Why Latency Matters to Mobile Backhaul - O3b Networks

Why Latency Matters to
Mobile Backhaul
by O3b Networks and Sofrecom
This paper is presented by O3b Networks to provide an
understanding of the impact of latency on mobile backhaul
services. Several common voice and data applications are
analyzed to assess the impact of latency on the quality of
the end user’s experience. The paper compares the Quality
of Experience (QoE) based on latency differences between
Geosynchronous satellites and Medium Earth Orbit satellites.
Contents
About O3b Networks Ltd. 2
About Sofrecom 2
Proprietary Statements 2
Mobile Industry Landscape 3
Latency for satellite backhaul and impact of Latency on
Quality of Experience 6
Conclusion 14
Annexes 15
www.o3bnetworks.com 1

Why Latency Matters to Mobile Backhaul
About O3b Networks Ltd.
O3b Networks is a global satellite service provider building a new fiberquality,
global Internet backbone for telecommunications operators,
internet service providers, enterprise and Government customers in
emerging markets. The O3b Networks system will combine the global reach
of satellite with the speed of a fiber-optic network providing billions of
consumers and businesses across 177 countries with low-cost, high-speed,
low latency internet and mobile connectivity. With investments and
operational support from SES, Google Inc. Liberty Global, Inc. HSBC Principal
Investments, Northbridge Venture Partners and Allen & Company, the O3b
system will provide telcos and ISPs with a low-cost, high-speed alternative
to connect their 3G, WiMAX and fixed-line networks to the rest of the
world. O3b Networks is headquartered in
St. John, Jersey, Channel Islands.
About Sofrecom
Sofrecom, a France Telecom – Orange Group Company, is a world leader in
telecommunications consulting and engineering.
With more than 45 years international experience, Sofrecom has acquired a
unique know-how of successfully managing change in the
telecommunications industry. Sofrecom has led major transformation
programs and has participated in the green field launch of multiple
operators, including all Orange mobile operators over the past 5 years.
This rich international experience provides Sofrecom with an unparalleled
expertise in all areas of telecommunications and makes it the natural
strategic partner for operators, governments, investors and international
financial organizations.
Proprietary Statements
O3b provides this information in good faith, and has taken all reasonable
steps to ensure its accuracy and completeness as of the date hereof.
However, O3b expressly disclaims any and all liability which may be based
on the use of such information, errors therein, changes and omissions
thereto. O3b shall not be liable for damages of any kind, including special,
incidental or consequential damages or damages for loss of goodwill or loss
of prospective profits, on account of omissions or errors contained herein.
O3b® and O3b Networks® are registered trademarks of O3b Networks
Limited. Other product names, company names, brand names, and trade
names mentioned within the corporate information center may be the
trademarks of their respective holders. All rights in the service marks,
company names, trade names, and logos used for the products or services
of O3b or of third parties belong exclusively to O3b or their respective
owners, and are protected from reproduction, imitation, dilution, or
confusing or misleading uses under national and international trademark
and copyright laws.
www.o3bnetworks.com 2

Why Latency Matters to Mobile Backhaul
Mobile Industry Landscape
The Mobile industry is changing rapidly, as it evolves from 2G to 3G and on
to 4G networks. This evolution is driven by the need to provide better
performance in three main areas:
− Provide subscribers with higher data rates
− Support a wider variety of end user applications
− Reduce the latency of the mobile network
Increasing sales of 3G smartphones, USB modems, tablets and PCs with
built in wireless radios is pushing data traffic on mobile networks to record
levels. While email, social networks and Internet browsing are very popular
among nomadic users, the deployment of mobile broadband services has
the biggest impact on network traffic. In rural areas, mobile networks are
often the only way to support applications historically delivered over
copper based networks. Video streaming is a major contributor to the boost
in traffic, with the success of Internet services such as YouTube and
DailyMotion. The latest Cisco Visual Networking Index forecasts
unprecedented global mobility demand.
Figure 1 – Mobile Data Growth
www.o3bnetworks.com 3

