Professional Documents
Culture Documents
ECS455 - Premidterm - U2
ECS455 - Premidterm - U2
Introduction
Asst. Prof. Dr. Prapun Suksompong
(«.¦.¦³¡´r » ¤°)
prapun@siit.tu.ac.th
Office Hours:
BKD, 6th floor of Sirindhralai building
Tuesday 14:20-15:20
Wednesday 14:20-15:20
1 Friday 9:15-10:15
2
Wired Communication
y Cup-and-string communication
Wireless communication
4
Wireless communication
y Duncan Wilson's Cup Communicator
Mobile?
y The term “mobile” has historically been used to classify all
radio terminal that could be moved during operation.
y More recently,
y use “mobile” to describe a radio terminal that is attached to a
high speed mobile platform
y e.g., a cellular telephone in a fast moving vehicle
y use “portable” to describes a radio terminal that can be hand-
held and used by someone at walking speed
y e.g., a walkie-talkie or cordless telephone inside a home
6
[Rappaport, 2002, p 9]
[Goldsmith, 2005, Section 1.1]
History (1)
y The first wireless networks
were developed in the pre-
industrial age.
y These systems transmitted
information over line-of-sight
distances (later extended by
telescopes) using smoke
signals, torch signaling, flashing
mirrors, signal flares, or
semaphore flags.
Semaphore
8
History: Radio
y Early communication networks were replaced first by the
telegraph network (invented by Samuel Morse in 1838) and
later by the telephone.
y In 1895, Marconi demonstrated the first radio transmission.
y Very high transmission power (> 200 kW)
y Early radio systems transmitted analog signals.
y Today most radio systems transmit digital signals
composed of binary bits.
y A digital radio can transmit a continuous bit
stream or it can group the bits into packets.
y The latter type of radio is called a packet radio and is
characterized by bursty transmissions
History: ALOHANET
y The first network based on packet radio, ALOHANET, was
developed at the University of Hawaii in 1971.
y ALOHANET incorporated the first set of protocols for
channel access and routing in packet radio systems, and many
of the underlying principles in these protocols are still in use
today.
y Lead to Ethernet and eventually wireless local area
networks (WLAN).
10
History: Pre-Cellular (1)
y The most successful application of wireless networking has
been the cellular telephone system.
y 1915: Wireless voice transmission (wireless telephony)
between New York and San Francisco was first established by
AT&T.
y 1946: First public mobile telephone service
was introduced in 25 cities across the US.
The equipment was expensive at $2,000.
y More than the price of a typical new car
(at that time).
[Goldsmith, 2005, p 3]
11 [http://earlyradiohistory.us/1915ATT1.htm]
12
History: Pre-Cellular (3)
y These initial systems used a single central transmitter to cover
an entire metropolitan area.
y High-powered transmitter & Large tower
y Inefficient!
y FM push-to-talk
y 1976: (30 yrs after the introduction of the service in 1946),
y the New York system (10M people) could only support 543 paying
customers.
y 3,700 on the waiting list
y The mobile units weighed about 10 kilograms and put out a steady
20-25 watts.
y The central transmitters that communicate with the mobile units
broadcast 200 to 250 watts.
13 [Klemens, 2010, Chapter 3]
Region 2
Region 1
Region 2
Region 4
Uplink signals from user of a Region 3
cell can reach the BS of a
different region.
16 [Klemens, 2010, p 46]
History: Pre-Cellular System (3.4)
y Regions need to be well-separated!
y As a result, the central transmitters had to be at least 100
miles apart, leaving a 50 mile blank space between them.
y So a customer could use the sporadic and unreliable service
only within the confines of one area.
Region 1 Region 2
Region 4
Region 3
17
[http://communities-dominate.blogs.com/brands/2009/11/celebrating-30-years-of-mobile-phones-thank-you-ntt-of-japan.html]
y Japan had the world’s first commercially available cellular
phone system.(Japan's national monopoly telecoms operator/carrier)
y Nippon Telegraph and Telephone (NTT)
created a cellular test system for Tokyo in 1975,
with the result coming to market in 1979.
y The first trial in America of a complete, working cellular system
was held in Chicago in the late 1970’s.
y 1983: Advanced Mobile Phone System (AMPS)
y First US cellular telephone system
y Deployed in 1983 by Ameritech in Chicago, IL.
y Worked well. (FM, FDMA)
y May even have worked too well.
y Its satisfactory performance lowered the demand for a better system,
allowing Europe to take the lead by creating a digital cellular system first.
19
20
[http://www.google.com/patents?id=nO8tA
AAAEBAJ&printsec=abstract&zoom=4&hl=th
21 #v=onepage&q&f=false]
22
First Cell Phone Call in 1973
https://www.youtube.com/watch?v=7A6RqCYuKac
23
Motorola’s DynaTAC
First commercially
available cell phone in
1983
25
https://www.youtube.com/watch?v=0WUF3yjgGf4
26
Motorola DynaTac 8000X
y The DynaTAC phone can be seen in action in several
Hollywood films
27
[http://recombu.com/mobile/article/hands-on-review-motorola-dynatac-8000x_M13438.html]
28 [https://www.engadget.com/2008/02/18/analog-cellular-networks-r-i-p-1983-2008/]
Video Demo of 1G Phone
A Tektronix CMD 80 cellular service monitor is used to
simulate a cell tower.
https://www.youtube.com/watch?v=GH_t2ZIX794
29
http://gizmodo.com/it-only-costs-400-to-build-your-own-cell-phone-network-1791282980
http://hackaday.com/2017/01/16/shmoocon-2017-dig-out-your-old-brick-phone/
30
History: 2G Cellular
y The first-generation (1G) systems introduced in the 1980s
were characterized by analog speech transmission.
y The second generation (2G) of cellular systems, first deployed
in the early 1990’s, were based on digital communications.
y The shift from analog to digital was driven by its higher capacity
and the improved cost, speed, and power efficiency of digital
hardware.
y 1991: US Digital Cellular (USDC – IS-54 > IS-136)
y Three times capacity of AMPS via digital modulation, speech coding,
and TDMA
y While second generation cellular systems initially provided mainly
voice services, these systems gradually evolved to support data
services such as email, Internet access, and short messaging.
32
GSM World Coverage Map
33
[http://www.worldtimezone.com/gsm.html]
History (Thailand)
NMT450
y 1G
y 1986 (2529): NMT470 (TOT)
y Nordic Mobile Telephone System @ 470MHz
y AMPS (Advanced Mobile Phone System)
y 1990 (2533): Cellular 900 (AIS)
y Worldphone 800 (TAC)
[http://3g.siamphone.com/articles/2009/3g/page.htm]
34
[http://www.scriptdd.com/diary/windows_live_for_mobile.html]
GSM Enhancement
y Want to deliver data as well as voice.
y 2.5G: General Packet Radio Service (GPRS)
y Provide connectivity to IP networks (Internet).
y Each slot can handle up to 20 kb/s. Each user may be allocated
up to 8 slots
y Data rates up to about 160 kb/s per user are possible.
y A single time slot may be shared by multiple users for
transferring packet mode data.
y 2.75G: Enhanced Data Rates for GSM Evolution
(EDGE)
y Support IP-based services in GSM at rates up to 384 kb/s
35
2.5G: GPRS
y General Packet Radio Service
y The first commercial launches for GPRS took place in 2001.
y Construction of a packet switched core network, to run
alongside the circuit switched network that was originally
built for GSM.
y "always on" connection that remains active as long as the phone
is within range of the service.
y A single time slot may be shared by multiple users for
transferring packet mode data.
36
2.5G: GPRS
y A good approximation for throughput in “average” conditions
is 10 Kbps per time slot. [Korhonen, 2003]
y Especially suitable for non-real-time applications, such as e-
mail and Web surfing.
y Bursty data is well handled with GPRS, as it can adjust the
assigned resources according to current needs.
y Not well-suited for real-time applications
y Resource allocation in GPRS is contention based
y Cannot guarantee an absolute maximum delay.
37
2.75?G: EDGE
y Enhanced Data Rates for GSM Evolution
y Originally this acronym stood for Enhanced Data rates for GSM Evolution,
but now it translates into Enhanced Data rates for Global Evolution,
as the EDGE idea can also be used in systems other than GSM [Korhonen,
2003]
y Higher modulation efficiency
y Eight-phase shift keying (8PSK)
y Can only be used effectively over a short distance.
y For wide area coverage, the old GMSK (Gaussian minimum shift key) is still needed.
y Only requires a software upgrade to base stations
y if the RF amplifiers can handle the non-constant envelope modulation with
EDGE’s relatively high peak-to-average power ratio (PAPR).
y EDGE is popular in North America, where the allocation of carrier
frequencies has made it hard for GSM operators to upgrade to UMTS.
38
The Original iPhone (also known as the iPhone 2G)
y On January 9th, 2007, Steve Jobs unveiled the iPhone.
y Quad-band GSM/GPRS/EDGE (850, 900 1,800 1,900
MHz)
y Released in June of 2007. Discontinued in 2008.
39
News: 2017
y January 9th, 2017 is a
milestone day in the
technology world: It’s the
10th anniversary of Apple's
iPhone.
y On January 1st, 2017, AT&T shut down its 2G wireless
network.
y In the U.S. the original iPhone was available exclusively on
AT&T.
y This means that AT&T has killed the original iPhone.
y Almost 10 years to the day after it was unveiled, the original
iPhone is dead. [ http://about.att.com/innovationblog/2g_sunset ]
[ https://www.engadget.com/2017/01/17/att-shuts-down-2g-network/ ]
40 [ http://www.makeuseof.com/tag/original-iphone-dead-because-att/ ]
[ http://www.theverge.com/2017/1/17/14298744/att-2g-network-shutdown-iphone ]
Motivation
41
43
[https://www.ericsson.com/mobility-report]
3G
y Studies started even before the earliest 2G systems arrived on the
market.
y International Mobile Telecommunications-2000 (IMT-2000)
y A subgroup of the International Telecommunication Union (ITU)
y Published a set of performance requirements of 3G (for both packet-
switched and circuit-switched data):
y A minimum data rate of 144 Kbps in the vehicular environment
y A minimum data rate of 384 Kbps in the pedestrian environment
y A minimum data rate of 2 Mbps in the fixed indoor and picocell environment
y There are several wireless standards and systems that qualify as
third generation (3G) systems.
y UMTS
y CDMA2000
44
(Collaboration between groups of telecommunications
3gpp2.org
3gpp.org
Scope
3G: IMT-2000 W-CDMA CDMA2000
International Mobile
Telecommunications
systems The 3G technologies standardized by 3GPP are often
referred to collectively as WCDMA.
3GPP uses two other acronyms to describe its
International specifications:
Telecommunication UMTS (Universal Mobile Telecommunications System)
Union applies to the entire cellular network contained in
ITU-R hundreds of 3GPP specifications; and
UTRAN (Universal Terrestrial Radio Access Network)
R = Radiocommunication Standardization Sector refers to the collection of network elements, and their
interfaces, used for transmission between mobile
45 terminals and the network infrastructure.
cdma2000
y Multicarrier, direct-sequence CDMA FDD system.
y Backward-compatible with its previous 2G iteration IS-95
(cdmaOne).
y CDMA2000 1X (IS-2000)
y also known as 1x and 1xRTT
y 1x = Spreading Rate 1 = use the same chip rate of IS-95 (i.e., 1.2288
Mcps).
y Same RF bandwidth as IS-95: a duplex pair of 1.25 MHz radio channels.
y Core CDMA2000 wireless air interface standard.
y Almost doubles the capacity of IS-95 by adding 64 more traffic
channels to the forward link, orthogonal to (in quadrature with) the
original set of 64.
46
UMTS
y Universal Mobile Telecommunications System (UMTS)
y The research activity on UMTS started in Europe at the beginning of
the 1990s.
y Even before the earliest 2G systems arrived on the market
y Designed to support wideband services with data rates up to 2Mbit/s.
y Developed from GSM
y Keep the core network more-or-less intact
y Change the air interface to use CDMA
y Compatibility between UMTS and GSM:
y Most UMTS mobiles also implement GSM, and the network can hand them
over from a UMTS base station to a GSM one if they reach the edge of the
UMTS coverage area.
y However, network operators cannot implement the two systems in the same
frequency band, so they are not fully compatible with each other.
