how are resources shared? - ku ittc are resources shared? • review general access network topology...
TRANSCRIPT
#8 1
Victor S. FrostDan F. Servey Distinguished Professor
Electrical Engineering and Computer ScienceUniversity of Kansas2335 Irving Hill Dr.
Lawrence, Kansas 66045Phone: (785) 864-4833 FAX:(785) 864-7789
e-mail: [email protected]://www.ittc.ku.edu/
How are resources shared?
#8
All material copyright 2006Victor S. Frost, All Rights Reserved
#8 2
How are resources shared?• Review general access network topology• Resource sharing principles• Resource reservation (call) model
– Dedicated resources– Shared after reservation
• Always-on model– Polling– Random Access
• Asymmetric mechanisms– Assumptions– General descriptions– Scheduling in the downstream– Contention in the upstream
• Scheduling
#8 3
General access network topology
• Sharing the upstream resources requires a “distributed” mechanism– Mulitpoint-to-point– Subject to collisions
• Sharing the downstream resources requires a scheduling mechanism– Point-to-multipoint
InternetAccessMedium
Downstream
Upstream
InternetAccessMedium
Downstream
Upstream
#8 4
Resource sharing principles• Complex
– Demand for resources • Diverse
– Video– Voice– Short messages
• Dynamic, changes with time– Requirements for services
• Real-time• Near-real time• Non-real-time• Loss tolerant
– Desire for efficient use of resources– Basic tradeoff between:
• Providing “service”• Efficient use of resources
#8 5
Resource sharing principles
• A large number of demands for resources will present an aggregate demand equal to the sum of the average of the individual demands
• Which is better, sharing– One large capacity link or– Several smaller capacity links?
#8 6
Resource sharing principles
λ
λ
λ
µ
µ
µ
mseparatesystems
versus One
consolidatedsystem
mλ
mµλ
λ
λ
µ
µ
µ
mseparatesystems
versus One
consolidatedsystem
mλ
mµ
λ = Packet arrival rate (packets/sec)µ= Packet service rate (packets/sec)ρ = load per server= λµρΤ = Total load = mρ
Buffers Servers
• One consolidated system is better• Therefore it is better to have a large number of users sharing a single server
#8 7
Impact on resource sharing • Goal:
– Access the “entire” channel bandwidth– Through a “global buffer”
• Want– Global knowledge of who wants to send– Allow each to send according to some schedule,
e.g. FIFO• However,
– Users are geographically distributed– There is no perfect knowledge of system state– The messages to coordinate the transmissions of
the users must also use the same media
#8 8
Impact on resource sharing
M/M/1 Delay
05
101520253035
0 0.2 0.4 0.6 0.8 1
Load
Del
ay
Ideal MACPerformance
#8 9
Impact on resource sharing
Tran
sfer
Del
ay
Load
E[T]/E[X]
ρρmax-2 1
1
ρmax-1
MAC Protocol 1
MAC Protocol 2
Adapted from: Leon-Garcia & Widjaja: Communication Networks
#8 10
Classification of MAC Schemes
MAC
Always-on Fixed Allocation
PollingContention
Hybrid
TDMA
CDMA
FDMA
SDMA
#8 11
Classification of MAC Schemes
MAC
Dynamically Scheduled Fully Scheduled
PollingRandom Access
Hybrids
TDMA
CDMA
FDMA
Another perspective
SDMA
#8 12
Fully scheduled (call) modelCircuit switching
• In fixed allocation– User requests connection via
signaling – Connection is established– Speakers converse– User(s) hang up– Network releases connection
resources
Signal
Source
Signal
Release
Signal
Destination
Goahead Message
Adapted from: Leon-Garcia & Widjaja: Communication Networks
#8 13
Resource reservation (call) model
• Circuit switching is a form of fully scheduled
• After call set up resources are dedicated for duration of “call”
• Signaling messages and user information may use different channels
• Signaling facilitates mobility• Enables billing per resources (min) used• Time wasted doing signaling
#8 14
Circuit Switching
Example: Find the time to transmit a 37.5 Mbyte message coast-to-coast is the USA (3000Km) on a 600 Mb/s link
Using Circuit Switching530ms
Key issue is holding time relative to call set-up time
A B
Call Set-up
Data Transmission
#8 15
Example
Air interface
AC = authentication centerBSS = base station subsystemEIR = equipment identity register
HLR = home location register
MSC
PSTN
BSS
STP SS7HLR
VLR
EIRAC
MSC = mobile switching centerPSTN = public switched telephone network
STP = signal transfer pointVLR = visitor location register
Physical Connection
Signaling PathInformation Path
Adapted from: Leon-Garcia & Widjaja: Communication Networks
Simplified Cellular System
#8 16
• In fully scheduled there are N resources available– Channels– Time slots– Codes
• Typically there for M users with access to N resources where M>>N
• Performance is measured in terms of probability of requesting a resource when all are busy
Performance of circuit switching
#8 17
∑=
=== N
n
n
N
B
n
NNkPP
0 !
