) DBA, which has been proposed in
Ref. [19] to optimize XG-PON to transport mobile backhaul
traffic. Based on our initial evaluation, we conclude that
(1) neither RR-DBA nor gGIANT satisfies the latency re-
quired for mobile front-haul and (2) in spite of the fact that
RR-DBA shows better delay performance compared to
gGIANT, it still suffers the drawback of poor utilization
of XG-PON upstream bandwidth under a bursty front-haul
traffic condition. To this end, we introduce optimized-RR
DBA to tackle this problem. Then, we evaluate its perfor-
mance against gGIANT and RR-DBA in a virtualized
small-cell C-RAN architecture. We further introduce an
application use case of optimized-RR DBA to support
5G-ultra-reliable low-latency communication (URLLC) ser- Fig. 1. Small-cell transport network: (a) conventional small-cell
vices over a 100 MHz 5G channel in distributed C-RAN BS with backhaul and (b) virtualized small cell with front-haul
architecture employing XGS-PON front-haul. and backhaul.
then connected to the vBBU pool (GPP servers) to enable C. Statistical Multiplexing of Mobile Uplink Traffic
front-haul data to be exchanged between vBBU-pool and in XG-PON Based Front-Haul Transport
small-cell RRHs. The OLT and vBBU are placed at the cen-
tral office (in the macro cell base station site) and connected
In an XG-PON front-haul network connecting RRHs and
to the mobile core network via a backhaul network.
vBBUs (Fig. 2), the mobile downlink traffic from the vBBU
(XG-PON downstream traffic) is broadcasted by the OLT to
B. Traffic Model for Mobile Uplink Data That all RRHs via ONUs that share the same XG-PON network
Transmits via XG-PON Front-Haul Transport medium, while mobile uplink traffic (XG-PON upstream
traffic) from RRHs is multiplexed by the OLT using the
time division multiple access (TDMA) scheme. To avoid
Virtualized small-cell mobile front-haul data traffic upstream collisions between multiple RRHs transmitting
can be modeled using the same model for backhaul traffic at the same time, XG-PON uses the DBA mechanism to
when functional split architecture (e.g., MAC-PHY split) is control upstream transmission.
considered between BBUs and RRHs because in this case
front-haul data traffic varies dynamically according to the Due to the fact that XG-PON is very common in FTTH
LTE user’s traffic similar to backhaul traffic. The Next access networks, most XG-PON-compliant DBA algorithms
Generation Mobile Networks (NGMN) alliance has devel- were designed to accommodate near-deterministic traffic
profiles that depict fixed broadband last-mile users.
oped a model to estimate backhaul traffic in small-cell LTE
However, in case the last-mile user is a mobile base station
eNodeBs [20]. In this model, small-cell base station back-
(i.e., small-cell RRH), the mobile uplink traffic transmit-
haul traffic is characterized using two different loading
ting via XG-PON front-haul in the upstream exhibits a
conditions that depend on the spatial distribution of mobile
high degree of temporal variation or burst-ness (as we
users within the cell, which exhibits a heavy-tailed pattern
mentioned earlier). Therefore, an advanced bandwidth
[21]. These two loading conditions are busy times and quiet
allocation scheme is required for efficient statistical multi-
times. During busy times, there are many users sharing the
plexing of XG-PON upstream bandwidth between RRHs,
cell’s spectral resources. Meanwhile, during quiet times,
while accommodating the burst-ness of mobile front-haul
there may often be only one user accessing the cell and uti-
uplink traffic.
