Content uploaded by Gabriel Otero
Author content
All content in this area was uploaded by Gabriel Otero on Feb 04, 2021
Content may be subject to copyright.
1
Delay Analysis of Fronthaul Traffic in 5G Transport
Networks
Gabriel Otero P´
erez, Jos´
e Alberto Hern´
andez, and David Larrabeiti L´
opez
Department of Telematics Engineering, Universidad Carlos III de Madrid, Spain
Email: {gaoterop, jahgutie, dlarra}@it.uc3m.es
Abstract—Cloud Radio Access Network (C-RAN) architecture
claims to reduce capital costs and facilitate the implementation
of multi-site coordination mechanisms. This paper studies the
delay constraints imposed by the Common Public Radio Interface
(CPRI) protocol in ring-star topologies used by mobile operators.
Simulations demonstrate that centralised implementations are
feasible via functional split in the baseband processing chain.
We derive theoretical expressions for propagation and queueing
delay, assuming a G/G/1 queueing model. Then, we examine
the properties of the fronthaul traffic flows and their behaviour
when they are mixed. We show that the theoretical queueing
delay estimations are an upper bound on the simulation output
and accurate under certain conditions. Based on our results, we
further propose a packetisation strategy of the fronthaul traffic
which helps reduce the worst case aggregated queueing delay
by 30%. Also, the benefits of a bidirectional ring topology are
shown, achieving a worst average queueing delay 10 times lower
than that of unidirectional topologies.
Index Terms—5G, CPRI, C-RAN, Fronthaul, Delay Analysis
I. INTRODUCTION
A. Motivation
ACCORDING to the Visual Networking Index (VNI)
Global Mobile Data Forecast [1] released by Cisco
in February 2016, there will be 5.5 billion global mobile
users by 2020. Nielsen’s law [2] advocates that the required
bandwidth in networks is steadily increasing, approximately,
50%every year. In the near future, a high-bandwidth, low
latency interconnection network will be mandatory. In order
to cope with the ever increasing traffic load that the networks
will need to support, a new approach for planning cellular
deployments should be followed.
Cloud Radio Access Network (C-RAN) architecture pre-
sented by China Mobile [3], introduces the idea of a cloud
computing-based processing of baseband signals on cellular
networks. Experiments testing the C-RAN framework reveal
that significant savings in both operational expenditure (OPEX)
and capital expenditure (CAPEX) can be achieved. This concept
represents large-scale centralised base station deployments,
achieving significant cost reductions by separating the radio
equipment of each base station from the elements that process
the signals, which now are centralised and possibly virtualised.
In this scenario, the Common Public Radio Interface (CPRI) [4]
protocol provides an interface between the radio transceivers,
Remote Radio Heads (RRHs), and the processing units, i.e.,
Baseband Units (BBUs) to transport the so-called fronthaul
traffic generated at the RRH through the backhaul network.
CPRI is an industry standard that can be used to implement
the Digitized radio-over-fiber (DROF) concept proposed in,
e.g., [5], [6]. Additionally, there exist several packetisation
projects which aim at developing protocols for the transport
of radio samples over packet-switching networks. It is worth
highlighting the work being carried out in the Time-Sensitive
Networking for Fronthaul IEEE 802.1CM standard [7], which
pursues to enable the transport of time-sensitive fronthaul
streams over Ethernet bridged networks. However, to the date,
no characterisation of fronthaul traffic and the aggregation of
fronthaul flows has been performed, which is a preliminary
step to fulfill the stringent delay and jitter requirements of
this type of traffic.
We address these questions as follows: Section II presents
the problem scenario, assumptions and analyses the propa-
gation delay. In Section III, we analyse the behaviour of the
fronthaul traffic in a simulated ring-star topology and compute
the end-to-end queueing delay. We conclude our paper in
Section IV.