47
3G 4 4S 5 5C
48
HSPA 3.5G?
y High Speed Packet Access (HSPA) is a collection of two
mobile telephony protocols
y High Speed Downlink Packet Access (HSDPA) and
y High Speed Uplink Packet Access (HSUPA)
y Extend and improve the performance of existing
WCDMA/UMTS protocols.
y Current HSDPA deployments support down-link speeds of
1.8, 3.6, 7.2 and 14.0 Megabit/s.
y Many HSPA rollouts can be achieved by a software
upgrade to existing 3G networks, giving HSPA a head start
over WiMAX, which requires dedicated network
infrastructure. 3.9G?
49
y There is also a further standard, Evolved HSPA (HSPA+).
HSPA+
y HSPA+ provides data rates up to
y 168 Megabits per second (Mbit/s) to the mobile device (downlink) and
y 22 Mbit/s from the mobile device (uplink).
y Technically these are achieved through the use of MIMO and higher
order modulation (64QAM) or combining multiple cells into one with a
technique known as Dual-Cell HSDPA.
y The 168 Mbit/s and 22 Mbit/s represent theoretical peak speeds.
y Only in very good radio conditions (very close to cell tower) or if the
terminal and network both support either MIMO or Dual-Cell HSDPA
y The actual speed for a user will be lower.
y Deliver significant battery life improvements and dramatically quicker
wake-from-idle time – delivering a true always-on connection.
50
3G in Thailand: HSPA, HSPA+
y Truemove H: 850 MHz
y Launched HSPA+ in September 2011
y DTAC: 850 MHz
y Launched HSPA+ in September 2011
y AIS: 900 MHz
y Launched HSPA+ in April 2011
y TOT3G: 2.1 GHz
y Different bands
51 [http://en.wikipedia.org/wiki/List_of_HSPA%2B_networks]
UMTS to LTE
associations (partners))
Scope
3G: IMT-2000 UMTS (W-CDMA) CDMA2000
International Mobile
Long Term
Telecommunications Evolution
systems 3GPP Release 8 and 9
International
Telecommunication
Union
IMT-Advanced LTE-Advanced LTE-A
3GPP Release 10
ITU-R
R = Radiocommunication Standardization Sector
52
LTE Around the World (in the beginning)
54
[http://www.radio-electronics.com/info/cellulartelecomms/lte-long-term-evolution/lte-frequency-spectrum.php]
Operating bands specified for LTE in
3GPP below 1 GHz
57
[http://www.apple.com/iphone/LTE/]
LTE on iPhone 5
59
Important?
y July 2011: A survey by Retrevo
found that (34% of) iPhone 4
owners actually believe that the “4”
in their iPhone 4 stands for 4G
y Dec 2011: A survey from Analysys
Mason showed that
y 46 percent of iPhone 4 users believe
that they already have 4G capability,
even though they don't.
y Over 50% of respondents were
unsure as to their network
technology.
[http://www.reuters.com/article/2012/09/14/us-apple-europe-spectrum-idUSBRE88D0NX20120914]
[http://www.retrevo.com/content/blog/2011/07/confusion-and-skepticism-may-impede-4g-adoption]
60 [http://www.pipelinepub.com/0112/OSS_BSS/Race-To-4G-1.php]
[http://www.analysysmason.com/About-Us/News/Insight/Insight_How_to_sell_4G_Dec2011/]
Upto 150Mbps
download speed.
62
y Announced on September 7, 2016.
y Released on September 16, 2016.
[ https://www.ifixit.com/Teardown/iPhone+7+Teardown/67382 ]
[ https://www.qualcomm.com/products/snapdragon/modems/4g-lte/x12 ]
64 [ http://www.fudzilla.com/news/mobile/41616-ifixit-reveals-qualcomm-modem-in-iphone-7-plus ]
[ http://appleinsider.com/articles/16/11/18/apple-confirmed-limiting-iphone-7-qualcomm-modem-to-keep-performance-on-par-with-intel-chip ]
Transceivers in iPhone 7 Model A1660
65
[ https://d3nevzfk7ii3be.cloudfront.net/igi/NlUTLePVEVEnEoqy.huge ]
66
[ http://www.techinsights.com/about-techinsights/overview/blog/apple-iphone-7-teardown/ ]
iPhone 7: Two modem suppliers
y A1660 and A1661 models use Qualcomm modem chips
y Sold in the U.S. (Verizon and Sprint), China, Hong Kong,
Puerto Rico
y A1778 and A1784 models use Intel modem chips
y Sold in the U.S. (AT&T and T-Mobile) and most other countries
in the world.
y 30-75% poorer network speeds
y Tested by New York-based Cellular Insights
y Does not support CDMA
[ https://www.theguardian.com/technology/2016/oct/21/32gb-
iphone-7-slower-than-more-expensive-versions-tests-show ]
[ http://cellularinsights.com/iphone7/ ]
[ http://appleinsider.com/articles/16/11/18/apple-confirmed-
limiting-iphone-7-qualcomm-modem-to-keep-performance-on-par-with-
67 intel-chip ]
[ https://www.youtube.com/watch?v=LnttbKInCQI ]
68
Apple vs. Qualcomm
y Jan 17, 2017: The lawsuit, filed by
the Federal Trade Commission
(FTC), accused Qualcomm of anti-
competitive movements to keep
Apple using its modems.
y Jan 20, 2017: Apple sued Qualcomm
to the tune of $1 billion over a
variety of patent loyalties
y “Qualcomm insists on charging Apple
at least five times more in payments
than all the other cellular patent
licensors we have agreements with
http://www.iphonehacks.com/2017/01/ftc-qualcomm-apple-modems.html
http://www.iphonehacks.com/2017/01/apple-suing-qualcomm-modems.html
http://www.iphonehacks.com/2017/01/qualcomm-says-apples-claims-1-billion-lawsuit-baseless.html
https://www.youtube.com/watch?v=kiL867OEkXQ
combined”
https://www.cnet.com/news/qualcomm-forced-apple-into-exclusive-chip-deal-ftc-says/
https://www.youtube.com/watch?v=9FEvvZ9o0ec
https://techcrunch.com/2017/01/21/qualcomm-fires-back-against-apple-over-lawsuit-and-ftc-action/
69
70 [ http://www.forbes.com/sites/aarontilley/2016/09/16/inside-iphone-7-teardown-intel-modem/ ]
LTE Adoption
71 [ https://en.wikipedia.org/wiki/LTE_%28telecommunication%29 ]
72 [http://www.worldtimezone.com/4g.html]
4G in Thailand
73
74
[ https://th.wikipedia.org/wiki/ÃÁ·É¨_°ÈÁÈ_°¤¤¼·Á´É ]
DTAC
Freq. Genera Launch End of
BW Tech. Status
Band tion Date License
·Ê» ¡.«.
1800 MHz 5 MHz GSM/GPRS/ED
GE 2G Active 1 .¥.
2532
¡.«.
2561 (´¤µ)
·Ê» ¡.«.
2100 MHz 10 MHz UMTS/HSPA+
42Mbps 3G Active 8 ¡..
2556
¡.«. 2570
(Ä°»µ)
·Ê» ¡.«.
850 MHz 10 MHz UMTS/HSPA+
42Mbps 3G Active 8 ¡..
2556
¡.«.
2561 (´¤µ)
·Ê» ¡.«.
1800 MHz 20 MHz LTE 150Mbps 4G Active ¡.«. 2558
2561 (´¤µ)
·Ê» ¡.«.
2100 MHz 5 MHz LTE 42Mbps 4G Active 8 ¡..
2556
¡.«. 2570
(Ä°»µ)
·Ê» ¡.«.
75 1800 MHz 20 MHz ŤnÅo Äo µ
2561 (´¤µ)
[ https://th.wikipedia.org/wiki/°ªµr_°·Ã¢¦r_Á°¦rª· ]
AIS
Freq. Launch End of
BW Tech. Generation Status
Band Date License
·Ê» ¡.«.
GSM/GPR 1 .. ¡.«.
900 MHz 5 MHz S/EDGE 2G Active
2533
2574
(Ä°»µ)
·Ê» ¡.«.
UMTS/HS 8 ¡.. ¡.«.
2100 MHz 5 MHz PA+ 3G Active
2556
2570
(Ä°»µ)
·Ê» ¡.«.
LTE 10 .. ¡.«.
1800 MHz 15 MHz 150Mbps 4G Active
2558
2576
(Ä°»µ)
·Ê» ¡.«.
10 .. ¡.«.
2100 MHz 10 MHz LTE 4G Active
2558
2570
(Ä°»µ)
True Move-H
Freq. Genera Launch End of
BW Tech. Status
Band tion Date License
·Ê» ¡.«.
UMTS/HSPA+ 8 ¡..
850 MHz 15 MHz 42Mbps 3G Active
2556
¡.«. 2568 (´µ
Ánµ)
UMTS/HSPA+ 8 ¡..
2100 MHz 5 MHz 21Mbps 3G Active
2556
¡.«.
·Ê» ¡.«.
2570
8 ¡..
2100 MHz 10 MHz LTE 75Mbps 4G Active
2556
¡.«. (Ä°»µ)
1 ¡..
900 MHz 5 MHz LTE 4G Active
2559
¡.«. ·Ê» ¡.«.
2574
GSM/GPRS/E
900 MHz 5 MHz DGE 2G Active ¡.«. 2533 (Ä°»µ)
·Ê» ¡.«.
4 ..
1800 MHz 15MHz LTE 150Mbps 4G Active
2558
¡.«. 2576
77 (Ä°»µ)
78
Mobile Operators in Thailand
Freq. Band DTAC TRUE Move-H AIS TOT CAT
15 MHz (3G)[2025] 15 MHz
850 MHz 10 MHz (3G)[2018]
(from CAT) (3G)[2025]
5 MHz (2G)[2031] 5 MHz (2G)[2031]
900 MHz 5 MHz (4G)[2031] 5 MHz (4G)[2031]
5 MHz (2G)[2018]
1800 MHz 20 MHz (4G)[2018]
15 MHz (4G)[2033] 15 MHz (4G)[2033] 20 MHz [2018?]
79
80
[Zarrinkoub, 2014]
Peak data rates
81
[Zarrinkoub, 2014]
82
The evolution of LTE
[ https://www.youtube.com/watch?v=8UFODEdoNHw ]
83
[ https://opensignal.com/reports/2016/11/state-of-lte ]
OpenSignal's
availability metric
tracks the
proportion of
time users have
access to a
particular
84 network.
[ https://opensignal.com/reports/2016/11/state-of-lte ]
85
5G
86
5G field testing, trials, research,
development world coverage map
87 [http://www.worldtimezone.com/5g.html]
ECS 455 Chapter 1
Introduction
1.2 Wireless Channel (Part 1)
Office Hours:
BKD, 6th floor of Sirindhralai building
Tuesday 14:20-15:20
Dr.Prapun Wednesday 14:20-15:20
1 prapun.com/ecs455 Friday 9:15-10:15
Wireless Channel
y Large-scale propagation effects
1. Path loss
[ ]
2. Shadowing
y Small-scale propagation effects [ ]
Path loss
y Caused by
y dissipation of the power radiated by the transmitter
y effects of the propagation channel
y Models generally assume that it is the same at a given
transmit-receive distance.
y Variation occurs over large distances (100-1000 m)
[ ]
[Goldsmith, 2005, Fig 2.1]
[ ]
4
Path Loss (PL)
7UDQVPLWWHGSRZHU Pt
PL
$YHUDJHUHFHLYHGSRZHU Pr
J
Pr §d ·
Simplified Path Loss Model Pt
K¨ ¸
©d ¹
P Captures the essence of
J
ORJ r ORJ Kd J ORJ d signal propagation without
[dB] Pt resorting to complicated
path loss models, which are
y K is a unitless constant which depends on only approximations to the
the antenna characteristics and real channel anyway!
the average channel attenuation
ఒ ଶ
y for free-space path gain at distance d0 assuming
ସగௗబ
omnidirectional antennas
y d0 is a reference distance for the antenna far-field (Near-field has scattering
phenomena.)
y Typically 1-10 m indoors and 10-100 m outdoors.
y ɀ is the path loss exponent.