lizing the entirety of the cell’s spectral resources. In a
densely deployed small-cell C-RAN environment, the num-
ber of cells is large, and the backhaul traffic results from IV. DYNAMIC BANDWIDTH ALLOCATION OF FRONT-HAUL
aggregation of a large number of cells with busy times UPLINK TRAFFIC TRANSMITTING VIA XG-PON UPSTREAM
and quiet times, and formulates on–off processes with
heavy-tailed on-times or off-times. The DBA mechanism in XG-PON is used to assign an
It has been shown in Ref. [22] that the aggregation of a upstream transmission opportunity to the ONU queues
large number of on–off processes with heavy-tailed on- (i.e., transmission containers or T-CONTs). For QoS pur-
times or off-times results in long-range dependence. The poses, XG-PON defines four different bandwidth types
existence of long-range dependence in real-world LTE associated with T-CONTs, namely, fixed (T-CONT-1), as-
and LTE-advance base station backhaul traffic has also sured (T-CONT-2), non-assured (T-CONT-3), and best-effort
been confirmed in Ref. [23]. Reference [24] shows that (T-CONT-4). The DBA operations in XG-PON start by send-
the Poisson Pareto burst process (PPBP) can be used to ing initial grants in the downstream frame from the OLT to
generate a long-range dependent arrival process that the ONU/T-CONTs, allowing them to report their buffer oc-
matches statistical properties of real-life network traffic. cupancy. After receiving the buffer occupancy reports from
Therefore, in this study we adopt PPBP as a traffic source the ONU/T-CONTs, the DBA engine selects the T-CONT to
to simulate the burst-ness of front-haul data traffic trans- be served, determines the size of grant allocation for the T-
mitting between vBBUs and RRHs (Fig. 2) in both uplink CONT, calculates the start time for each grant allocation,
and finally broadcasts these decisions to the ONUs through
and downlink. Table I shows the parameters we use to gen-
a header in the downstream frame that is updated at regu-
erate the PPBP traffic that is used in our simulations in
lar time intervals, called the DBA cycle or maximum poll-
this study.
ing interval. Different DBA mechanisms use different
methods to assign the bandwidth for T-CONTs.
In the XG-PON mobile front-haul network (Fig. 2), the
TABLE I mobile uplink traffic that transmits via XG-PON front-haul
TRAFFIC SIMULATION PARAMETERS in the upstream direction is highly bursty, and accordingly
the bandwidth demand for ONUs/T-CONTs is very unpre-
Parameters Details dictable. Thus, evaluating the performance of XG-PON-
Number of bursts 5000 standard-compliant DBAs over such bursty traffic is highly
Mean burst time length 2 ms desirable. The following subsections present the perfor-
Poisson arrival rates [95, 110, 125–200 Mbps] mance evaluation of gGIANT and RR-DBA over simulated
Hurst parameter 0.8 bursty mobile front-haul traffic. Also, they introduce opti-
Pareto shape parameter 1.4 mized-RR DBA to accommodate mobile front-haul traffic
Packet size 1470 bytes
busrt-nesss.
Mikaeil et al. VOL. 9, NO. 11/NOVEMBER 2017/J. OPT. COMMUN. NETW. 987
heavily loaded T-CONTs at the last step of the algorithm standard deviation of the difference in end-to-end delay
P
using Exjt BW − K j
j0 W t ∕N, and finally formulates
between every successive packet’s reach at the OLT, ignor-
the bandwidth allocation map and broadcasts it back to ing lost packets; and (3) packet loss ratio.
the T-CONTs at the ONU side. (Note: After the first allo- Figure 5 shows the overall performance comparison of
cation cycle Exjt is no longer a zero, and also the excess the three algorithms when increasing the per-RRH/ONU
share calculated in this allocation cycle will be used to com-
pute the T-CONT updated limit of the upcoming allocation
cycle, i.e., W j2 W max Exj1 , W j3 W max Exj2 ; …; etc:)
In succeeding allocation cycles, similar to the first allo-
cation cycle, the DBA engine will assign grant allocation to
T-CONTs, then calculate the T-CONT updated limit for the
upcoming allocation cycle using the excess shares calcu-
lated from one cycle backward, and finally calculate the
excess shares for one cycle ahead.
traffic load as follows: [95, 110, 125 ∼ 200 Mbps]. As we can window size to accommodate more data and efficiently
see, optimized-RR outperforms both gGIANT and RR-DBA utilize the upstream frames in every allocation cycle (i.e.,
in all performance evaluation metrics (upstream delay, exploit the unallocated remainder of the upstream frames
jitter, and packet loss ratio). in every allocation cycle). This decreases the waiting time
From Fig. 5(a), we can see that optimized-RR performs for the heavily loaded T-CONTs and reduces the total num-
ber of frames that are needed to serve all T-CONTs, as well
better than both gGIANT and RR-DBA in terms of up-
as the total number of allocation cycles that each heavily
stream delay. Such a performance improvement is due to
loaded T-CONT needs to clear its buffer.