II. PROBLEM STATEMENT AND ASSUMPTIONS
A. Reference scenario
Figure 1 illustrates the topology used to measure and evalu-
ate the performance of the different approaches. Since network
coverage and spatial reuse are fundamental issues of wireless
networks, operators traditionally followed a hexagonal grid
deployment architecture, in which a target field is partitioned
into hexagonal grids. Assuming a hexagonal cell coverage
area, we have groups of 7cells that, from now on, we will
refer to as hives. Every cell comprises three sectors, each
served by a 120◦sector antenna. Regarding the uplink (the
one that imposes the most stringent delay requirements [12]),
the traffic originating from the three sectors is aggregated at
the centre of the cell in a tree topology, using direct optical
fibre links. Later, the traffic of each cell in a hive is mixed
together at its central point (see numbered circles). The green
star at the center of Fig. 1 represents the point where the
flows coming from all topolgy sectors meet. It is also part
of a higher level (ring) network, connecting another groups of
hives or clusters. In addition, it is a great candidate for hosting
the BBU processing unit for the entire cluster. We number, in
clockwise descending order each hive starting in the yellow
shouth westernmost one. We will refer to this hive as hive #7.
Then, the last hive in the ring is hive #1.
One of the end-to-end delay components is the propagation
delay of the packets. Assuming the physical properties of the
ª*&&&
Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on July 22,2020 at 19:32:16 UTC from IEEE Xplore. Restrictions apply.
2
Fig. 1. Regular hexagonal lattice deployment.
TABLE I
PROPAGATION DELAY
Propagation delay - Worst case - Ring-Star Topology
Unidirectional Bidirectional
Dense Urban (R = 300 m) 43.84 μs 22.22 μs
Urban (R = 500 m) 73.069 μs 38.70 μs
Rural (R = 1500 m) 219.21 μs 116.10 μs
optical links to be fixed, the propagation delay depends only
on the total physical distance between the radio equipment
and the processing units. Regarding the cell coverage radius,
common values used in the related literature [8] are 300 m, 500
m and 1500 m for Dense-urban,Urban and Rural scenarios,
respectively. On the other hand, we must take into account
the queueing delay, which appears whenever we packetise the
user’s uplink radio signals. Since the communication process
occurs in a ring we should have in mind that it might be
unidirectional or bidirectional. In the first case, we will have
a worst case scenario where the packets originating in hive
#7are suffering the highest propagation and queueing delays
since they have to travel the longest distance to reach the
final aggregation point in hive #1. In the bidirectional ring
case, the propagation delay is expected to be half of that
in the unidirectional case and the queueing delay would be
affected by half the number of aggregation points. Trivial
geometry calculations [9] lead to a closed form expression
for the propagation delay, which is defined as follows
dprop =R·cos(30◦)·(N√28 + 2)
vprop
(1)
where Rdenotes the radius of a hexagonal cell, Nrepresents
the number of hives a given packet needs to cross until it
reaches its final destination and vprop is the speed of light in op-
tical fiber. Since the refractive index of glass is approximately
1.5, we assume vprop 2·108m/s, which means a 5μs/km
delay. Table I shows the results of applying the derived analytic
expression (see eq. (1)) for the propagation delay in the
proposed topology. Worst-case scenarios have been taken into
Fig. 2. Uplink LTE processing chain.
account to compute the final delay values for each case, i.e.,
the propagation delay values in the table are those experienced
by the packets originating in the furthest cells from the central
BBU in the ring topology. 100 μs is the target maximum delay
budget determined by the IEEE 802.1CM working group. Half
of it can be spent in the propagation delay, which allows a
target RRH-BBU distance up to 50 μs
5μs/km =10km. As for the
other half, it should be enough to allow a reasonable amount of
hops, each one of them adding both processing and queueing
delays.