Why Latency Matters to Mobile Backhaul
Data services are projected to grow by a staggering 78% annually over just
the next four years. It is also clear that the mix of applications is changing,
particularly with the shift towards mobile video.
While the increasing data rates and the changing application mix are easily
understood, the impact of latency on performance requires a closer look.
How does latency affect application quality and the end user experience
We will explain the significant difference lower latency can make in the
delivery of voice, video and data services. We will strive to answer the
question: Why does latency matter to mobile backhaul
QoE is the new challenge
Fierce market competition is driving network operators to identify true
differentiators capable of separating them from the rest. The innovators, in
turn, are intensifying efforts to boost the Quality of Experience for their
subscribers. As a result, QoE is now one of the latest buzzwords in the
mobile industry.
The QoE experienced by a mobile subscriber for a voice call is influenced by
a number of factors, including mouth-to-ear delay, voice call set-up delay
and audio quality/noise. The QoE for data services is influenced by web
page response time and download speeds.
Network segment QoS indicators
For mobile network elements, such as components of the access, backhaul
and core networks, several indicators are used to define and monitor the
target performance.
− Latency - Latency is the duration of time for information to transit from
one point to another point of the network. Depending on the type of
service, it can be measured either in one direction (Ear-to-mouth or
One-way latency) or as a round trip time (RTT or Two-way latency).
− Jitter - Jitter is the variation of the measured latency. It is generated by
the variation of the load on the network due to usage levels.
− Packet loss - Packet loss is the ratio of packets being lost during transit
from one point to another point of the network. It is generally caused
by transmission errors, congestion of links, faults or the routing process.
In modern IP based networks jitter is easily compensated for by
implementing buffers in the endpoints. Likewise packet loss is very low on
today’s networks, which employ high performance error correction codes
such as LDPC which is used in DVB-S2. Latency, however, is more difficult to
avoid. The root cause is linked to long distances and the speed of light.
Once present in a system, latency cannot be removed. This is exactly why
the mobile industry has consistently engineered each new technology
generation with lower latency, as seen in Figure 2.
www.o3bnetworks.com 4

Why Latency Matters to Mobile Backhaul
Figure 2 – Latency in Mobile Technology Evolution
Every new mobile technology generation has reduced network latency. The
right side of the chart shows the latency of satellite backhaul for traditional
GEO satellite and O3b’s MEO constellation. It is easy to see that for later
generation mobile networks satellite backhaul is the dominant contributor
to end-to-end latency and the resulting subscriber QoE.
www.o3bnetworks.com 5

Why Latency Matters to Mobile Backhaul
Latency for satellite backhaul and impact of Latency on
Quality of Experience
End-to End latency Model
Each network component contributes to latency. The chart below, extracted
from a study by Nokia Siemens Networks (NSN), shows the latency
distribution for a mobile network (exclusive of backhaul or international
network latency).
Figure 3 – Latency Distribution for Mobile Network (NSN 1 )
(1) Extracted from “The impact of latency on application performance, Nokia Siemens Networks”
With each release there has been a tremendous engineering effort to
optimize each network element and terrestrial link to provide the lowest
possible latency. Individual components have been optimized, frame
lengths reduced, protocols streamlined, and in the case of LTE, an entire
network layer has been eliminated.
The satellite backhaul element is much more difficult to optimize.
Technologies, such as ComtechEFData’s VersaFEC, have improved the
modem processing delay, but the fundamental distance of the path length
to the satellite is what drives the delay.
www.o3bnetworks.com 6