8
[Goldsmith, 2005, Table 2.2]
Path Loss Exponent J
y 2 in free-space model
y 4 in two-ray model
[Goldsmith, 2005, eq. 2.17]
[ ]
11
Shadowing (Analogy)
[ http://spacegrant.montana.edu/MSIProject/NDVI.html ]
[ https://www.flickr.com/photos/pokoroto/4045274462 ]
12
Shadowing (Analogy)
https://brightside.me/creativity-photography/16-truly-remarkable-photos-everyone-needs-to-see-113155/#image7456160
Shadows
thousands
of miles long
cast
by clouds
on Earth’s
surface.
13
&RQWRXUVRI&RQVWDQW5HFHLYHG3RZHU
14
Log-normal shadowing
y Random variation due to blockage from objects in the signal
path and changes in reflecting surfaces and scattering objects
Æ random variations of the received power at a given
distance 4 – 13 dB with higher values in urban
areas and lower ones in flat
Pt rural environments.
ORJ & P V
Pr
in dB
Object 2
K¨ ¸
Object 3
r
i
Pt ©d ¹ i
J
P §d ·
ORJ r ORJ K ¨ ¸ ¦ORJ D i
Pt ©d ¹ i
By CLT, this is
16 approximately
Gaussian
PDF of Lognormal RV
y Consider a random variable
Pt
R
Pr
y Suppose
ORJ R & P V Here, it should be clear
why the unit of V is in
dB.
y Then,
§ ORJ r P ·
¨
° V
¸
e © ¹
r !
fR r ® SV OQ r
°
¯ RWKHUZLVH
18
PDF of Lognormal RV (Proof)
Suppose c ORJ b Y & P V
c
Let X c ORJ bY Note that X c ORJ bY OQ Y k OQ Y
X
OQ b
c
Then, Y e k where k
OQ b
Recall, from ECS315 that to find the pdf of Y = g(X) from the pdf of
X, we first find the cdf of Y and then differentiate to get its pdf:
ª Xk º
FY y P >Y d y @ P «e d y » P ª¬ X d k OQ y º¼ FX k OQ y
¬ ¼
§ k OQ y P ·
¨
d k k © V
¸
¹
fY y FX k OQ y f X k OQ y e
dy y SV y
19
dx
f X x dx fY y dy fY y fX x
dy
k
This gives fY y f X c ORJ b y (same as what we found earlier).
y
20
Ray tracing (a prelude)
y Approximate the solution of Maxwell’s equations
y Approximate the propagation of electromagnetic waves by
representing the wavefronts as simple particles.
y Thus, the reflection, diffraction, and scattering effects on the
wavefront are approximated using simple geometric equations
instead of Maxwell’s more complex wave equations.
y Assumption: the received waveform can be approximated by the
sum of the free space wave from the transmitter plus the reflected
free space waves from each of the reflecting obstacles.
x t Pt FRV S f c t y t "
Tx Rx
21
d
22
Review: Power Calculation
ଶ
ଶ
Periodic with period ܶ
்బ
ೖ
where the ܽ ݐare orthogonal
(e.g., do not overlap in the
frequency domain)
23
ଶSౡ ௧ ଶ
where the fk are distinct
ೖ
where the ܣ ݂ േ ݂ ’s do not
overlap
24
Power Calculation: Additional Formula
ଶ
ͳ
ܽଵ ʹߨ݂ ݐ ߶ଵ ൌ ȁܽ ݁ థభ ܽଶ ݁ థమ ȁଶ
ܽଶ ʹߨ݂ ݐ ߶ଶ ʹ ଵ
ͳ ͳ
ൌ ܽଵଶ ܽଶଶ ܽଵ ܽଶ ߶ଶ െ ߶ଵ
ʹ ʹ
25
y Multipath Reception
D § § d ··
y t Pt FR V ¨ S f c ¨ t ¸ ¸
d © © c ¹¹
26
Ex. One reflecting wall (1/4)
y There is a fixed antenna transmitting the sinusoid ݐ ݔ, a
fixed receive antenna, and a single perfectly reflecting large
fixed wall.
y Assume that the wall is very large, the reflected wave at a
given point is the same (except for a sign change) as the free
space wave that would exist on the opposite side of the wall if
the wall were not present
Wall
Tx Rx
27
d
n
D § § r ··
y t ¦R k Pt FRV ¨ S f c ¨ t k ¸ ¸
© c ¹¹
k rk ©
Tx Rx
28
d
Ex. One reflecting wall (3/4)
D § § d ·· D § § w d · ·
y t Pt FRV ¨ S f c ¨ t ¸ ¸ Pt FRV ¨ S f c ¨ t ¸ S ¸
d © © c ¹ ¹ w d © © c ¹ ¹
§ § D · § D · D ·
Pt ¨ ¨ ¸ ¨ FRV 'I ¸¸
¨ © d ¹ © w d ¹¸
Py
© d w d ¹
w d
'I S f c S S w d S form constructive
c O
and destructive
interference
pattern
w
Wall
Tx Rx
29
d
u
u
u
f 0+]
3 G u
w NP
u
u
30
Ex. Two-Ray Model r r
Delay spread = c c
D § § r ·· D § § r ··
y t Pt FRV ¨ S f c ¨ t ¸ ¸ Pt FRV ¨ S f c ¨ t ¸ ¸
r © © c ¹ ¹ r © © c ¹¹
r r r
Py D j S f c
r
D j S f c D D j S f c
e c
e c
e c
Px r r r r
ht hr
Tx Rx
Floor (Ground)
3\ 3[
f 0+]
ht P
hr P
G [
32
Ex. Two-Ray Model
3\ 3[
f 0+]
ht P
hr P
ORJG
33
3\ 3[
f 0+]
ht P
hr P
ORJG
34
Ex. Two-Ray Model (Approximation)
ht hr ht hr d ht hr
Py D D j S D j S
| e O
| e Od
Px r r d
§D · § h h · D h h § SD ht hr ·
| ¨ ¸ ¨ j S t r ¸ j S t r ¨ ¸
©d¹ © Od ¹ d Od © Od
¹
§ GTxGRx ht hr ·
¨¨ ¸¸ v
© d ¹ d
d
ht hr
Tx Rx
Floor (Ground)
35
3\ 3[
ORJG
36
Ex. Two-Ray Model
([DFW
$SSUR[
1RUHIOHFWLRQ
8SSHUERXQG
/RZHUERXQG
3\ 3[
G [
37
dBm
y The range of RF power that must be measured in cellular phones
and wireless data transmission equipment varies from
y hundreds of watts in base station transmitters to
y picowatts in receivers.
y For calculations to be made, all powers must be expressed in the
same power units, which is usually milliwatts.
y A transmitter power of 100 W is therefore expressed as 100,000mW.
A received power level of 1 pW is therefore expressed as
0.000000001mW.
y Making power calculations using decimal arithmetic is therefore
complicated.
y To solve this problem, the dBm system is used.
38
[Scott and Frobenius, 2008, Fig 1.1]
ܲ
ܲ W ൌ ͳͲଵ dBm
ͳmW
ͳͲͲܹ݉
ͳͲଵ ൌʹͲ dBm
ͳmW
dB and dBm
y The decibel scale expresses factors or ratios logarithmically.
y Unitless dB value
P
y Represent power ratio: ORJ
P
y dB value with a unit
y Represent the signal power itself:
P P
P > G%: @ ORJ P > G%P @ ORJ
: P:
y Note that P > G%P @ P > G%: @
40
Remark
y Adding dB values corresponds to multiplying the underlying
factors, which means multiplying the units if they are
present.
y It is therefore appropriate to add unitless dB values to a dB
value with a unit (such as dBm)
y The result is still referred to that unit.
y Ex: 17 dBm + 13 dB – 6 dB = 24 dBm
y Correspond to 50 mW × 20 / 4 = 250 mW.
41
D § § r· ·
Pt FRV ¨ S f c ¨ t ¸ I ¸
r © © c¹ ¹
y If moving, r becomes r(t).
y If moving away at a constant velocity v, then r t r vt
D § § r vt · · D § § v· r ·
FRV ¨ S f c ¨ t I FRV ¨ S ¨ f c f c ¸ t S f c I ¸
r t © © c ¸¹ ¸
¹ r t © © c¹ c ¹
Frequency shift
v
'f
42
O
Review: Instantaneous Frequency
For a generalized sinusoid signal
A FRV T t
the instantaneous frequency at time t is given by
d
f t T t
S dt
§ r t ·
When T t S f c ¨ t ¸ I
© c ¹
d f d d
f t T t fc c r t fc r t
S dt c dt O dt
Frequency shift
43
Big Picture
Transmission impairments in cellular systems
(Path Loss)
Office Hours:
BKD, 6th floor of Sirindhralai building
Tuesday 14:20-15:20
Dr.Prapun Wednesday 14:20-15:20
1 prapun.com/ecs455 Friday 9:15-10:15
Electromagnetic Spectrum
3 MHz 3 GHz
100 m 10 cm
c fO
u PV Wavelength
Frequency
2
[http://www.britannica.com/EBchecked/topic-art/585825/3697/Commercially-exploited-bands-of-the-radio-frequency-spectrum]
Radio-frequency spectrum
y Commercially exploited bands
Cellular Bands
y Cellular phone networks worldwide use a portion of the radio
frequency spectrum designated as ultra high frequency (UHF) (300
MHz to 3 GHz)
y The UHF band is also used for television, GPS, Wi-Fi, and Bluetooth
transmission.
y Due to historical reasons, radio frequencies used for cellular networks differ
in the Americas, Europe, and Asia.
y Frequency bands that are currently identified for IMT (International
Mobile Telecommunications) in all three ITU Regions:
y 450 – 470 MHz Additional frequency bands identified for IMT on a Regional or National
basis:
y 790 – 960 MHz 698-790 MHz (Region 2)
y 1710 – 2025 MHz 610 – 790 MHz (9 countries in Region 3: Bangladesh, China, Rep. of
y 2110 – 2200 MHz Korea, India, Japan, New Zealand, Papua New Guinea, Philippines
and Singapore.)
y 2300 – 2400 MHz 3400 – 3600 MHz (Over 80 Administrations in Region 1 plus 9 in
y 2500 – 2690 MHz Region 3 including India, China, Japan and Rep. of Korea).
4 [ https://www.itu.int/ITU-D/tech/MobileCommunications/Spectrum-IMT.pdf ]
ITU Regions
y The ITU divides the world into three ITU regions for the
purposes of managing the global radio spectrum.
Region 1
Region 2
Region 3
2016
8 [ https://www.ntia.doc.gov/category/spectrum-management ]
Thailand Freq. Allocations
[ https://www.youtube.com/watch?v=VzD-L24vg4U ]
9
10
[ nbtc.go.th/spectrum_management/¤n/µ¦µÎµ®¨ºÉªµ¤¸É®nµ·.aspx ]
National Table of Frequency Allocation
11
[ nbtc.go.th/spectrum_management/¤n/µ¦µÎµ®¨ºÉªµ¤¸É®nµ·.aspx ]
12
http://www.ntc.or.th/uploadfiles/freq_chart_thai.htm
Spectrum Allocation
y Spectrum is a scarce resource.
y “Radio spectrum will be the first of our finite resources to run
out, long before oil, gas or mineral deposits.”
y Spectrum is allocated in “chunks” in frequency domain.
y “Chunks” are licensed to (cellular/wireless) operators.
y Within a single cellular operator, the chunk is further divided
into many channels.
y Each channel has its own band of frequency.
y Mobile networks based on different standards may use the
same “frequency chunk”.
y For example, AMPS, D-AMPS, N-AMPS and IS-95 all use the
800 MHz “frequency chunk”.
y This is achieved by the use of different channels.