B. Functional split B
The first option to connect a RRH and the BBU consists
of transmitting the pure sampled radio signal [10]. Since no
further processing is done at the remote radio equipment,
overhead information , such as the cyclic prefix (CP), is being
sent over the link to the BBU. Complex processing devices
are no longer needed at the RRH because all the functions
required to decode the signal are located at the BBU. The data
rate assuming a single carrier implementation is:
RSplit A =2·fs·Nov ·Nbits ·Nant (2)
where Nov is the oversampling factor. Nbits and Nant repre-
sent the bit resolution we use to quantise the signal samples
and the number of receiving antennas, respectively. A factor
of 2 accounts for the complex nature of each I/Qsample.
For instance, a 2-branch antenna system with 10 bits per
sample, assuming an oversampling factor of 2 and a sampling
frequency of 30.72 MHz requires 2.46 Gbit/s per sector. In
order to relax the bandwidth burden we may remove the cyclic
prefix from the quantised signals and perform a Fast Fourier
Transform (FFT) to decode the OFDM subcarriers. Subcarriers
used as a guard band, typically around 40%, are no longer
necessary. Assuming a resolution of 10 bits per sample, a
subcarrier spacing of 15 KHz and a symbol rate of 66.7μs
to maintain orthogonality, the new rate is
RSplit B =2·fs·NSub ·Nov ·Nbits ·Nant 720 Mb/s (3)
where 1/fs=66.7μs and 1200 active subcarriers are con-
sidered. In Split B, the bandwidth requirement is clearly
Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on July 22,2020 at 19:32:16 UTC from IEEE Xplore. Restrictions apply.
3
reduced. Further processing of the signals at the base station
may leverage the fact that only part of the resource blocks
of the base station are being assigned to users. This is the
case of functional splits C, D and E. Bandwidth requirements
obviously decrease as we leave more processing equipment at
the RRHs but, on the other hand, deployment and maintenance
costs increase and we progressively lose the benefits of cen-
tralisation [11]. Detailed investigations of the above-mentioned
splits have been conducted in [12].
C. Queueing theory review
M/M/1 and M/G/1 models are attractive because closed
expressions can be obtained for the main metrics of interest,
such as average waiting time in queue, mean number of users
in the queue, the server load, etc. However, exponentially
distributed interarrival times assumptions are made. Time be-
tween arrivals is not exponentially distributed (until we merge
enough number of constant bit rate flows, see Section III-A)
and our service time is constant and does not follow an
exponential distribution. This is the reason why the G/G/1
model, generalising both interarrival and service times, is
preferred to obtain the estimations of the mean wait time
in queue in our multihop tolopogy. Unfortunately, no closed
expressions exist for the mean waiting time in queue under
these assumptions. Let Tbe the random variable modeling
the interarrival times of packets at the queue and S, the
service time random variable. Also, we write ρfor the system
utilisation. Defining the squared coefficient of variation of a
random variable Xas C2[X]=Var [ X]
E[X]2, an upper bound on this
parameter is [13]
Wq≤E[S]·ρ
1−ρ·C2[T]+C2[S]
2(4)
which claims to be a good approximation to the mean queue
waiting time when ρ→1. There also exists a lower bound
which is not very useful since it often gives negative results,
which are trivial outcomes.
III. EXPERIMENTS
After analysing the bandwidth requirements in Section
II-B, we focus our study on Split B. Recall that, in Split
B, traffic follows a Constant bitrate (CBR) pattern with a
rate of 720 Mb/s (90 MB/s) and an OFDM symbol is sent
every 66.7μs. Accordingly, we can compute the burst size
as 90 MB/s ·(15KHz)−1= 6000 bytes. Figure 3 shows
the initial packetisation scheme chosen to conduct our study,
which consists of four back-to-back packets. The efficiency
of each packet is given by
η=Packet payload
Packet Header +Packet Payload (5)
Consequently, packets with a 46 bytes (RoEover MAC-in-
MAC) header and 1500 bytes of payload, lead to an efficiency
η0.97. Table II shows the efficiency values for different
payload sizes. Additionally, we implemented a custom discrete
event simulator so as to assess the validity of the theoretical ap-
proximations as well as to unveil the behaviour and properties
Fig. 3. Split B burst.