Why Latency Matters to Mobile Backhaul
Mobile Service Classification
From a mobile network perspective, the ITU and the 3GPP have grouped
the end-user services into four classes based on their sensitivity to latency.
Figure 4 – User services classification
This paper focuses on the three major service classes for mobile operators
and the practical impact of latency on:
− Conversational Voice
− Interactive Applications, such as cloud-based applications, online
shopping, instant messenger and online search.
− Non real-time data, such as web browsing, email SMS/Text messaging,
file sharing, and video streaming.
Conversational Voice Services
Voice quality is measured in the ITU model by the R factor, which can also
be translated into a Mean Opinion Score (MOS) score, a scale widely used
to evaluate the quality of a voice call. The MOS score and voice call quality
degrades rapidly when the ear-to-mouth latency increases beyond 200ms.
Annex 1 provides a detailed explanation of the ITU model.
A latency comparison for a UMTS Release 99 voice call for two cases using
MEO and GEO satellites is shown below.
Figure 5 – End to end latency distribution for a voice call with O3b satellite backhaul
www.o3bnetworks.com 7

Why Latency Matters to Mobile Backhaul
Figure 6 – End to end latency distribution for a voice call with GEO satellite backhaul.
For O3b’s MEO constellation, the fiber return path adds 30ms of latency,
representing the case of more than 5000km of optical fiber.
The following diagram shows the resulting MOS scores for these two cases.
Figure 7 – Voice Quality / MOS score degradation as a function of latency
A side-by-side comparison of end-to-end latency values results in an MOS
score of 4.4 for O3b’s MEO constellation and a MOS score of 3.8 for
traditional GEO satellite.
www.o3bnetworks.com 8

Why Latency Matters to Mobile Backhaul
Interactive user services
The impact of latency on interactive services is more subjective. As the
delay increases, the applications continue to function. However, the sense
of immediacy and the usability degrades. At some point, the user becomes
less tolerant of waiting for the response, eventually giving up and moving
on to something more satisfying. A great deal of research has been done
by companies whose revenues are impacted when users abandon their
transactions.
The importance of latency on web commerce was first showcased by Zona
Research in 1999 with the popular “8-second rule”. Zona found that more
than $4 billion e-commerce sales were lost due to poor site performance,
which ultimately led to transaction abandonment. In a follow up study in
2006, Jupiter Research concluded that the threshold had been reduced to 4
seconds, as more users experience the performance of high-speed
broadband.
Google discovered that the number of web searches and ad revenue
declined as they increased the number of user search results from 10 to 30.
Further research found the result was due to the additional time required to
compute and present more results to the user. Increasing the results from
10 to 30 required an additional 500ms to compute and resulted in a 25%
drop in the number of searches done by users. Following this initial study,
Google performed additional tests focused exclusively on the latency. They
injected artificial delay into the display of the search results. An increase in
display time of 400ms actually decreased the number of searches per users
by 0.44% to 0.76%. While these drops in usage appear small, they had a
direct and meaningful impact on ad revenue.
Similarly with Amazon: “We tried delaying the page in increments of
100 milliseconds and found that even very small delays would result in
substantial and costly drops in revenue,” said researcher Greg Linden.
Linden provides a precise figure: -1% sales for 100ms or more latency.
(Greg Linden in “Make Data Useful”, Data Mining course Stanford in 2006).
Steve Souders from O’Reilly adds more examples in a blog posting
http://radar.oreilly.com/2009/07/velocity-making-your-site-fast.html.
In “Performance Related Changes and their User Impact,” Eric Schurman
(Bing) explains that Bing researchers have conducted tests adding a static
delay of 50 ms to 2 seconds to their servers. A negative impact on the Bing
performance indicators resulted as soon as the delay increased by more
than 50ms. The degradation noticed was linear.
www.o3bnetworks.com 9