13
14
Nov 2015: Thailand 1800MHz Auction
y | 40bn baht
y 15 MHz (UL) + 15 MHz (DL)
y | 1.3 billion baht per MHz
y | 74 million baht per MHz per year
(| 2.5u)
15
(forfeit)
16
Interesting Book
y Spectrum Wars: The Policy and
Technology Debate
17
Office Hours:
BKD, 6th floor of Sirindhralai building
Tuesday 14:20-15:20
Dr.Prapun Wednesday 14:20-15:20
18 prapun.com/ecs455 Friday 9:15-10:15
Unlicensed bands
y Frequency bands that are free to use
y according to a specific set of etiquette rules.
y The purpose of these unlicensed bands is to encourage
innovation and low-cost implementation.
y Many extremely successful wireless systems operate in
unlicensed bands, including wireless LANs, Bluetooth,
and cordless phones.
y Major difficulty:
y If many unlicensed devices in the same band are used in close
proximity, they generate much interference to each other,
which can make the band unusable.
19
22
Section 17.4.6.3 (Channel
Numbering of operating
channels) of the IEEE Std 802.11
(2012) states “In a multiple cell
network topology, overlapping
and/or adjacent cells using
different channels can operate
simultaneously without
interference if the distance
between the center frequencies
is at least 25 MHz.”
23 [http://en.wikipedia.org/wiki/IEEE_802.11]
24
Unlicensed 60 GHz Frequency Band
y A lot of bandwidth
available
Worldwide
spectrum
availability
Office Hours:
BKD, 6th floor of Sirindhralai building
Tuesday 14:20-15:20
Dr.Prapun Wednesday 14:20-15:20
26 prapun.com/ecs455 Friday 9:15-10:15
LTE in unlicensed spectrum
y The use of the 4G LTE radio communications technology in
unlicensed spectrum, such as the 5 GHz band already
populated by Wi-Fi devices.
LTE
(5 GHz)
27
28
LTE-U: Controversy
y June 2015: Google sent the FCC a protest
y August 2015: the Wi-Fi Alliance also voiced opposition
y The technical concern with LTE-U is that LTE is a “rude”
technology.
y WiFi includes a “politeness protocol” that LTE lacks.
y WiFi will back off if it senses interference from other users.
y Eventually rude LTE operating in WiFi’s polite bands could take
over the band.
29
“A better
neighbor to Wi-Fi
than Wi-Fi itself ”
30
ECS455 Chapter 2
Cellular Systems
http://gizmodo.com/what-the-world-would-look-like-if-you-could-see-cell-ph-1481106278
Office Hours:
BKD, 6th floor of Sirindhralai building
Tuesday 14:20-15:20
Dr.Prapun Suksompong Wednesday 14:20-15:20
1 prapun.com/ecs455 Friday 9:15-10:15
Pre-Cellular System
y Regions need to be well-separated!
Region 1
Region 2
Region 4
Region 3
4
ECS455 Chapter 2
Cellular Systems
2.1 Frequency Reuse
Dr.Prapun Suksompong
5 prapun.com/ecs455
Reuse
Spectral
efficiency
Idea (1)
y Suppose the whole system has S = 70 frequency channels
y Pre-cellular:
All
70
All
70
All
“Capacity” of the system
70 = # users the system can
support simultaneously
= 70u3 = 210
8
G B
F A C
Idea (2) E D
S = 70 frequency channels
y Cellular:
y Split 70 channels into 7 groups (A,B,C,D,E,F,G).
y Each group has m = 10 channels. Cells using the same groups
are far apart.
G B
Note:
10
Less interference
G B F 10 A C Cells can overlap.
(Recall that Pr is
10 10 10 10
inversely F 10 A C E D
proportional to dJ.)
10 10 10 G 10 B 10
E D
10 10 F 10 A 10 C
“Capacity” of the system
10 10 10 = # users the system can
E D
support simultaneously
10 10 = 70u3 = 210
9
G B
10
Cluster G B F 10 A C
10 10 10 10
F 10 A C E D
10 10 10 G 10 B 10 Reuse Distance (D)
E D = minimum distance
10 10 F 10 A 10 C between the centers
Cluster size 10 10 10 of cells that use the
E D
N=7 same channel set
10 10
10
Idea (4)
y To support more users (increase capacity), simply use smaller
cell size (area).
G B
10 10
G B F A C
10 10 10 10 10
G B F C A E D
10 10 10 10 G 10 B 10
10
G B F A C E D
10 10 10 10 10 G 10 B 10 F 10 A 10 C
F A C E D
10 10 10 10 B 10 F 10 A 10 C 10 E 10 D 10
E D G
10 10 F 10 A 10 C 10 E 10 D 10 G 10 B 10
10 10 10 G 10 B 10 F 10 A 10 C
E D
10 10 F 10 A 10 C 10 E 10 D 10
10 10 10 G 10 B 10
E D
10 10 F 10 A 10 C “Capacity” of the system
10 10 10
E D = # users the system can
10 10
support simultaneously
>> 210
11
[ http://rathishnair.com/techblog/cell-handoff-basics/ ]
12
Can we keep reducing the cell size?
y While smaller cells generally increase capacity, they also have
their disadvantages.
y Smaller cell size increases the rate at which handoffs occur,
which increases the dropping probability if the percentage of
failed handoffs stays the same.
y Smaller cells increase the load on the backbone network.
y More cells per unit area requires more base stations, which
can increase system cost.
y Propagation characteristics typically change as cell size
shrinks, so the system does not scale perfectly.
another.
y Connect the entire cellular system to
the PSTN (public switched telephone
network) for landline calls and
Internet access.
16
How a Cell Phone Call Works
17
[http://cellphones.org/blog/how-cell-phone-calls-work/]
19
(æ´«r£µ¡ºÊ·)
Terrestrial TV in BKK
зњѥєщѝ Ѩѷ Ѥррѥцѱъіъ Ѥћьҙ VHF.(Low Band) зњѥєщѝѷѨ Ѥррѥцѱъіъ Ѥћьҙ UHF.(Band 5)
Channel Picture Channel Picture
Bandwidth. Audio Carrier. Bandwidth. Audio Carrier.
. Carrier. . Carrier.
2 47 - 54 48.25 53.75 35 582 - 590 583.25 588.75
3 54 - 61 55.25 60.75 MUX 2 36 590 - 598 591.25 596.75
4 61 - 68 62.25 67.75 37 598 - 606 599.25 604.75
зњѥєщѝѨѷ Ѥррѥцѱъіъ Ѥћьҙ VHF.(Hight Band) 38 606 - 614 607.25 612.75
Channel Picture 39 614 - 622 615.25 620.75
Bandwidth. Audio Carrier.
. Carrier. MUX 3 40 622 - 630 623.25 628.75
5 174 - 181 175.25 180.75 41 630 - 638 631.25 636.75
6 181 - 188 182.25 187.75 42 638 - 646 639.25 644.75
7 188 - 195 189.25 194.75 43 646 - 654 647.25 652.75
8 195 - 202 196.25 201.75 MUX 4 44 654 - 662 655.25 660.75
9 202 - 209 203.25 208.75 45 662 - 670 663.25 668.75
10 209 - 216 210.25 215.75 46 670 - 678 671.25 676.75
11 216 - 223 217.25 222.75 47 678 - 686 679.25 684.75
12 223 - 230 224.25 229.75 48 686 - 694 687.25 692.75
зњѥєщѝ Ѩѷ Ѥррѥцѱъіъ Ѥћьҙ UHF.(Band 4) 49 694 - 702 695.25 700.75
Channel Picture 50 702 - 710 703.25 708.75
Bandwidth. Audio Carrier.
. Carrier. 51 710 - 718 711.25 716.75
26 510 - 518 511.25 516.75 MUX 1 MUX 5 52 718 - 726 719.25 724.75
27 518 - 526 519.25 524.75 53 726 - 734 727.25 732.75
28 526 - 534 527.25 532.75 54 734 - 742 735.25 740.75
29 534 - 542 535.25 540.75 55 742 - 750 743.25 748.75
30 542 - 550 543.25 548.75 56 750 - 758 751.25 756.75
31 550 - 558 551.25 556.75 57 758 - 766 759.25 764.75
32 558 - 566 559.25 564.75 58 766 - 774 767.25 772.75
33 566 - 574 567.25 562.75 59 774 - 782 775.25 780.75
20 34 574 - 582 575.25 580.75 60 782 - 790 783.25 788.75
Digital Terrestrial TV: MUX
“¤´¨·Á¡¨Èr” ®¤µ¥ªµ¤ªnµ æ nµ¥µ¤®¤µ¥² ¸É¦ª¦ª¤¦·µ¦¦³µ¥Á¸¥®¦º°Ã¦´«r ®¦º°¦·µ¦
´µ°ºÉĸɵΠÁÈ Á¡ºÉ°n®¦º°nµ¥°nµn°´µÁ¸¥ª¡¦o °¤´
(¦¤¦³µ´¤¡´r Ã¥µ¸ª·¥»Ã¦´«r®n¦³Á«Å¥)
(°rµ¦¦³µ¥Á¸¥Â¨³Â¡¦n£µ¡µµ¦³Â®n¦³Á«Å¥ åæ´«rÅ¥¡¸¸Á°)
MCOT Kiddie
21
22
Freq. Set: D2
ªµ¤¸ªÉ · ¥»µÎ ®¦´·µ¦Ã¦´«r£µ¡ºÊ ·Ä¦³·· °¨:
µ¦´¨»¤n n°ªµ¤¸É
23 [. ª.102-2555]
24
39.2 ÿčĕĀÜðćéĊ T-D2 44 48 28 32 40
4.4 ÝĆîìøïčøĊ D1 47 31 35 39 43
21.5
22.0
22.1
ðćÜöąñšć
úĞćðćÜ
đëĉî
D2
D2
D3
32
44
45
36
48
49
40
32
33
26
36
37
29
40
41
39.3
39.4
îøćíĉüćÿ/÷ĊęÜĂ
÷ąĀøĉęÜ
T-D2
T-D2
44
44 [. ª.102-2555]
48
48
28
28
32
32
40
40
ªµ¤¸ªÉ · ¥»µÎ ®¦´·µ¦Ã¦´«r£µ¡ºÊ ·Ä¦³·· °¨
(´¸É 2): µ¸®¨´
®¤µ¥Á¨ n°ªµ¤¸Éª· ¥»ÄÁ¦ºÉ°®¤µ¥ªÁ¨È ®¤µ¥¹ ®¤µ¥Á¨ n°ªµ¤¸Éª· ¥»É¹³Îµ¤µÄo µÁÈ µ¦´Éª¦µªn°¥»·µ¦¦´n
´µª·¥»Ã¦´«rĦ³Â°³¨È° ¦³®ªnµ¸É®¤µ¥Á¨ n°ªµ¤¸Éª· ¥» o µ®o µÁ¦ºÉ°®¤µ¥ªÁ¨È¥´Å¤n°»µÄ®o µÎ ¤µÄo µ
26
Frequency Reuse (Review)
y Cellular radio systems rely on an intelligent allocation and reuse of
channels throughout a coverage region
y Each cellular BS is allocated a group of radio channels to be used
within the corresponding cell.
y BSs in adjacent cells are assigned channel groups which contain
completely different channels than neighboring cells.
y By limiting the coverage area to within the boundaries of a cell,
the same group of channel may be used to cover different cells that
are separated from one another by distances large enough to keep
interference levels within tolerable limits.
y The minimum distance between two cells that use the same
channel set is called the reuse distance.
27 [Rappaport, 2002]
Cell Shape
y The actual radio coverage of a cell is known as the footprint.
y Determined from field measurements or propagation prediction models.
y In reality, it is not possible to define exactly the edge of a cell.
y Signal strength gradually reduces, and towards the edge of the cell
performance falls.
y MSs have different levels of sensitivity, this adds a further greying of the edge
of the cell.
y Impossible to have a sharp cut-off between cells.
y In some areas they may overlap, whereas in others there will be a hole in
coverage.
y Although the real footprint is amorphous in nature, a regular cell shape
is needed for systematic system design and adaptation for future growth.