TABLE II
Payload Size Efficiency (η)
500 bytes 0.916
1000 bytes 0.956
1500 bytes 0.970
2000 bytes 0.978
0 100 200 300 400 500 600 700 800
Number of merged flows
1
2
4
6
8
10
12
14
16
18
20
12 Packets per Burst
6 Packets per Burst
4 Packets per Burst
3 Packets per Burst
Poisson Arrival Process
Fig. 4. Arrivals squared coefficient of variation.
of the traffic under different conditions. We make use of this
simulator in the following experiments, considering 100 Gb/s
links for the ring-star topology explained in Section II-A.
A. Aggregation of multiple fronthault flows
The aim of this experiment is to check the steady state
convergence of the arrivals squared coefficient of variation, if
any, when we aggregate more and more flows in a given hive’s
hub. Flows are merged applying an offset to each deterministic
burst flow, uniformly distributed between 0and the burst
period, U(0,T), where T=66.7μs. In the worst case, two
flows are completely aligned, that is, their bursts arrive at the
same time to the aggregation point.
As shown in Figure 4, the squared coefficient of variation
of the packet arrivals, converges to unity as we increase
the number of mixed flows. This behaviour is explained by
the Palm-Khintchine theorem [14], which claims that if we
combine a large enough number of independent and not
necessarily Poissonian renewal processes, each with small
intensity, we encounter Poissonian properties1. It is rather
1Squared Coefficient of Variation: C21
Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on July 22,2020 at 19:32:16 UTC from IEEE Xplore. Restrictions apply.
4
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
System load [ ]
0.5
1
1.5
2
2.5
3
3.5
4
4.5
Estimation Ratio
3 Flows
10 Flows
50 Flows
100 Flows
150 Flows
Fig. 5. Queueing delay vs System Load. 4 packets per burst.
important to stress that the rate of convergence to the steady
state is different depending on payload size of the packets
we use to transport the burst. Note that the distribution of the
interarrival times is not Poissonian, (i.e, it does not converge to
one) until we merge, approximately, more than 600 flows with
the aforementioned bursty structure. Hence, we cannot assume
the M/M/1 nor the M/G/1 model as good approximations since
only (3 sectors ·7cells ·7hives = 147) fronthaul flows are
merged at the last hop of the topology, in the worst case. In
addition, for 150 flows, the squared arrival coefficient of
variation reduces from 8 while using 12 packets per burst
to roughly 2, when we use 3. This factor strongly affects
the queueing delay in view of equation (4). Also, Figure 3
shows that the service time of the packets is not exponentially
distributed, since they are all of the same size and, thus, it
follows a deterministic distribution. In order to take all these
aspects into account, assume both distributions as generalised
distributions characterised by the appropriate coefficients of
variation. This fact supports the decision of using the G/G/1
queueing model so as to estimate the theoretical queueing
delay.
B. Theoretical estimations vs Simulation
In this section, we assess the validity of the analytic estima-
tions and how close they are to the simulation outputs. Figure
5 shows the evolution of the ratio between the theoretical
G/G/1 estimation and the simulation results for the queue
waiting time as we increase the load of a given aggregation
point, for different number of aggregated flows. It is worth
noting that the more flows we merge, the more similar the
analytic estimations and the simulation outputs become until
we reach heavy load states. The ratio approaches to unity
for system loads ρ≥0.4while merging 50 or more flows.
Additionally, we observe that the theoretical G/G/1 queueing
model is, indeed, an upper bound on the simulation outputs.
Since theoretical values are only approximations and consid-
ering the gap between the estimated and the simulated values
for some load conditions, we present only the simulation
outputs from now on.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Simulated Mean Aggregated Delay
Hive#1 Hive#2 Hive#3 Hive#4 Hive#5 Hive#6 Hive#7
0
2
4
6
8
10
12 packets per burst
6 packets per burst
4 packets per burst
3 packets per burst
(a) Unidirectional ring
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Simulated Mean Aggregated Delay
Cell Hive#4 Hive#5 Hive#6 Hive#7
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
12 packets per burst
6 packets per burst
4 packets per burst
3 packets per burst
(b) Bidirectional ring
Fig. 6. End-to-end queueing delay.