!][ρ
ρ
Performance of circuit switching
Erlang BBlocking Formula λ = Call arrival rate (call/sec)µ= Call service rate (call/sec)ρ= load (Erlangs)
PB=Blocking Probability=
#8 18
Example simulation
3V 1 2
D
T U
D
T U
D
T U Exit#
Exit#
Exit#
D
T U1 2 3
Rand
Count
#r
Exit#
Number Calls Generated
Exit#
Number Calls Generated
AverageHolding Time =
Arrival Rate (call/min)=
C
C
Telephone Trunks
Blocked Calls C
#8 19
0 166.6667 333.3333 500 666.6667 833.3333 10000
1.822917
3.645833
5.46875
7.291667
9.114583
10.9375
12.76042
14.58333
16.40625
18.22917
20.05208
21.875
23.69792
25.52083
27.34375
29.16667
Time
% Blocked CallsPlotter, Discrete Event
Solid Blue GrayPat Red GrayPat Green ltGrayPat Black
Holding time=3min, Arrival rate=0.833 calls/min ->PB = 0.15, Simulated PB = 0.198
Example simulation
#8 20
Performance Evaluation: Example
• A department has 140 phones, each phone is busy 10% of the time during the busy hour.
• How many lines do you need to buy from the phone company to keep the blocking probability less than 2%.
#8 21
Performance Evaluation: Example
• Traffic intensity = 14 Erlangs
• From Erlang Table –14 Erlangs & 2% –Blocking ==> 21 lines
#8 22
Performance Evaluation: Example
• Design of a building phone system. The design goal is to minimize the number of lines needed between the building and the phone company. The blocking QoS is specified to be 5%.
• A building has four floors, on each floor is a separate department. Each department has 22 phones, each busy 10% of the time during the busy hour.
#8 23
Performance Evaluation: Example-Case A
• Acquire one telephone switch for each floor.
• 2.2 Erlangs/floor & B=5% gives:
• 5 lines/floor or 20 lines for the building.
#8 24
Performance Evaluation: Example-Case B
• Acquire one telephone switch for the building.