28 [Rappaport, 2002]
Hexagonal cell shape
y Simplistic model of the radio coverage for each BS.
y Universally adopted
y Permit easy and manageable analysis
29
G B
F A C
Why hexagon? E D
y Omnidirectional BS antenna and free space propagation Æ
Circular radiation pattern.
y Adjacent circles cannot be overlaid upon a map without
leaving gaps or creating overlapping regions.
y Tessellating Cell Shapes: When considering geometric
shapes which cover an entire region without overlap and
with equal area, there are three sensible choices: a square, an
equilateral triangle, and a hexagon.
Diamond and
rectangles are also
tessellating shapes.
30
Why hexagon? (2)
y A cell must be designed to serve the weakest mobiles
within the footprint, and these are typically located at the
edge of the cell.
y For a given distance between the center of a polygon and its
farthest perimeter points, the hexagon has the largest area of
the three.
y By using the hexagon geometry, the fewest number of cells can
cover a geographic region
31
32
Since 1985, there are 14 types
y The hunt to find and classify the pentagons that can tile the
plane has been a century-long mathematical quest
33 [http://www.mathpuzzle.com/tilepent.html]
34
Frequency Reuse Plan
y The frequency reuse plan is overlaid upon a map to
indicate where different channel sets are used.
y Cells labeled with the same letter use the same group of
channels.
y Create co-channel interference
35
Clusters
y The total coverage area is divided into clusters.
y The number of cells (N ) in a cluster is called the cluster
size.
y Cells in a cluster collectively use the complete set of
available frequencies.
y No co-channel interference within a
cluster.
y Replicated over the coverage
area.
y Example: The picture shows
clusters of size N = 7, outlined in
bold.
36
Frequency Reuse (N = 4, N = 7)
Frequency reuse factor = 1/N
(Each cell within a cluster is only assigned 1/N of the total available channels in the system.)
37
“Capacity”
total # available duplex radio channels for the system
cluster size
S mN (Frequency reuse factor = 1/N)
Ability to handle
simultaneous #channels allocated to each cell
numbers of calls
AWRWDO AWRWDO S
“Capacity” C um u
AFHOO AFHOO N This formula suggests
that if the cluster size
N is reduced, more
total # cells in the system capacity is achieved.*
* Tradeoff: Small value of N may lead to large interference.
38
Cluster size (N)
y There are only certain cluster sizes and cell layouts which are
possible [Mac Donald,1979].
y N can only have values which satisfy
N i i u j j
where i and j are non-negative integers.
particular cell,
y move i cells along any chain of
hexagons and then
y turn 60 degrees counter-
clockwise and move j cells.
y Try N = 19
yi=3
yj=2
[Rappaport, 2002, Fig. 3.2]
[Goldsmith, 2005, Fig 15.6]
40
N=3
i=1
j=1
41
N=3
i=1
j=1
42
Locating co-channel cells (N = 3)
y To locate the nearest co-
channel neighbors of a
particular cell,
y move i cells along any chain of
hexagons and then
y turn 60 degrees counter-
clockwise and move j cells.
43
(i = 2, j =1)
(i = 2, j =0)
44
Co-Channel Cells: Ex. N = 3
A A A A A A
A A A A A A
A A A A A A
A A A A A A
A A A A A A
45
A
We can locate its co-channel A A A A A
cells by using the recipe in
the previous section.
A A A A A A
These cells are called co-
channel cells in the first
tier.
A A A A A A
A A A A A A
46
Co-Channel Cells: Ex. N = 3
A A A A A A
A
The recipe can be applied toA A A A A
each cell in the first tier to
find even more co-channel
cells. A A A A A A
These additional cells are
called co-channel cells in the
second tier.
A A A A A A
A A A A A A
47
A
The recipe can be applied toA A A A A
each cell in the first tier to
find even more co-channel
cells. A A A A A A
These additional cells are
called co-channel cells in the
second tier.
A A A A A A
A A A A A A
48
Co-Channel Cells: Ex. N = 3
A A A A A A
A A A A A A
A A A A A A
A A A A A A
A A A A A A
49
A A A A A A
A A A A A A
A A A A A A
A A A A A A
50
ECS455 Chapter 2
Cellular Systems
2.2 Co-Channel Interference
Dr.Prapun Suksompong
(Intercell)
Co-Channel Interference
y Frequency reuse Æ co-channel interference
y Consider only nearby
interferers.
y Power decreases rapidly as
the distance increases.
y In a fully equipped
hexagonal-shaped cellular
system, there are always K = 6
cochannel interfering cells in
the first tier.
2
Three Measures of Signal Quality
Pr
y For noise-limited systems, 615
Signal-to-noise (power) ratio PQRLVH
Signal-to-interference-plus-noise (power) ratio
Pr
y Consider both noise & interference: 6,15
PLQWHUIHUHQFH PQRLVH
y The best cellular system design places users that share the
same channel at a separation distance (as close as possible)
where the intercell interference is just below the maximum
tolerable level for the required data rate and BER.
y Good cellular system designs are interference-limited,
meaning that the interference power is much larger than the
noise power. P
6,5 r
Signal-to-interference (power) ratio PLQWHUIHUHQFH
3
We will soon revisit and use these numbers for some more specific
calculations
4
Review: Simplified Path Loss Model
Capture the essence of
J J
Pr § d · Pt E d k signal propagation
E¨ ¸ Pr v without resorting to
Pt ©d ¹ dJ dJ dJ complicated path loss
models, which are only
y E is a unitless constant which depends on the approximations to the
antenna characteristics and the average real channel anyway!
channel attenuation
y d0 is a reference distance for the antenna far-
field
y Typically 1-10 m indoors and 10-100 m
outdoors.
y ɀ is the path loss exponent.
y 2 in free-space model
y 4 in two-ray model [Goldsmith, 2005, eq.
2.17]
5
[Goldsmith, 2005, Table 2.2]
¦P
PLQWHUIHUHQFH
RIWKHi th LQWHUIHUHU
i
y Pr = the desired signal power from the desired base station
y Pi = the interference power caused by the ith interfering co-
channel cell base station.
y Often called the carrier-to-interference ratio: CIR.
6 [Rappaport, 2002]
SIR: N = 3
Hexagon
5
R
5 5
R
R
5
R 5 5
R
R
5
§ ·
$UHD u u¨ u Ru R¸ R | R
8 © ¹
SIR: N = 3 (Ignore co-channel cells
that are too far away)
SIR: N = 3
SIR: N = 7
SIR: N = 3
Approximation
y Consider only first tier.
y Worse-case distance
6,5 | J
§D ·
¦i ¨© Ri ¸¹
y Use the same D for Di
12
Approximation
y Consider only first tier.
y Worse-case distance
6,5 | J
§D ·
¦i ¨© Ri ¸¹
y Use the same D for Di
J
§D·
6,5 | J
| J ¨ ¸
§D· §D· K©R¹
¦i ¨© R ¹¸ K¨ ¸
©R¹
R i j ij R N
i R This distance, D,
j R
is called reuse
distance.
ϭϮϬq
A B
15
¦P
PLQWHUIHUHQFH
RIWKHi th LQWHUIHUHU
i
J
§D·
| J
| J ¨ ¸
§ Di · §D· K©R¹
¦i ¨© R ¹¸ K¨ ¸
©R¹
J
N
K
As the cell cluster size (N) increases, the spacing (D) between interfering
16 cells increases, reducing the interference.
Summary: Quantity vs. Quality
S = total # available duplex radio channels for the system
AWRWDO S J
“Capacity” C u 6,5 | N
AFHOO N Tradeoff K
m = # channels allocated to
each cell.
17
SIR: N = 7
Better approximation…
R J
6,5 | J J
DR DR D J
J J
Q Q Q J
D
$JDLQQ N
R
18
Comparison
6,5 | J
§ Di ·
¦
¨ ¸
i © R ¹
6,5>G%@
6,5 | QJ
'L
6,5 | J J
'
'
Q Q Q J
1
D
19
Q
R
SIR Threshold
y The SIR should be greater than a specified threshold for
proper signal operation.
[Schwartz, 2005, p 64]
20
ECS455 Chapter 2
Cellular Systems
2.3 Sectoring
Dr.Prapun Suksompong
1 prapun.com/ecs455
2
Sectorization (sectoring)
y Use directional antennas instead of omnidirectional antennas.
y When 120° sectorization is used, one cell that usually covers
360° is divided into three 120° regions.
y When 60° sectorization is used, one cell that usually covers
360° is divided into six 60° regions.
y These regions are called sectors.
Analogy: Flashlight
Sectoring (N = 7)
y Ex.
y With no sectoring, suppose we have m = 18 channels/cell
y With 120o sectoring, we have 6 channels/sector
y With 60o sectoring, we have 3 channels/sector
120° Sectoring 60° Sectoring
[Rappaport, 2002,
Figure 3.10]
5
The value of K changes from 6 to 1!
J
6,5 | N
Sectoring (N = 3 , 60q) K
K=2 K=1
6
J
6,5 | N
Sectoring (N = 3, 120q) K
K=3 K=2
Sectoring (N = 7, 120q)
Assuming seven-cell reuse,
for the case of 120q sectors,
the number of interferers in
the first tier is reduced from
six to two.
K=2
8
Summary:
S = total # available duplex radio channels for the system
Frequency reuse with cluster size N Path loss exponent
J
AWRWDO S S kR J §D· J
“Capacity” C u | ¨ ¸ N
AFHOO N Tradeoff I K u kD J K©R¹ K
m = # channels allocated to
each cell.
Omni-directional: K = 6
120q Sectoring: K = 2
60q Sectoring: K = 1
Sectoring
y Advantages
y Reduce interference by reducing K
y Increase SIR (better call quality).
y The increase in SIR can be traded with reducing the cluster size (N)
which increase the capacity.
y Disadvantages
y Cost: Increase number of antennas at each base station.
y Next section: Decrease trunking efficiency due to channel
sectoring at the base station.
y The available channels in the cell must be subdivided and dedicated to a
specific antenna.
10
Location of the BS
y Center vs. Corner
11
12 [http://motherboard.vice.com/blog/here-are-what-cell-phone-signals-would-look-like-if-you-could-see-them]
Sector
13 radiation patterns from two hypothetical base-station sites on the Capitol. The strongest signals are at
the center of the radiation patterns.
Cell sites on top of buildings provide much wireless coverage for New York City.
14
The hexagon pattern is a theoretical grid for antenna placement.
How cellular signals might appear in the Hollywood Hills if we could see the electromagnetic radiation at these
frequencies as we can in the visible spectrum. The long-distance tower is radiating three channel combinations in three
directions
15 indicated by the red, yellow and blue radiation patterns. In the background are cell stations each of which have
3 sector antenna radiation patterns as well.
A regular, hexagonal grid of cellular base-station sites is conceptualized for Chicago, with stations at the
corners of the hexagons. Near the downtown area more users are likely to be found and the hexagonal
cells
16 are smaller to serve approximately the same numbers of users found in larger cells elsewhere. Antenna
Chicago
signals extending beyond the original cells provide coverage over part of Lake Michigan.
Coverage provided by a base station located at the Herbert C. Hoover Building in
Washington
17 D.C. Hexagonal cells and their related coverage are also shown in the
background.
18
ECS455 Chapter 2
Cellular Systems
2.4 Traffic Handling Capacity
and Erlang B Formula
Dr.Prapun Suksompong
1 prapun.com/ecs455
2
Capacity Concept: A Revisit
y If we can “give” the time that “User 1” is idle to other users,
y then one channel can support users!!
y True?
4
New Concepts for a New Look at
Capacity
y We can let more than one user share a channel by using it at
different times.
y Blocked call happens if a user requests to make a call when all
the channels are occupied by other users.
y Probability of (call) blocking: Pb
y The likelihood that a call is blocked because there is no available
channel.
y 1%, 2%, 5%
y In which case, the number of users that a cell can support can
exceed ௌ ே.
y How much larger depends strongly on the value of Pb that can be
tolerated.