C. End-to-end delay: deciding on the packet size
We now focus on measuring the average and worst case end-
to-end delay of a packet, that is, in the unidirectional case, we
measure the queueing delay a given packet experiences on its
path originating in a RRH of hive #7all the way throught to
the BBU facilities, located at hive #1. Note that the fiber link
between hive #1and the BBU is the bottleneck here due to the
fact that it has to deal with the largest amount of traffic of all.
In the unidirectional case it serves the traffic resulting from
the aggregation of every single hive’s flow. In the bidirectional
case, we may consider each half of the ring separately. One
side merges the traffic flows coming from hives #5,#6and
#7. The other one, has to aggregate the traffic from hives #4,
#3,#2and #1, which is the worst case scenario. Figures 6(a)
and 6(b) show, respectively, the mean aggregated queueing
delay of a packet for the unidirectional and bidirectional worst
cases, as it traverses the ring topology. Note that the aggregated
queueing delay does not grow linearly as we approach the final
destination, the BBU located at the center of hive #1. Also, the
Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on July 22,2020 at 19:32:16 UTC from IEEE Xplore. Restrictions apply.
5
Cell Hive#7 Hive#6 Hive#5 Hive#4 Hive#3 Hive#2 Hive#1
0
5
10
15
20
25
Fig. 7. Queueing delay percentiles. Unidirectional Ring & 12 packets per
burst.
aggregated queuing delay is below 2μs until the 5th hop in
both cases. With the aim of assessing the effects of different
packetisation policies on the queueing delay, we illustrate the
behaviour of the system for different packet payload sizes.
Closer inspection of the plots, reveals that the more packets we
use to transport a given burst, the higher the average queueing
delay. In addition, we find that at the first hops of the topology,
the packetisation strategy does not really affect the aggregated
queueing delay. Conversely, when many flows are mixed in the
last hops at the very end of the ring, choosing the right number
of packets per burst can make the difference. In this scenario,
choosing 12 packets to encapsulate the OFDM symbol leads to
an average queue waiting time of 10.48 μs (see Figure 6a). On
the other hand, employing 3packets to encapsulate the burst
means, on average, 7.22 μs which represents an approximate
31% saving in terms of waiting time at the concentrator’s
packet queue. It is important to have in mind that, in the
unidirectional ring case, the system load at the last hop (hive
#1) is close to unity. Additionally, note that, by considering
a bidirectional ring, the average worst case queueing delay
reaches approximately 1μs at the last hop, which is 10 times
lower than that of the unidirectional case. Figure 7 shows
the aggregated queueing delay statistics obtained from the
simulator in the unidirectional worst case for the 12 packets
per burst case. Notice that interquartile range increases as we
traverse the ring towards the BBU. Furthermore, regarding the
last hop (hive #1), 5% of the packets are likely to suffer an
aggregated queueing delay of more than 23 μs. As a rule
of thumb, the experiments show that we can obtain a better
performance by using packets with a greater payload and, thus,
decreasing the overhead and increasing efficiency.
IV. SUMMARY AND CONCLUSION
The benefits of a centralised processing architecture are
clear from the mobile operator perspective. Users may also
benefit from it by using cloud’s powerful and specialised
hardware. Enhancements may include robust and complex
forward error correction algorithms, parallel computing, more
sophisticated coordination multi-point algorithms, etc. We con-
clude that a real world deployment is achievable using Split B.
We found that the rate of convergence of the arrivals squared
coefficient of variation is different depending on the packet
payload size. Furthermore, when aggregating 150 flows, it can
be reduced by a factor of 4 by using 3-packets bursts, instead
of 12. Regarding the tightness of the theoretical estimations,
we show that they are close to the simulation outputs for
system loads ρ≥0.4while combining more than 50 flows.