Why Latency Matters to Mobile Backhaul
Distinct
Queries/User
Query
Refinement Revenue/User Any Clicks Satisfaction
Time to Click
(increase
in ms)
50ms – – – – – –
200ms – – – -0.3% -0.4% 500
500ms – -0.6% -1.2% -1.0% -0.9% 1200
1000ms -0.7% -0.9% -2.8% -1.9% -1.6% 1900
2000ms -1.8% -2.1% -4.3% -4.4% -3.8% 3100
Figure 8 – Bing table of latency impact for user search service
Phil Dixon (Shopzilla.com) presented results from the Shopzilla site
redesign at the Velocity 2009 Conference. He illustrated that a reduction in
page load time from 7 to 2 seconds resulted in a 25% increase in page
views and a 7-12% increase in revenue.
Below, a mobile subscriber is using a HSPA session over a satellite backhaul
to access Internet data services.
Figure 9 - End to end latency for data services using O3b backhaul
Figure 10 - End to end latency data services using GEO satellite backhaul
www.o3bnetworks.com 10

Why Latency Matters to Mobile Backhaul
The increase in latency of 420ms between O3b’s MEO solution and a GEO
satellite has a significant impact on interactive services.
As demonstrated, revenue loss is a significant consequence of the quality
degradation of a users’ experience caused by download delays. End users
perceive latency increases of a few hundred milliseconds in a negative
way, which translates into a loss of business for providers and advertisers.
Users are unaware whether the delay in response time is due to a slow
server or a satellite link. The end result in both cases is the same: a poor
user Quality of Experience.
TCP in a nutshell
The Internet uses the Transmission Control Protocol (TCP) to provide end-toend
services. In particular, TCP ensures the error free sequenced delivery of
data transmitted from a source to a destination.
Data is sent by the transmitter to the receiver in fragments called TCP
segments. A maximum size is defined for the segments. The receiver will
confirm the correct reception of sent TCP segments by acknowledging them
(sending the transmitter a response “ACK message” with a correct
sequence number). If an ACK message is not received or if it is not correct,
the transmitter will return the message ensuring a reliable transfer. At the
end of the session, it is correctly closed by exchange of FIN messages.
The key factor in this context is that TCP will only send a limited amount of
data before it needs an acknowledgement. The amount of data is governed
by the TCP buffer size, which in Windows 7 is 64 kbytes. If 64 kbytes of
data has been sent but the acknowledgements have not yet been received,
TCP will wait for an acknowledgement before it sends another packet. This
acknowledgement mechanism is what limits the transmission rate over
high latency links.
Non interactive data services
The major complexity of TCP comes from its flow and congestion control
mechanisms. The TCP flow control mechanism adjusts the TCP window size
(the number of segments that can be sent and not yet be acknowledged)
depending on the status of the end devices and congestion/latency on the
link during the transfer.
www.o3bnetworks.com 11

Why Latency Matters to Mobile Backhaul
Figure 11 – Messages exchanged for TCP connections.
Each message is subject to transmission delay. For complex transmissions,
such as web pages with multiple images and tables, the delay is
cumulative. If a page has ten elements, a 500ms increase in latency will
produce a 5 second increase in the page load time.
TCP throughput and transfer time
TCP throughput, and consequently file transfer duration, is a function of
end-to-end latency, packet loss rate and maximum segment size (more
details are available in Annex 2).
− Throughput is increased by a maximum segment size increase.
− Throughput is decreased by packet loss and/or latency increase.
Figure 12 – Maximum TCP throughput decrease due to latency
www.o3bnetworks.com 12