Trunking
y Allow a large number (n) of users to share the relatively
small number (m) of channels in a cell (or a sector) by
providing access to each user, on demand, from a pool of
available channels.
Omnidirectional 120q Sectoring 60q Sectoring
#sectors/cell 1 3 6
#channels/sector ௌ ௌ
݉ ൌ ܵȀܰ ݉ൌ Ȁ͵ ݉ൌ Ȁ
ே ே
8
Erlang B Formula
m = Number of trunked channels
Am
Pb m
m
Ai
¦
i i
Call blocking A WUDIILFLQWHQVLW\RUORDG >(UODQJV@
probability
O
O = Average # call
P attempts/requests per unit
time
In MATLAB, use
erlangb(m,A) H $YHUDJHFDOOOHQJWK
P
We use the MATLAB code from http://infohost.nmt.edu/~borchers/erlang.html to evaluate the Erlang B formula.
9
Exercise
y Use the Erlang B formula to find Pb for the values of m and A
specified in the table below.
Am m A 1 2 f
Pb m 1
m
Ai
¦
i i
2
10
M/M/m/m Assumption
y Blocked calls cleared
y No queuing for call requests.
y For every user who requests service, there is no setup time and the user is given
immediate access to a channel if one is available.
y If no channels are available, the requesting user is blocked without access and is
free to try again later.
y Calls arrive as determined by a Poisson process.
y There are memoryless arrivals of requests, implying that all users, including
blocked users, may request a channel at any time.
y There are an infinite number of users (with finite overall request rate).
y The finite user results always predict a smaller likelihood of blocking. So,
assuming infinite number of users provides a conservative estimate.
y The duration of the time that a user occupies a channel is
exponentially distributed, so that longer calls are less likely to occur.
y There are m channels available in the trunking pool.
y For us, m = the number of channels for a cell (C) or for a sector
11
Am
m
Erlang B Formula and Chart Pb
¦
m
Ai
Number of Trunked Channels (m) i i
12
Example 1a
Number of Trunked Channels (m)
5
Example 2.1
y Consider a cellular system in which
y an average call lasts two minutes
y the probability of blocking is to be no more than 1%.
y If there are a total of 399 traffic channels for a seven-cell
reuse system, there will be 57 traffic channels per cell.
y From the Erlang B formula, can handle 44.2 Erlangs or 1326
calls per hour.
Example 2.2
y Now employing 120° sectoring,
there are only m = 57/3 = 19 channels per sector.
y For the same probability of blocking and average call length,
each sector can handle 11.2 Erlangs or 336 calls per hour.
y Since each cell consists of three sectors, this provides a cell
capacity of 3 × 336 = 1008 calls per hour, which amounts
to a 24% decrease when compared to the unsectored case.
y Thus, sectoring decreases the trunking efficiency while
improving the SIR for each user in the system.
1% 0.1%
m
19
[Rappaport, 2002, Table 3.4]
Summary of Chapter 2: Big Picture
S = total # available duplex radio channels for the system
Frequency reuse with cluster size N Path loss exponent
J
AWRWDO S S kR J §D· J
“Capacity” C u | ¨ ¸ N
AFHOO N Tradeoff I K u kD J K©R¹ K
Omni-directional: K = 6
m = # channels allocated to 120q Sectoring: K = 2
each cell. 60q Sectoring: K = 1
Trunking m = # trunked channels
O = Average # call attempts/requests per unit time
Am
O
Call blocking
Pb m A WUDIILFLQWHQVLW\RUORDG>(UODQJV@
probability m
Ai P
20
¦
i i
H $YHUDJHFDOOOHQJWK
Erlang-B formula P
y SIR t 15 dB
SIR = 10*log10(1/K*((sqrt(3*N)).^y));
plot(N,SIR)
J end
6,5 | N
K
K=1oN=3
K=2oN=3
K=6oN=7
SIR (dB)
N
22
Make sure that you understand where numbers in this table come from!
Example 3 (3)
Omnidirectional 120q Sectoring 60q Sectoring
K 6 2 1
N 7 3 3
SIR [dB] 18.7 16.1 19.1
#channels/cell ͶͲͲȀ = 57 ͶͲͲȀ͵ = 133 ͶͲͲȀ͵ = 133
#sectors/cell 1 3 6
m = #channels/sector 57 ସ ସ
Ȁ͵ = 44 Ȁ = 22
ଷ ଷ
A [Erlangs]/sector 51.55 38.56 17.13
A [Erlangs]/cell 51.55 38.56u3 = 115.68 17.13u6 = 102.78
#users/cell 18558 41645 37001
Assume that each user makes 2 calls/day and 2 min/call on average o 1/360 Erlangs.
For the same N, we see that 120q sectoring and 60q sectoring give much
better SIR. However, sectoring reduces the trunking efficiency and
therefore suffer reduced value of A.
24
Idea: The values of SIR are too high for the cases of 120q sectoring and 60q sectoring. We can
further reduce the cluster size. This increases the number of channels per cell and hence per sector.
Omnidirectional 120q Sectoring 60q Sectoring
K 6 2 1
N 7 3 3
SIR [dB] 18.7 16.1 19.1
#channels/cell ͶͲͲȀ = 57 ͶͲͲȀ͵ = 133 ͶͲͲȀ͵ = 133
#sectors/cell 1 3 6
m = #channels/sector 57 ସ ସ
Ȁ͵ = 44 Ȁ = 22
ଷ ଷ
25
Sirindhorn International Institute of Technology
Thammasat University
School of Information, Computer and Communication Technology
In this note, we look at derivations of the Erlang B formula and the Engsett formula,
both of which estimate the call blocking probability when trunking is used. To do this,
we need to borrow some concepts from queueing theory. Moreover, some basic analysis of
stochastic processes including Poisson processes and Markov chains is needed. For com-
pleteness, working knowledge on these processes is summarized here as well. However, we
do assume that the readers are familiar with concepts from basic probability course.
Contents
1 Poisson Processes 4
1.1 Discrete-time (small-slot) approximation of a Poisson process . . . . . . . . 7
1.2 Properties of Poisson Processes . . . . . . . . . . . . . . . . . . . . . . . . 9
3 Markov Chains 20
4 Engsett Model 26
1
In performance evaluation of cellular systems or telephone networks, Er-
lang B formula [1, p 23], to be revisited in Section 2, is a formula for
estimating the call blocking probability for a cell (or a sector, if sectoring is
used) which has m “trunked” channels and the amount of (“offered”) traffic
is A Erlang: m
A
m!
Pb = m . (1)
Ai
P
i!
i=0
It is directly used to determine the probability Pb that call requests will be
blocked by the system because all channels are currently used. The amount
of traffic (A) can be found by the product of the total call request rate λ
and the average call duration (1/µ).
When we design a cellular system, the blocking probability Pb should be
less than some pre-determined value. In which case, the function above can
be used to suggest the minimum number of channels per cell (or sector). If
we already know the number of channels per cell (or sector) of the system,
the (inverse of this) function can also be used to determined how many users
the system can support.
The Erlang B formula (1) is derived under the following M/M/m/m
assumptions:
(a) Blocked calls cleared
• No queue for call requests.
• For every user who requests service, there is no setup time and the
user is given immediate access to a channel if one is available.
• If no channels are available, the requesting call is blocked without
access and has no further influence on the system.
(b) Call generation/arrival is determined by a Poisson process.
• Arrivals of requests are memoryless.
(c) There are an infinite number of users (with finite overall request rate).
• The finite user results always predict a smaller likelihood of block-
ing. So, assuming infinite number of users provides a conservative
estimate.
2
(d) The duration of the time that a user occupies a channel is exponen-
tially distributed.
(e) The call duration times are independent and they are also independent
from the call request process.
(f) There are m channels available in the trunking pool.
• For us, m = the number of channels for a cell or for a sector.
3
and some probability concepts will be reviewed in Section 1. Most of the
probability review will be put in footnotes so that they do not interfere with
the flow of the presentation.
1 Poisson Processes
In this section, we consider an important random process called Poisson
process (PP). This process is a popular model for customer arrivals or calls
requested to telephone systems.
1.1. We start by picturing a Poisson Process as a random arrangement of
“marks” (denoted by “x” or ) on the time axis. These marks usually indi-
cate the time that customers arrive in queueing models. In the language of
“queueing theory”, the marks denote arrival times. For us, they indicate
the time that call requests are made:
Time
1.2. In this class, we will focus on one kind of Poisson process called ho-
mogeneous1 Poisson process. Therefore, from now on, when we say
“Poisson process”, what we mean is “homogeneous Poisson process”.
1.3. The first property that you should remember is that
12
1
This is a special case of Poisson processes. More general Poisson process (which are called non-
homogenous Poisson processes) would allow the rate to be time-dependent. Homogeneous Poisson processes
have constant rates.
2
For homogeneous Poisson process, λ is a constant. For non-homogeneous Poisson process, λ is a
function of time, say λ(t). Our λ is constant because we focus on homogeneous Poisson process
4
Therefore, if we consider many non-overlapping intervals on the time axis,
say interval 1, interval 2, and interval 3 below,
Time
0, 1, 2, 3, . . . .
Because we don’t know their exact values, we describe them via the likelihood or probability that they
will take one of these possible values. For example, for N1 , we describe it by
where P [N1 = k] is the probability that N1 takes the value k. Such list of numbers is a bit tedious. So,
we define a function
pN1 (k) = P [N1 = k] .
This function pN1 (·) tells the probability that N1 will take a particular value (k). We call pN1 the probability
mass function (pmf) of N1 . At this point, we don’t know much about pN1 (k) except that its values will be
between 0 and 1 and that
X∞
pN1 (k) = 1.
k=0
These two properties are the necessary and sufficient conditions for any pmf.
4
By saying that something are independent, we mean it in terms of probability. In particular, when we
say that N1 and N2 are independent, it means that
P [N1 = k and N2 = m]
5
1.5. Do we know anything else about N1 , N2 , and N3 ? Again, we have only
one parameter λ for a Poisson process. So, can we relate λ with N1 , N2 , and
N3 ?
Recall that λ is the average number of arrivals per unit time. So, if λ = 5
arrivals/hour, then we expect that N1 , N2 , and N3 should conform with this
λ, statistically.
Let’s first be more specific about the time durations of the intervals that
we have earlier. Suppose their lengths are T1 , T2 , and T3 respectively.
T1 T2 T3
N1 = 1 N2 = 2 N3 = 1
Time
EN = λT. (2)
5
Recall that EN1 is the expectation (average) of the random variable N1 . Here, the random variables
are discrete. Therefore, formula-wise, we can calculate EN1 from
∞
X
EN1 = k × P [N1 = k] ;
k=0
that is the sum of the possible values of N1 weighted by the corresponding probabilities
6
1.1 Discrete-time (small-slot) approximation of a Poisson process
1.6. The next key idea is to consider a small interval:
Imagine dividing a time interval of length T into n equal slots.
Time
Then each slot would be a time interval of duration δ = T /n. For example,
if T = 20 hours and n = 10, 000, then each slot would have length
12
T 20
δ= = = 0.002 hour.
n 10, 000
Why do we consider small interval? The key idea is that as the interval
becomes very small, then it is extremely unlikely that there will be more
than 1 arrivals during this small amount of time. This statement becomes
more accurate as we increase the value of n which decreases the length of
each interval even further. What we are doing here is an approximation6 of
a continuous-time process by a discrete-time process.7
To summarize, we will consider the discrete-time approximation of the
(continuous-time) Poisson process. In such approximation, the time axis is
divided into many small time intervals (which we call “slots”).
When the interval is small enough, we can assume that at most 1
arrival occurs.
6
You also do this when you plot a graph of any function f (x). You divide the x−axis by many (equally
spaced) values of x and then evaluate the values of the function at these values of x. You need to make
sure that the values of x used are “dense” enough such that no surprising change in the function f is
overlooked.
7
If we want to be rigorous, we would have to bound the error from such approximation and show that
the error disappear as n → ∞. We will not do that here.