Aggregated queueing delay is, on average, far from exceeding
the 50 μs budget envisioned in Section II-A. Also, the worst
average queuing delay is 10 times smaller in the bidirectional
case, compared to the unidirectional ring topology. Neverthe-
less, two aspects must be analysed carefully. Regarding the
statistical properties of the queueing delays, some packets may
suffer from a queueing delay which is much higher than the
average. Secondly, more sources of delay should be added to
the propagation and queueing delays, such as switching delay,
packet processing delays, etc. Dynamic optimal allocation of
flows into different dedicated optical circuits and studies about
the optimal location of the BBU are natural extensions to this
work.
ACKNOWLEDGEMENTS
The authors would like to acknowledge the support of the
Spanish project TEXEO (grant no. TEC2016-80339-R) and the
EU-funded 5G-Crosshaul project (grant no. H2020-671598).
REFERENCES
[1] Cisco, Visual Networking Index, February 2016, White paper at cisco.com
[2] J. Nielsen, Nielsen’s Law of Internet Bandwidth, April 5, 1998
https://www.nngroup.com/articles/law-of-bandwidth
[3] China Mobile, The Road Towards Green RAN (White Paper),
http://labs.chinamobile.com/cran/wp-content/uploads/2014/06/
20140613-C- RAN-WP- 3.0.pdf
[4] Common Public Radio Interface (CPRI), Interface Specification v5.0,
http://www.cpri.info/spec.html.
[5] B. R. Ballal and D. Nema, ”Performance Comparison of Analog and
Digital Radio Over Fiber Link,” International Journal of Computer
Science & Engineering Technology (IJCSET), 2012.
[6] S. Kuwano and Y. Suzuki, ”Digitized Radio-over-Fiber (DROF) System
for Wide-Area Ubiquitous Wireless Network,” IEEE Xplore, April 2007
[7] IEEE Time-Sensitive Networking for Fronthaul 802.1CM: http://www.
ieee802.org/1/pages/802.1cm.html
[8] S. Purge, D. Breuer, ”Convergence of fixed and Mobile Broadband
access/aggregation networks,” Technical Report, October 2016
[9] P. Suksompong, ”The Cellular Concept,” Technical Report, November
2009, http://www2.siit.tu.ac.th/prapun/tcs455/Cellular455.pdf
[10] A. De la Oliva, J. A. Hern´
andez, D. Larrabeiti, et al, ”An overview of the
CPRI specification and its application to C-RAN-based LTE scenarios,”
IEEE Communications Magazine, 2016
[11] D. Wubben et al., ”Benefits and impact of cloud computing on 5g signal
processing: Flexible centralization through cloud-ran,” Signal Processing
Magazine, IEEE, 2014.
[12] U. D¨
otsch, M. Doll, H. P. Mayer, F. Schaich, J. Segel, and P. Sehier,
Quantitative analysis of split base station processing and determination
of advantageous architectures for LTE, Bell Labs Tech. J., May 2013.
[13] J. Kingman, ”The single server queue in heavy traffic”. Mathematical
Proceedings of the Cambridge Philosophical Society, 1961
[14] J. A. Hern´
andez, P. Serrano: Probabilistic models for computer networks:
tools and solved problems, 2013 (ISBN: 978-1291546873)
[15] F. Cavaliere, P. Iovanna, et al , ”Towards a Unified Fronthaul-backhaul
Data Plane for 5G The 5G-Crosshaul Project Approach,” Comput. Stand.
Interfaces, Elsevier Science Publishers B. V., March 2017
[16] C. Divya, K. Koteswararao Kondepu, ”5G Fronthaul Latency and Jitter
Studies of CPRI Over Ethernet,” Opt. Commun. Netw., February 2017
Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on July 22,2020 at 19:32:16 UTC from IEEE Xplore. Restrictions apply.