Why Latency Matters to Mobile Backhaul
Figure 12 shows that the maximum TCP throughput degrades rapidly as
latency increases.
The same illustration for an interactive service is considered: an HSPA
access over satellite to Internet service (c.f. Figure 9 and Figure 10 for detail
on each case).
The maximum throughput for TCP Reno will be 2.1 Mbps for O3b and
775 kbps for GEO satellite. This will be the maximum achievable rate for a
TCP connection, such as a single user communicating with a server. Table 5
estimates the transfer time for various media for both backhaul links for a
single TCP stream.
O3b
2 MB Song 10 seconds 28 seconds
1.5GB Movie 1.66 Hours 4.5 Hours
3.8GB HD Movie 4.25 Hours 11.4 Hours
Table 1: Estimated transfer time comparison between O3b and GEO backhaul
The subscriber QoE for these types of services is directly related to the
download time. The lower latency of the MEO case results in a huge
improvement in download times and subscriber QoE.
GEO
Non real time video streaming
Video streaming was traditionally transported over UDP, but major web
sites such as YouTube and DailyMotion now use either Adobe Flash Video or
HTML5:
− HTML5 is delivered using HTTP over TCP.
− Adobe Flash Video can be delivered using:
− FLW or SWF Files downloaded using a browser with HTTP over TCP
− Progressive download over HTTP over TCP
− Real Time Messaging Protocol over TCP directly or through HTTP tunnel
over TCP
− HTTP Live Streaming format over TCP.
With all the major providers of web video streaming now using TCP
transport for video streaming, the latency impact on the degraded user
experience will be exactly the same for Internet video streaming as it is for
data file transfer.
www.o3bnetworks.com 13

Why Latency Matters to Mobile Backhaul
Conclusion
In the past, a boost in the subscriber data rate was all that was needed,
and operators could simply focus on increasing voice minutes and data
volume. But as consumers become more sophisticated and networks
become more complex with the introduction of mobile broadband, Quality
of Experience has become an important indicator of network performance.
Latency is the critical factor in improving QoE across all services, including
traditional voice services and the latest data services (i.e. interactive cloudbased
applications and movie downloads).
− Voice quality is measurably improved using the ITU model
− The response of interactive applications is dramatically improved
− File download times are reduced by over 60%
The latency of O3b’s MEO constellation is substantially better than
traditional GEO satellites. This has a direct improvement on QoE and
provides a competitive advantage for operators who are deploying modern
mobile networks in rural areas.
www.o3bnetworks.com 14

Why Latency Matters to Mobile Backhaul
Annexes
Annex 1 Voice Codec, E-Model and MOS score
The quality of voice services has been defined in several ITU standards,
namely G.107, G.108 and G.109. Quality of the voice services is evaluated
using a quality factor “R” ranging from 0 to 100.
The R factor can be directly translated to a MOS score, as defined in the ITU
G.107 recommendation. The mapping formula can be represented by the
following curve.
Figure 13: MOS Score and R Rating mapping (ITU G. 107 / P. 800)
www.o3bnetworks.com 15

Why Latency Matters to Mobile Backhaul
End to end latency, or so called “mouth to ear” has a strong impact on the
voice quality, as illustrated by the following figure extracted from the G.114
recommendation.
Figure 14: End-to-end latency effect on R factor (ITU G.114)
The higher the latency, the lower the voice quality. And as shown in the
right side of the graph, voice quality is directly related to user’s QoE and
overall satisfaction.
www.o3bnetworks.com 16

Why Latency Matters to Mobile Backhaul
Finally, the absolute quality of the voice and latency impact is related to
the codec selected. Although in any case, higher latency means a decrease
in quality. For Mobile backhaul, AMR-12.2 is widely used and corresponds
to GSM-EFR codec.
Ie= 0 5 7 10 15 19 19 20 26
ms
G.711 GSM-EFR
G.726
@32 G.729
G.728
@16
~0 94 87
G.723.1
@6.3
G.729A
+VAD
w/2% loss
G.723.1
@5.3
G.723.1
@
6.3+VAD
w/1% loss
GSM-FR
IS-54
G.729A
+VAD
w/4% loss
50 93 86 83 74 67
100 92 87 85 82 77 73 73 72 66
150 90 85 83 80 75 71 71 70 64
200 87 82 80 77 72 68 68 67 61
250 80 75 73 70 65 61 61 60 54
300 74 69 67 64 59 55 55 54 48
350 68 63 61 58 53 49 49 48 42
400 63 58 56 53 48 44 44 43 37
450 59 54 52 49 44 40 40 39 33
Note 1 – R-values in this table have been calculated using the indicated values for Ie and T (T=Ta=Tr/2) along
with the default values from Table 6 for all other parameters.
Note 2 – Unless indicated otherwise, examples do not include packet loss or Voice Activity Detection (VAD).
Note 3 – Blackened cells indicate combinations of delay and codec that are impossible to realize.
Figure 15: R rating values for various voice codecs depending on End to end delay (ITU G. 108
Annex B)
Annex 2 TCP details and throughput calculation
This annex gives more details on TCP Flow Control and TCP Congestion and
Maximum TCP throughput and TCP transfer duration.
Flow control
Flow control is achieved using a feature called “sliding windows”.
The receiver advertises the amount of data it can accept through time
(generally related to buffering capacity). The transmitter is allowed
to transmit data up to this advertised value until it receives an ACK
message.
www.o3bnetworks.com 17