7
1.7. Let’s look at the small slots more closely. Here, we let N1 be the
number of arrivals in slot 1, N2 be the number of arrivals in slot 2, N3 be
the number of arrivals in slot 3, and so on as shown below.
Time
Then, these Ni ’s are all Bernoulli random variables because they can only
take the values 0 or 1. In which case, for their pmfs, we only need to specify
one value
12 P [Ni = 1]. Of course, knowing this, we can calculate P [Ni = 0]
by P [Ni = 0] = 1 − P [Ni = 1].
Recall that the average EX of any Bernoulli random variable X is simply
P [X = 1].8 So, if we know EX for Bernoulli random variable, then we know
right away that P [X = 1] = EX and P [X = 0] = 1 − EX.
Now, it’s time to use what we learned about Poisson process. The slots
that we consider before are of length δ = T /n. So, the random variables
N1 , N2 , N3 , . . . share the same expected value
EN1 = EN2 = EN3 = · · · = λδ.
For example, with λ = 5, T = 20, and n = 10, 000, the expected number of
arrivals in a slot is λδ = λ Tn = 0.01 arrivals.
Because these Ni ’s are all Bernoulli random variables and because they
share the same expected value, we can conclude that they are identically
distributed; that is their pmf’s are all the same. Furthermore, because the
slots do not overlap, we also know that the Ni ’s are independent. Therefore,
for small non-overlapping slots, the corresponding number of ar-
rivals Ni ’s are i.i.d. Bernoulli random variables whose pmf’s are
given by
p1 = P [Ni = 1] = λδ and p0 = P [Ni = 0] = 1 − λδ,
where δ is the length of each slot.
8
For Bernoulli random variable X, the average is
EX = 0 × P [X = 0] + 1 × P [X = 1] = P [X = 1] .
For conciseness, we usually let p0 = P [X = 0] and p1 = P [X = 1]. Hence, EX = p1 .
8
1.8. Discrete-time approximation provides a method for MATLAB to generate
a Poisson process with arrival rate λ. Here are the steps:
(a) First, we fix the length T of the whole simulation. (For example, T = 20
hours.)
(b) Then, we divide T into n slots. (For example, n = 10, 000.)
(c) For each slot, only two cases can happen: 1 arrival or no arrival. So,
we generate Bernoulli random variable for each slot with p1 = λ × T /n.
(For example, if λ = 5 arrival/hr, then p1 = 0.01.)
To do this for n slots, we can use the command rand(1,n) < p1 or
binornd(1,p1,1,n).
1.9. Note that what we have just generated is exactly Bernoulli trials
whose success probability for each trial is p1 = λδ. In other words, a Poisson
process can be approximated by Bernoulli trials with success probability
p1 = λδ.
Time
Let N be the number of arrivals during this time interval. In the picture
above, N = 4.
16
Again, we will start with a discrete-time approximation; we divide T
into n small slots of length δ = Tn . In the previous subsection, we know that
the number of arrivals in these intervals, denoted by N1 , N2 , . . . , Nn can be
well-approximated by i.i.d. Bernoulli with probability of having exactly one
arrival = λδ. (Of course, we need δ to be small for the approximation to be
9
precise.) The total number of arrivals during the original interval of length
T can be found by summing the values of the Ni ’s:
(3)
10
fact, in the limit as n → ∞ (and hence δ → 0), the random variable N is
governed by P(λT ). Recall that the expected value of P(α) is α. Therefore,
λT is the expected value of N . This agrees with what we have discussed
before in (2).
In conclusion,
the number N of arrivals in an interval of length T is a Poisson
random variable with mean (parameter) λT
1.11. Now, to sum up what we have learned so far, the following is one of
the two main properties of a Poisson process
The number of arrivals N1 , N2 , N3 , . . . during non-overlapping time
intervals are independent Poisson random variables with the mean
being λ× the length of the corresponding interval.
1.12. Another main property of the Poisson process, which provides an
alternative method for simulating Poisson process, is that
The lengths of time between adjacent arrivals W1 , W2 , W3 , . . . are
i.i.d. exponential12 random variables with mean 1/λ.
Time
W1 W2 W3 W4
1 − e−λx , x > 0,
FX (x) ≡ P [X ≤ x] =
0, x ≤ 0.
Often, we also talked about the complementary cumulative distribution function (CCDF) which, for expo-
nential random variable, is given by
−λx
e , x > 0,
P [X > x] ≡ 1 − FX (x) =
1, x ≤ 0.
11
Time
Time
N1 = 1 N2 = 2 N3 = 1
Time
W1 W2 W3 W4
The lengths of time between adjacent arrivals W1, W2, W3 ,… are i.i.d.
exponential random variables with mean 1/.
13
You may recall from your introductory probability class that, in Bernoulli trials, the number of trials
until the next success is geometric.
12
Small Slot Analysis (Poisson Process)
(discrete time approximation)
T1 T2 T3
N1 = 1 N2 = 2 N3 = 1
Time
W1 W2 W3 W4
In the limit, there is at most one arrival in any slot. The numbers of arrivals on the slots are
i.i.d. Bernoulli random variables with probability p1 (= ) of exactly one arrivals where is
the width of individual slot.
Time
D1 The total number of arrivals on n slots is a
The number of slots between adjacent binomial random variable with parameter
arrivals is a geometric random variable. (n,p1)
In the limit, as the slot length gets smaller, geometric exponential
binomial Poisson
11
Figure 3: Small slot analysis (discrete-time approximation) of a Poisson process
13
2.2. Let’s first divide the time into small slots (of the same length δ) as we
have done in the previous section.
Time
Then, consider any particular slot. Suppose that at the beginning of this
time slot, there are K channels that are currently used.14 We want to find
12
out how this number K changes as we move forward one slot time. This
random variable K will be called the state of the system15 . The system
moves from one state to another one as we advance one time slot.
Example 2.3. Suppose there are 5 persons using the channels at the be-
ginning of a slot. Then, K = 5.
(a) Suppose that, by the end of this slot, none of these 5 persons finish
their calls.
(b) Suppose also that there is one new person who wants to make a call at
some moment of time during this slot.
Then, at the end of this time slot, the number of channels that are occupied
becomes
So, the state K of the system changes from 5 to 6 when we reach the end
of the slot, which can now be regarded as the beginning of the next slot.
2.4. Our current intermediate goal is to study how the state K evolves
from one slot to the next slot. Note that it might be helpful to label the
state K as K1 (or K[1]) for the first slot, K2 (or K[2]) for the second slot,
and so on.
As shown in Example 2.3, to determine how the Ki ’s progress from K1
to K2 , K2 to K3 , and so on, we need to know two pieces of information:
14
The value of K can be any integer from 0 to m.
15
This is the same “state” concept that you have studied in digital circuits class.
14
Q1 How many calls (that are ongoing at the beginning of the slot under
consideration) end during the slot that we are considering?
Q2 How many new call requests are made during the slot under consider-
ation?
Note that Q1 depends on the characteristics of the call duration and Q2
depends on the characteristics of the call request/arrival process. After we
know the answers to these two questions, then we can find Ki via
p1 = λδ.
2.6. For Q1, we need to consider the call duration model . The M/M/m/m
assumption states that the call duration16 is exponentially distributed with
parameter (rate) µ. Let’s consider the call duration D of a particular call.
Recall that the probability density function (pdf) of an exponential ran-
dom variable X with parameter µ is given by
−µx
µe , x > 0,
fX (x) =
0, x ≤ 0,
16
In queueing theory, this is sometimes called the service time. The parameter µ then captures the
service rate.
15
An important property of an exponential random variable X is its mem-
oryless property17 :
For example,
P [X > 7|X > 5] = P [X > 2] .
What does this memoryless property mean? Suppose you have a light-
bulb and you have used it for 5 years already. Let it’s lifetime be X.
Then, of course, X is a random variable. You know that X > 5 be-
cause it is still working now. The conditional probability P [X > 7|X > 5]
is the probability that it will still work after two more years of use (given
the fact that currently it has been working for five years). Now, if X is
an exponential random variable, then the memoryless property says that
P [X > 7|X > 5] = P [X > 2]. In other words, the probability that you can
use it for at least two more years is exactly the same as the probability that
you can use a new lightbulb for more than two years. So, your old lightbulb
essentially forgets that it has been used for 5 years, It always performs like
a new lightbulb (statistically). This is what we mean by the memoryless
property of an exponential random variable.
2.7. To answer Q1, we now return to our small slot approximation. Again,
consider one particular slot. At the beginning of our slot, there are K = k
ongoing calls. The probability that a particular call, which is still ongoing
at the beginning of this slot, will be unfinished by the end of this slot is
e−µδ .
17 P (A∩B)
To see this, first recall the definition of conditional probability:P (A |B ) = P (B) . Therefore,
e−µ(x+δ)
P [X > x + δ |X > x ] = = e−µδ = P [X > δ] .
e−µx
16
To see this, consider a particular call. Suppose the duration of this call
is D. By assumption, we know that D is exponential with parameter µ.
Let s be the length of time from the call initiation time to the beginning
of our slot. Note that the call is still ongoing. Therefore D > s. Now, to
say that this call will be unfinished by the end of our slot is equivalent to
requiring that D > s + δ. By the memoryless property, we have
P [D > s + δ |D > s] = P [D > δ] = e−µδ .
Recall that we have K = k ongoing calls at the beginning of our slot. So,
by the end of our slot, the probability that none of them finishes is
k
e−µδ = e−kµδ .
The probability that exactly one of them finishes is
(b) the probability that exactly one of them ends during our slot is
17
Magically, these two probabilities sum to one. So, we don’t have to consider
other events/cases.
2.8. Summary:
(a) Call generation/request/initiation process:
k-1 k k+1
1 1 k k 1 k
P[0 new call request] ≈ 1 -
The labels on the arrows are 18 call request] ≈
P[1 new
transition probabilities. P[0 old-call end] ≈ 1 - k
2 P[1 old-call end] ≈ k
Note that the labels on the arrows indicate transition probabilities which
are conditional probabilities of going from some value of K to another value.
2.9. When there are m trunked channels, the possible values of K are
0, 1, 2, . . . , m. We can combine the diagram above into one diagram that
includes all possible values of K:
Note that the arrow λδ which should go out of state m will return to state
m itself because it corresponds to blocked calls which do not increase the
value of K.
The resulting diagram is called the Markov chain state diagram for
Erlang B.
Small
Example 2.10. slot
Erlang Analysis:
B (Markov chain)Markov Chain
state diagram when m = 2:
Case: m = 2
1 0 1 2 1 2
2
1
19
3
What we need from our model is to find out the behavior of the system
in the long run. The complete discussion of the Markov chain itself deserves
a whole textbook. (See e.g. [2].) In the next section, we only introduce the
concept of Markov chain and some necessary properties.
3 Markov Chains
3.1. Markov chains are simple mathematical models for random phenomena
evolving in time. Their simple structure makes it possible to say a great
deal about their behavior. At the same time, the class of Markov chains
is rich enough to serve in many applications. This makes Markov chains
the most important examples of random processes. Indeed, the whole of the
mathematical study of random processes can be regarded as a generalization
in one way or another of the theory of Markov chains. [2]
3.2. The characteristic property of Markov chain is that it retains no mem-
ory of where it has been in the past. This means that only the current state
of the process can influence where it goes next.
3.3. Markov chains are often best described by their (state) diagrams. You
have seen a Markov chain in Example 2.10.
Example 3.4. Draw the (state) diagram of a Markov chain which has three
states: A,B and C. It moves from state A to state B with probability 1.
From state C, it moves either to A or to B with equal probability 1/2, and
from B it jumps to C with probability 1/3, otherwise it stays at B.
3.5. The Markov chains that we have just seen in the Example 3.4 and in
the previous section (Example 2.10) are all discrete-time Markov chains.
However, the Poisson process and the state K(t) which we have studied ear-
lier are examples of a continuous-time Markov chain. Nonetheless, equipped
with our small-slot approximation (discrete-time approximation) technique,
20
we may analyze the Poisson process or any continuous-time Markov chain
as a limit of a discrete-time Markov chain as well.