• 88 phones @ .1 Erlang/phone = 8.8 Erlangs
• 8.8 Erlangs & B=5% gives:• 13 lines for the building• Select Case B
– More traffic sharing more resources
#8 25
Virtual Circuit Packet Switching
• Use signaling process to set up a call• Resources are not necessarily
reserved for the flow • A “logical connection” is established
between the source and destination• All packets flow over the same route
through the network• Packets still “statistically share” link
#8 26
Virtual Circuit Packet Switching
• Forwarding decisions are made based on a “virtual circuit identifier” not on the full address
• Packet share transmission facilities• Switches save state/connection• State is saved for duration of the
connection• QoS can be guaranteed• Facilities billing
#8 27
Virtual Circuit Packet Switching
*Note: Do not need the same VCI end-to-end
#8 28
Fully Scheduled
• Efficient when resource demands have long holding times, e.g., movies, telephone calls
• Resource being scheduled can be:– Frequency band (FDMA)– Time slot (TDMA)– Code (CDMA)– Space– Combinations of the above
#8 29
sec40
50sec/10
81017
6
=bits
bitxx
Potential for improvement
• Example– A common transmission media has a rate of 10
Mb/s and supports 50 users. The system uses fully scheduled allocation. A user has a 1 Mbytefile to transmit. The file transfer time is:
#8 30
• Dynamically scheduled– Suppose you send a message to all the
other 49 users saying, ‘I need the whole channel for about 1sec, do not use it, please’
– As long as the overhead incurred in sending the message is less than 39 sec. the user will get better performance.
– The essence of dynamically scheduled mechanisms is their distributed coordination of transmissions
Potential for improvement
#8 31
CallArrival
CallDuration
VoIPPacket Arrivals
VoIPPacket Lengths
Dynamically Scheduled
Session Interarrivals
Session Duration
Packet Interarrivals
Packet LengthsResources are requested on a burst/packet basis.
VoIP Example
Time
#8 32
Dynamically Scheduled
• Approaches– Polling– Contention (Random)– Hybrids
• Suitable for Access Networks– Geographically small
networks (few Km)– Owned by one
organization• Cable company• Telephone company• Power company
InternetAccessMedium
Downstream
Upstream
InternetAccessMedium
Downstream
Upstream
#8 33
Deterministic:Polling, Token Ring &Token Bus
• Advantage: the maximum time between users chances to transmit is bounded. (assuming a limit on the token holding time)
• Disadvantage: Time is wasted polling other users if they have no data to send.
• The technology does not scale– With geographic size– Network Speed– Number of users
#8 34
Deterministic Protocols
• Roll Call Polling– Master/slave arrangement– Master polls each node; Do you
have data to send?– If the polled node has data it is
sent otherwise next node is polled.
#8 35
Deterministic Protocols
Master
Node
Node
Node
Node
#8 36
Deterministic Protocols
• Hub Polling– No master station– Each nodes polls the next node in turn
Node
Node
NodeNode
Node
#8 37
Deterministic Protocols
• Example:– # nodes = 10– Link rate = 1 Mb/s– Packet Size = 1000 bits– Low load no queueing– 0.1 ms between nodes– Find the effective transmission rate and efficiency.
• On average destination is 5 nodes away .5 ms• Time to transmit 1000 bits = 0.5 ms + 1 ms = 1.5 ms• Effective transmission rate = 1000 bits/ 1.5 ms = 666Kb/s• Efficiency = (666 Kb/s)/(1000 Kb/s) = 0.66
– Repeat for link rate = 10 Mb/s• On average destination is 5 nodes away .5 ms• Time to transmit 1000 bits = 0.5 ms + .1 ms = .6 ms• Effective transmission rate = 1000 bits/ .6 ms = 1.67 Mb/s• Efficiency = (1.67 Mb/s)/(10 Mb/s) = 16.7%
– Conclusion Polling does not scale with link rate
#8 38
Random Access
• Each node sends data with limitedcoordination between users:
No explicit permission to transmit• Total chaos: Send data as soon as ready• Limited chaos: Listen before sending data, if the
channel is busy do not send.• Further Limiting chaos: Listen before sending
data, continue listening after sending and if collision with another transmission stop sending.
[Carrier Sense Multiple Access with Collision Detection CSMA/CD]
#8 39
Random Access
• Advantage: Simple• Disadvantage:
–No guarantee that you will ever get to send.
–The MAC protocol technology does not scale
#8 40
Random Access Protocols
• Assumptions– Overlap in time and space of two or
more transmissions causes a collision and the destruction of all packets involved.