Why Latency Matters to Mobile Backhaul
For instance, the receiver advertised a maximum window size of 10
segments. The transmitter will send the 10 segments (at a variable rate
described in the congestion avoidance) and will then stop until ACK is
received. For each ACK (if window size stays constant) received, the
transmitter may send another segment. The algorithm may advertise
a variable window size depending on the actual state of the receiver
(i.e. buffers filling or emptying), allowing it to transmit more or fewer
segments and vary the actual throughput.
Congestion Control
Flow control enables the two end devices to regulate the transmission
throughput, but it does not take into account the status of the network links
between them. Congestion control mechanisms are required to regulate
the speed of the data transmission and the quality of the network link.
Using a slow start mechanism, the sender begins to transmit TCP segments
at a low rate, gradually increasing the number of segments simultaneously
transmitted until the full window size is reached or congestion is detected.
In case of congestion, an avoidance mechanism will drastically decrease
the rate and initiate the step increase again, creating a saw-tooth effect.
Various TCP flavors (Reno, Vegas, Cubic, Compound) have variations in the
algorithm used to define the increasing and decreasing rate, based on a
maximum segment size and/or queuing delay.
Other optimizations include retransmission of selected segments and fast
recovery to increase the transmission rate more quickly.
TCP maximum throughput estimation
TCP maximum throughput can be estimated using a formula based on TCP
RENO (RFC2001) as described by Padhye et. al. in "Modeling TCP
Throughput: a Simple Model and its Empirical Validation"
The model used is the following one:
⎛
⎜Wmax
B(
p)
≈ min ,
⎜ RTT RTT
⎝
2bp
3
+ T min 1,
0
1
( ) ( ) 3 bp
2
3 p 1+
32 p ⎟⎟ 8
⎠
⎞
www.o3bnetworks.com 18

Why Latency Matters to Mobile Backhaul
Where
B(p): approximate model of TCP throughput [packet/s]
Wmax: maximum window buffer size of receiver [packets]
RTT: Round Trip Time [sec]
b: number of packets that are acknowledged by a received ACK
p: probability that a packet is lost
T0: time-out for re-transmitting an unacknowledged (lost) packet [sec]
RTT should take into account the additional delay due to serial transmission
of the packet (especially for low rate interfaces) and various TCP receive
window sizes depending on the Operating System of the end-device.
TCP transfer duration estimation
Overall duration of file transfer can be approximated using the formula
given by N. Dukkipati et al in “An argument for Increasing TCP’s Initial
Congestion Window”
S being the size of the file
C being the bottleneck link-rate
RTT being the two-way end-to-end latency
being 1.5 if ACK are delayed of 2 else.
Init_cwnd is the initial congestion windows, generally no more than three segments or about 4kB.
The transfer time calculated does not take into account packet loss
(no packet loss) for simplicity. Packet loss further increases the duration of
transfer due to retransmission of error packets and congestion control
mechanisms.
www.o3bnetworks.com 19