3.6. We will now introduce the concept of stationary distribution, steady-
state distribution, equilibrium distribution, and limiting distribution. For
the purpose of this class, we will not distinguish these terms. We shall see
in the next example that for the Markov chains that we are considering, in
the long run, it will reach a steady state.
Example 3.7. Consider the Markov chain characterized by the state tran-
sition diagram below:
0 + 50,000 0 + 50,000
Time Slot # 1 2 3 4 5 6 7 Time Slot 2 50,000 50,000
2/5 3/5 1/2 1/2
B
B 20,000 + 25,000 30,000 +25,000
B Time Slot 3 45,000 55,000
B 2/5 3/5 1/2 1/2
Let’s try a thought experiment – imagine that you start with n = 100, 000
trials
1
of these Markov chain, all of which start in state B. So, during slot
1 (the first time slot), all trials will be in state B. For slot 2, about 50% of
these will move to state A; but the other 50% of the trials will stay at B.
By the time that you reach slot 6, you can observe that out of the 100,000
trials, about 45.5% will be in state A and about 55.5% will be in state B.
Mathematically, this evolution can be calculated simply by
h i
(1) (1) (1)
(a) Start with a row vector n = nA , nB containing the initial number
of trials in each state.
21
(b) For the kth time slot, the new row vector n(k) is the old row vector
n(k−1) multiplied by the transition matrix
Limiting distribution P=
2/5 3/5
.
1/2 1/2
The relative frequencies (of the two states) can be found from n1 n(k) .
Recall that when n is large, relative frequencies converge to probabilities.
Turn out that the relative frequencies [45.5%, 55.5%] stay roughly the
same
1 as you proceed to slot 7, 8, 9, and so on. Note also that it does not
matter how we start our 100,000 trials. We may start with 10,000 in state
A and 90,000 in state B. Eventually, [45.5%, 55.5%] will emerge.
In conclusion,
(a) If we observe the long-run behavior of this Markov chain at a particular
slot, then the probability that we will see it in state A is 0.455 and the
probability that you will see it in state B is 0.545.
(b) In addition, one can also show that if we observe the behavior of this
Markov chain for a long time, then the proportion of time that it stays
in state A is 45.5% and the proportion of time that it stays in state B
is 54.5%.
22
The distribution [0.455, 0.545] is what we referred to as stationary distri-
bution, steady-state distribution, equilibrium distribution, or limiting dis-
tribution above.
3.8. Another way to look at the convergence in Example 3.7 is to first look
at Pk . This is because n(k+1) = n(1) Pk .
Analytically, the convergence in Example 3.7 can be shown easily by realizing that the tran-
sition P can be decomposed into P = VDV−1 where D is a diagonal matrix of eigenvalues and
V is a matrix whose columns contain the corresponding eigenvectors. In MATLAB, the matrices V
and D can be found easily from the command [V,D] = eig(P). For Example 3.7,
2/5 3/5 −1/10 0 −6/5 1
P= , D= , and V = .
1/2 1/2 0 1 1 1
Therefore,
1
−1 k −1
k k k
V−1 .
lim P = lim VDV = lim VD V =V lim D
k→∞ k→∞ k→∞ k→∞
This gives
k 5/11 6/11
lim P = .
k→∞ 5/11 6/11
Therefore, it does not matter what we have for n(1) . As k → ∞,
(k+1) (1) k (1) 5/11 6/11 5 6
n =n P →n = 11 n 11 n
5/11 6/11
23
3.9. From Example 3.7, we can see that Discrete-time Markov chain can be
analyzed via its state transition diagram or its probability transition
matrix P.
Example 3.10. Consider an evolution of a Markov chain which has two
states (1 and 2):
2, 2, 1, 1, 1, 2, 2, 1, 2, 1
(a) Estimate its transition matrix P.
(b) Let pi denote the proportion of time that the system spends in state i.
Estimate pi .
24
To write down a balance equation, first define a boundary, then consider
the transfer of probabilities “in” and “out” of the boundary. To be at
equilibrium, there should not be any net transfer.
Example 3.14. Let’s reconsider the state diagram derived in Example 3.4.
Write the balance equation for each of the boundaries below.
3.15. The steady-state probabilities can be found from the balance equa-
tions together with the fact that the sum of them must be 1.
Example 3.16. For the the state diagram derived in Example 3.4. We can
use the balance equations expressed in Example 3.14 to find its steady-state
probabilities:
25
Example 3.17. Let’s reconsider Example 2.10 where m = 2.
1 0 1 2 1 2
2
Let pk be the long-term 1
probability that K = k.
p0 p1 p1 2 p2
p0 p1 p2 1
1 1 2
p0 2
, p1 Ap0 , p2 A p0
A 2
1 A
2
26
Here, pm is the same as call blocking probability Pb , which is the long-run
proportion of call requests that get blocked.
3.20. Convention: In this class, for any question that requires you to get
your answers (call blocking probability or steady-state probabilities) “from
the Markov chain” or “via the Markov chain”, make sure that you
(a) draw the Markov chain (with all the states and transition probabilities),
(b) set up the boundaries,
(c) write down the corresponding balance equations,
(d) use the equations to solve for the interested quantities.
3.21. Remark: Beyond this class, mathematicians do have more direct
ways to analyze continuous-time Markov chain without the discrete-time
approximation that we have been using here. However, the analysis requires
more background knowledge and hence we did not try to use it.
Here is a glimpse of what’s out there:
(a) Our version:
27
4 Engset Model
In this section, we will consider a more realistic system where the infinite-
user assumption in M/M/m/m is relaxed.
4.1. Modified (more realistic) assumptions:
(a) Finite number of users: n users/customers/subscribers.
(b) Each user independently generates new call request with rate λu .
• Total call request rate = n × λu = λ.
4.2. We need to modify our small slot analysis. Here, for each small slot,
each user do the following:
(a) If it is idle (not using the channel) at the beginning of the slot,
(i) it may request/generate a new call in a small slot with probability
λu δ.
i. If there is at least one available channel, then it may start its
conversation. (In which case, at the beginning of the next slot,
its call is ongoing.)
ii. If there is no channel available, then the call is blocked and it
is idle again (at the beginning of the next slot).
(ii) With probability 1 − λu δ, no new call is requested by this user
during this slot. (In which case, it is idle at the beginning of the
next slot.)
(b) If it is making a call at the beginning of the slot,
(i) with probability µδ, this call ends before the end of this slot. (In
which case, at the beginning of the next slot, it is idle.)
(ii) with probability 1 − µδ, this call is still ongoing at the end of this
slot (which is the same as the beginning of the next slot).
4.3. Observe that the call generation process for each user is not a Poisson
process with rate λu . This is because it get interleaved with the call duration
for each successful call request. Part of the Poisson assumption that is left
is that, in fact, for an idle user, the time until the next call request will be
exponential with rate λu .
28
4.4. A user can not generate new call when he/she is already involved in a
call. Therefore, if the system is in state K = k, there are only n − k users
that can generate new calls. Hence, the “total” call request rate for state
K = k is (n − k)λu .
• Earlier, when we consider the Erlang B formula, we always have λ as
the total new call request rate regardless of how many users are using
the channels. This is because we assumed infinite number of users and
hence having k users using the channels will not significantly change
the total call request rate.
4.5. Comparison of the state transition probabilities:
We can then write down the balance equations and solve for all the steady-
state probabilities.
4.7. It is tempting to conclude that the call blocking probability is pm .
However, this is not the case for us here. Recall that pm is the long-run
probability (and the long-run proportion of time) that the system will be
in state K = m. In this state, any new call request will be blocked. So, pm
gives the blocking probability in terms of the time (time congestion).
29
However, if you look back at how we define Pb which is the call blocking
probability, this is the probability that a call is blocked. So, what we need
to find out is, out of all the new calls that are requested, how many will be
blocked.
To do this, consider s slots. Here the value of s is very large. Then,. . .
(a) About pk × s slots will be in state k.
(b) Each of these slots will generate new call request with probability
(n − k)λu δ.
(c) So, the number of new calls request from slots that are in state k will
be approximately
(n − k)λu δ × (pk × s).
(e) However, the number of the new call requests that get blocked is
4.8. Remarks:
(a) If we keep the total rate λ constant and let n → ∞, then the call
blocking probability in the Engset model will be the same as the call
blocking probability in the Erlang model.
(b) If n ≤ m, the call blocking probability in Engset model will be 0.
30
References
[1] Andrea Goldsmith. Wireless Communications. Cambridge University
Press, 2005. (document)
[2] J. R. Norris. Markov Chains. Cambridge University Press, 1998. 2, 3.1
[3] Theodore S. Rappaport. Wireless Communications: Principles and
Practice. Prentice Hall PTR, 2 edition, 2002.
31
ECS455 Chapter 3
Call Blocking Probability
Office Hours:
BKD, 6th floor of Sirindhralai building
Tuesday 14:20-15:20
Dr.Prapun Suksompong Wednesday 14:20-15:20
1 prapun.com/ecs455 Friday 9:15-10:15
Introduction
y The English dictionary word with the most consecutive
vowels (six) is EUOUAE.
y It is also the longest English word consisting only of vowels
2
Introduction
y Words with five consecutive vowels include AIEEE,
COOEEING, MIAOUED, ZAOUIA, JUSSIEUEAN,
ZOOEAE, ZOAEAE.
Introduction
y Words with five consecutive vowels include AIEEE,
COOEEING, MIAOUED, ZAOUIA, JUSSIEUEAN,
ZOOEAE, ZOAEAE.
y Note: The longest common word without any of the five vowels is RHYTHMS.
y There are longer rare words: SYMPHYSY, NYMPHLY, GYPSYRY, GYPSYFY, and TWYNDYLLYNGS. WPPWRMWSTE and
GLYCYRRHIZIN are long words with very few vowels.
4
That Second “e”…
y You may recall the rule for changing a verb into its “–ing”
form from your English class…
y If the verb ends in an “e” we remove the “e” and add “-ing”:
y browsing, causing, changing, charging, choosing, giving, having,
hiring
6
Queueing vs. Queuing
ECS455 Chapter 3
Call Blocking Probability
3.2 Markov Chain
Dr.Prapun Suksompong
8 prapun.com/ecs455
Review: Discrete-Time Markov Chain
y We model the evolution in time of K by Markov chain.
y K(t) = the number of channels being occupied at time t
y Time is divided into small slots so that our analysis can be
done in discrete time.
y This only approximate the solution. However, the answers will
be accurate in the limit that the slot size G approaches 0.
y Discrete-time Markov chain can be specified via its state
transition diagram or its probability transition matrix P.
10
Simulating a Markov Chain in MATLAB
n = 1e1; % The number of slots to be considered
S = [1,2]; % Two possible states
P = [2/5 3/5; 1/2 1/2]; % Transition probability matrix
X1 = 2; % Initial state
X = MarkovChainGS(n,S,P,X1)
11
[MarkovChain_Demo1.m]
Example 3.10
>> MarkovChain_Demo1
X =
2 2 1 1 1 2 2 1 2 1
P_sim =
0.5000 0.5000
0.6000 0.4000
p_sim =
0.5000 0.5000
12
Exercises
>> MarkovChain_Demo1
X =
2 1 2 1 2 1 2 1 1 1
P_sim =
0.4000 0.6000
1.0000 0
p_sim =
0.6000 0.4000
>> MarkovChain_Demo1
X =
2 2 2 2 1 2 2 2 1 1
P_sim =
0.5000 0.5000
0.2857 0.7143
p_sim =
0.3000 0.7000
13
Example 3.7
n = 1e4; % The number of slots to be considered
S = [1,2]; % Two possible states
P = [2/5 3/5; 1/2 1/2]; % Transition probability matrix
X1 = 2; % Initial state
X = MarkovChainGS(n,S,P,X1);
14
[MarkovChain_Demo2.m]
Example 3.4
n = 1e4; % The number of slots to be considered
S = [1,2,3]; % Three possible states
P = [0 1 0; 0 2/3 1/3; 1/2 1/2 0]; % Transition probability matrix
X1 = 2; % Initial state
X = MarkovChainGS(n,S,P,X1);
16