[ No capture effects]
– One channel– For analysis no station buffering
#8 41
Random Access Protocols
– Time-Alternatives• Synchronization between users (Slotted
time)• No synchronization between users (unsloted
time)– Knowledge of the channel state-
Alternatives• Carrier sensing• Collision detection
#8 42
Random Access ProtocolsStrategies
• Aloha– No coordination between users– Send a PDU, wait for acknowledgment,
if no acknowledgment then backoff and retransmit
• Slotted Aloha– Same as Aloha only time is slotted
#8 43
Random Access ProtocolsStrategies
• p-persistent CSMA– Listen to channel, if idle or on transition from
busy to idle transmit with probability p– After sending the PDU, wait for
acknowledgment,if no acknowledgment then backoff and retransmit
• Non-persistent, if channel busy then reschedule transmission
• 1-persistent, Transmit as soon as idle
#8 44
Random Access ProtocolsStrategies
• CSMA/CD– 1-persistent but continue to
sense the channel, if collision detected then stop transmission.
– CSMA/CD is used in 10, 100 Mb/s, and 1 Gb/s Ethernet
#8 45
Limitations on Random Access Protocols
• Distance– Time to learn channel state
Propagation time
• Speed– Time to learn channel state
Clocking speed
#8 46
Random Access ProtocolsAnalysis of Aloha:
• Goal: Find Smax• Protocol Operation
– Packet of length L (sec) arrives at station i• Station i transmits immediately• Station i starts an acknowledgment timer
– If no other station transmits while i is transmitting then success
– Else a collision occurred– Station i learns that a collision occurred if the
acknowledgment timer fires before the acknowledgment arrives
#8 47
Random Access ProtocolsAnalysis of Aloha
– If collision detected then station iretransmitts at a later time, this time is pseudo-random and is determined by a backoff algorithm
• Design Issue:– Determine the maximum normalized
throughput for an Aloha system
#8 48
Random Access ProtocolsAnalysis of Aloha
Assumptions
λ1. = Average number of new message arrival to the system
2. Λ = Average number arrivals to the system, i.e.,new arrivals + retransmissions
3. The total arrival process is Poisson4. Fixed Length Packets
S = λL ≤ 1S = throughput
#8 49
Random Access ProtocolsAnalysis of Aloha
Collision Mechanism
Target Packet
2L
Target packet is vulnerable to collision for 2L Sec.
Arrival Arrival Arrival
#8 50
Random Access Protocols: Analysis of Aloha
18% is Alohafor t throughpuMaximum The
0.18=2e1 =Sor
21 =G when 0 =
S FindGe = Sor )e -G(1 + S =G
ThenL)=(S load Offered = L =G
Let)e -(1 + =
But e -1 =
sec] 2Lin arrivals Prob[no -1 =Collision ofy Probabilit
max
max
2L-2L-
2L-
2L-
dGdS
ΛΛ
Λ
Λ
Λ
ΛΛ
λ
λ
Load
Delay
0.18
#8 51
Random Access ProtocolsAnalysis of Slotted Aloha
Synchronization reduces the vulnerable period from 2L to L so the maximum throughput isincreases to 36%
Target Packet
#8 52
Number of transmissions required for success
71.2on Transmissi ofNumber Expected 1 GAt
)1(on Transmissi ofNumber Expected
)1( attemptsk in success ofy Probabilit
1Collision ofy Probabilit
1
===
=−=
−==
−==
∑∞
=
e
ePPk
PPP
eP
G
k
kcc
kcck
Gc
#8 53
Random Access ProtocolsPerformance of Unslotted and Slotted Aloha
From: “Computer Networks, 3rd Edition, A.S. Tanenbaum. Prentice Hall, 1996
#8 54
Random Access ProtocolsCSMA Protocols
• Listen to the channel before transmitting to reduce the vulnerable period
• Let D = maximum distance between nodes• Let R = the transmission rate (b/s)• Let c = speed of light = 3 x 108 m/s• The end-to-end propagation time = D/kc=τ
k is a constant for the physical media: k = .66 for fiber, k=.88 for coax
#8 55
Random Access ProtocolsCSMA Protocols