the efficiency of optimized-RR in utilizing XG-PON up-
stream bandwidth compared to the two other DBAs. This Also from Fig. 5(a), we can notice that both gGIANT
is clear from Fig. 6, which shows the amount of improve- and RR-DBA fail to satisfy the delay required for mobile
ment in XG-PON upstream utilization that can be added by front-haul for the 10 RRH front-haul case, as they both
optimized-RR DBA in comparison with gGIANT and RR- show upstream delay performance higher than 300 μs when
DBA—and also from Fig. 7, which shows the amount of per-ONU (RRH) load is in the range of 100 to 140 Mbps.
achieved average throughput by each DBA. In addition, Meanwhile, optimized-RR DBA attains an upstream delay
the improvement in the delay performance by optimized- performance in the range of 290–310 μs for the same range
RR DBA can be more easily explained as follows: because of per-ONU traffic loads. (Note: The 20 MHz/three sectors/
of the usage of the dynamic maximum allocation bytes limit four antennas LTE base station requires 123.2 Mbps front-
by optimized-RR DBA, T-CONTs (i.e., heavily loaded haul uplink throughput when the MAC-PYH split is consid-
T-CONTs) are allowed to increase their transmission ered as a split point between the BBU and RRH [27].)
Figure 5(b) shows the packet loss ratio performance of
the three algorithms. As we can see, optimized-RR DBA
achieves the lowest packet loss ratio compared to RR-
DBA and gGIANT. This is due to the reduction of network
congestion that can be achieved by optimized-RR DBA, by
allowing T-CONTs to use wider transmission windows to
transmit data and alleviate the packet drop rate.
Figure 5(c) shows the performance comparison of the
achieved jitter by each algorithm. From this figure we
can notice that the lowest packet jitter performance is also
achieved by optimized-RR DBA because packet jitter is
mainly impacted by the network congestion (i.e., the delay
between two successive received packets varies instead
of remaining constant when the network congestion in-
creases). As we can see, when the network load increases
and the congestion increases, packet jitter increases
accordingly.
Figure 6 shows the amount of utilization improvement
(as a percentage) that can be added by optimized-RR
Fig. 6. Amount of utilization improvement (as a percentage) that DBA at different per-ONU traffic loads when compared
can be added by optimized-RR DBA in comparison with RR-DBA with RR-DBA and gGIANT. From this figure we can notice
and with gGIANT, respectively. that the amount of utilization improvement (%) that can be
Fig. 7. Comparison of the achieved throughput by each algorithm at different per-ONU load: (a) 95 Mbps, (b) 125 Mbps, and
(c) 200 Mbps.
Mikaeil et al. VOL. 9, NO. 11/NOVEMBER 2017/J. OPT. COMMUN. NETW. 991
Fig. 8. Upstream delay versus the number of RRHs that can be Fig. 10. Distributed C-RAN architecture with multilevel
aggregated over XG-PON system by each algorithm. front-haul.
992 J. OPT. COMMUN. NETW./VOL. 9, NO. 11/NOVEMBER 2017 Mikaeil et al.
[32] Next Generation Fronthaul Interface 1914 Working Group [34] G. Fettweis and S. Alamouti, “5G: Personal mobile Internet
[Online]. Available: http://sites.ieee.org/sagroups‑1914/ beyond what cellular did to telephony,” IEEE Commun.
p1914‑1/. Mag., vol. 52, no. 2, pp. 140–145, 2014.
[33] “Common Public Radio Interface (CPRI); Interface [35] “Study on new radio access technology; Radio access architec-
Specification,” Oct. 2015 [Online]. Available: http://www. ture and interfaces (Release 14),” Tech. Rep. 3GPP TR 38.801-
cpri.info/downloads/CPRI_v_7_0_2015‑10‑09.pdf. 200, 2017.