• Assume node A transmits at time t and node B at t -x, where x 0(That is, Node B transmits right before it hears A)
• If after 2D/kc sec. no collision occurred, then none will occur
• Let a= τ/L=(D/kc)/L = normalized length of the bus
• Remember L(sec) = (Packet Length [bits])/R [b/s]
• As a --> 1, CSMA performance approaches Aloha performance
#8 56
Random Access ProtocolsCSMA Protocols
• Limits on a– Want a small to keep vulnerable period
short by having:• Short bus• Lower speeds• Long packets
– Lower limit (Minimum) packet length to upper bound a
– Maximum packet length to be fair
a= DR/Xkcwhere
X= packet length in bits
#8 57
Random Access ProtocolsPerformance
From: “Computer Networks, 3rd Edition, A.S. Tanenbaum. Prentice Hall, 1996
#8 58
Random Access ProtocolsCSMA Protocols
• Example: Ethernet– Rate = 100 Mb/s– Minimum packet size = 512 bits– Maximum packet size = 12144 bits– D (max per segment) = 500 m– a --> [0.001, 0.03]
• CSMA networks do not scale– Increase D performance degrades– Increase R performance degrades
#8 59
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.90.
02
0.03
0.06
0.13
0.25 0.
5 1 2 4 8 16 32 64
0.81
0.51
0.14
S
G
a = 0.01
Non-Persistent CSMA Throughput
a = 0.1
a = 1
• Higher maximum throughput than 1-persistent for small a
• Worse than Aloha for a > 1
Adapted from: Leon-Garcia & Widjaja: Communication Networks
#8 60
Performance of Random Access Protocols
0
0.2
0.4
0.6
0.8
1
0.01 0.1 1
ALOHA
Slotted ALOHA
1-P CSMA
Non-P CSMA
CSMA/CD
a
ρmax
• For small a: CSMA-CD has best throughput• For larger a: Aloha & slotted Aloha better throughput
From: Leon-Garcia & Widjaja: Communication Networks
#8 61
Random Access ProtocolsCSMA Protocols: States
From: “Computer Networks, 3rd Edition, A.S. Tanenbaum. Prentice Hall, 1996
#8 62
Collision Free Protocols
• Collision free protocols establish rules to determine which stations sends after a successful transmission.
• Assume there are N stations with unique addresses 0 to N-1.
• A contention interval is a period after a successful transmission that is divided into N time slots, one for each station.
#8 63
Collision Free Protocols• If a station has a PDU to send it sets
a bit to 1 in its time slot in the contention interval.
• At the end of the contention interval all nodes know who has data to send and the order in which it will be sent.
From: “Computer Networks, 3rd Edition, A.S. Tanenbaum. Prentice Hall, 1996
#8 64
Collision Free Protocols
• Example of using resources, time to “schedule” transmissions
• Problems:– Fairness– Flexibility
• Many systems use the basic approach of collision free protocols
#8 65
Hybrids
• Hybrid approaches combine:– Random access– Fully scheduled
• Idea is to limit the time (resources) involved in collisions
• Different protocols can be used in the upstream and downstream directions
#8 66
Reservation Aloha
• Consider a slotted system with N slots/Frame
• Each slot can be in one of three states:• Empty, i.e., not is use• Mine, i.e., in use by me• Other, i.e., in use by another node
• Protocol• If state is mine then continue to use it in
next frame• If state is other then do not send in that
time slot in next frame• If state is empty then contend for that slot
sin Al h in n xt f m
#8 67
Reservation Aloha
• At low loads the network performs like a random access systems, i.e., no waiting for permission to send. Send in next frame.
• At high loads the systems performs like a TDM system.
• Example – Some time “reserved” for contention.– Distributed algorithm
• This scheme has a problem with fairness• How are opportunities to transmit in time slots
granted?
#8 68
Random Access and Reservations
• Distributed systems: Stations implement a decentralized algorithm to determine transmission order, e.g., reservation Aloha
• Centralized systems: A central controller accepts requests from stations and issues grants to transmit– Frequency Division Duplex (FDD): Separate frequency bands
for uplink & downlink– Time-Division Duplex (TDD): Uplink & downlink time-share
the same frequency channel• The centralized system is used in many access
technologies, e.g.,– DOCSIS– IEEE 802.16– WCDMA HSDPA– CDMA2000 EV-DO
Adapted from: Leon-Garcia & Widjaja: Communication Networks
#8 69
Reservation Systems
TimeCycle n
Reservationinterval
Frame transmissions
r d d d r d d d
Cycle (n + 1)
r = 1 2 3 M
• Transmissions organized into cycles (or frames)• Cycle: reservation interval + frame transmissions• Reservation interval has a minislot for each station to request
reservations for frame transmissions; minislot can carry other information, e.g., number of frame to TX, station backlog, CQI
Adapted from: Leon-Garcia & Widjaja: Communication Networks
Upstream Transmissions
Minislot
#8 70
Reservation System Options• Centralized or distributed system
– Centralized systems: A central controller listens to reservation information, decides order of transmission, issues grants
– Distributed systems: Each station determines its slot for transmission from the reservation information
• Single or Multiple Frames– Single frame reservation: Only one frame transmission can
be reserved within a reservation cycle– Multiple frame reservation: More than one frame
transmission can be reserved within a frame• Channelized or Random Access Reservations
– Channelized (typically TDMA) reservation: Reservation messages from different stations are multiplexed without any risk of collision
– Random access reservation: Each station transmits its reservation message randomly until the message goes through
Adapted from: Leon-Garcia & Widjaja: Communication Networks
#8 71
Reservation System
• System Characteristics– Asymmetric
• Upstream– Minislots with requests for resources– Access Minislots via random access protocol
• Downstream– Accepts minislots and includes grants for
transmission– Grants control the flow on the upstream link– Order of grants established via a “scheduling”
algroithm
#8 72
Throughput• Let
– R = Link rate (b/s)– L = packet size (bits)– V = minislot size (sec)– M = Number of stations– X = L/R
• Assume– Propagation delay < X Access network– Heavy load stations have packets to send– All requests are granted– one minislot needed for each packet/station
• Time to transmit M packets = Mv+MX
XMXM
MXS vv +=
+=
1
1max
Adapted from: Leon-Garcia & Widjaja: Communication Networks
#8 73
Throughput
• If k frame transmissions can be reserved with a reservation message and if there are Mstations, as many as Mk frames can be transmitted in XM(k+v) seconds
XkMkXM
MkXS vv +=
+=
1
1max
Adapted from: Leon-Garcia & Widjaja: Communication Networks
#8 74
Throughput: with random access contention for Minislots
• Real systems have too many nodes for each to get a fixed minislot.
• Therefore a random access protocol is used to transmit in a minslot.– A station attempts to obtain a grant by
transmitting in a minslot in the upstream direction.
– If successful the station will get the grant on the down stream
– If unsuccessful then assume collision, backoffand retry.
#8 75
Throughput: with random access contention for Minislots
• Assume slotted Aloha is used for contention for minislots.
• On average, each reservation takes at least e = 2.71 minislot attempts
• Effect is just to make the minislotsseem longer
XX(1+ev)
1 1 + 2.71v/X
Smax = =
#8 76
A user perspective• Call model
– User connects to service– Then does activity– Examples
• Dial-up models• Cell phones
• Always-on model– User is “always” connected– Have packet just “send” it– “send” it happens in some
controlled way– No call process (dialing #) – No waiting for connection– Examples:
• Cable modems• WiFi• DSL
• Call model– Fully scheduled– Efficient for long holding times
• Always-on model– Dynamically scheduled– Needs coordination– Support large number of users– Often users can send at full link speed– Efficient for bursty traffic