atm_cong
DESCRIPTION
Congestion in atmTRANSCRIPT
![Page 1: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/1.jpg)
Raj JainThe Ohio State University
1
ATM Traffic ManagementATM Traffic Management
Raj JainProfessor of Computer and Info. Sciences
The Ohio State UniversityColumbus, OH 43210-1277
http://www.cis.ohio-state.edu/~jain/
Dollar Day Sale
One Megabit memory, One Megabyte disk,One Mbps link, One MIP processor, one
dollar each.....
![Page 2: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/2.jpg)
Raj JainThe Ohio State University
3
ATM Networks: OverviewATM Networks: Overview❑ STM = Synchronous Transfer Mode,
ATM = Asynchronous Transfer Mode
Allows any-speed and even variable rate connection
Broadband = Rate greater than primary rate (1.5 Mbps)
❑ ATM = Short fixed size 53-byte cells
❑ Connection oriented ⇒ Virtual Channels (VC)
![Page 3: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/3.jpg)
Raj JainThe Ohio State University
4
❑ Why worry about congestion?
❑ Congestion schemes for ATM
❑ Rate vs Credit: Key issues
❑ Explicit Rate-based Control
❑ ABR Traffic Management
Overview
![Page 4: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/4.jpg)
Raj JainThe Ohio State University
6
Why Worry About Congestion?Why Worry About Congestion?Q: Will the congestion problem be solved when:
❑ Memory becomes cheap (infinite memory)?
❑ Links become cheap (very high speed links)?
❑ Processors become cheap?
A: None of the above.
No buffer Old age19.2 kb/s 1 Mb/s
SS SS SS SS SS SS SS SS
Time = 7 hoursFile transfer time = 5 mins
![Page 5: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/5.jpg)
Raj JainThe Ohio State University
7
Conclusions:❑ Congestion is a dynamic problem.
Static solutions are not sufficient
❑ Bandwidth explosion⇒ More unbalanced networks
❑ Buffer shortage is a symptom not thecause.
AA
BBSS
CC
DDAll links 1 Gb/s
![Page 6: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/6.jpg)
Raj JainThe Ohio State University
8
Economic ReasonsEconomic Reasons❑ Network is a shared resource
Because it is expensive and needed occasionally(Like airplanes, emergency rooms)
❑ Most costs are fixed.Cost for fiber, switches, laying fiber andmaintaining them does not depend upon usage⇒ Underutilization is expensive
❑ But overutilization leads to user dissatisfaction.
❑ Need a way to keep the network maximally utilized
![Page 7: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/7.jpg)
Raj JainThe Ohio State University
9
One Scheme or Many?One Scheme or Many?❑ Fundamental principle of control theory:
❑ Control faster than feedback ⇒ InstabilityControl slower than feedback ⇒ non-responsivenessIdeal: Control rate ≈ Feedback rate
❑ Lesson: No scheme can cure a congestion lasting less than itsfeedback delay.
Time
State Fast
SlowTarget
![Page 8: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/8.jpg)
Raj JainThe Ohio State University
13
Classes of ServiceClasses of Service❑ CBR (Constant bit rate): User dec8 Tcres required rate.
Throughput, delay and delay variation guaranteed.
❑ VBR (Variable bit rate): User dec8ares average and max rate.
– rt-VBR (Real-time variable bit rate): Conferencing.Max delay and delay variation guaranteed.
– nrt-VBR (non-real time variable bit rate): Stored video.
❑ ABR (Available bit rate): Follows feedback instructions.Network gives maximum throughput with minimum loss.
❑ UBR (Unspecified bit rate):User sends whenever it wants. No feedback mechanism.No guarantee. Cells may be dropped during congestion.
![Page 9: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/9.jpg)
Raj JainThe Ohio State University
15
Traffic Management FunctionsTraffic Management Functions❑ Connection Admission Control (CAC):
Can requested bandwidth and quality of service be supported?
❑ Traffic Shaping: Limit burst length. Space-out cells.
❑ Usage Parameter Control (UPC):Monitor and control traffic at the network entrance.
❑ Network Resource Management:Scheduling, Queueing, virtual path resource reservation
❑ Selective cell discard:Cell Loss Priority (CLP) = 1 cells may be droppedCells of non-complient connections may be dropped
❑ Frame Discarding❑ Feedback Controls: Network tells the source to increase or
decrease its load.
![Page 10: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/10.jpg)
Raj JainThe Ohio State University
17
Initial Binary Rate-based SchemeInitial Binary Rate-based Scheme
❑ We invented DECbit scheme (1986).Implemented in many standards since 1986.
❑ Forward explicit congestion notification (FECN) in Framerelay
❑ Explicit forward congestion indicator (EFCI) set to 0 at source.Congested switches set EFCI to 1
❑ Every nth cell, destination sends an resource management (RM)cell to the source indicating increase amount or decrease factor
EFCI
DestinationDestinationSourceSource
RM
![Page 11: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/11.jpg)
Raj JainThe Ohio State University
19
The Explicit Rate SchemeThe Explicit Rate Scheme
❑ Sources send one RM cell every n cells
❑ The RM cells contain “Explicit rate”
❑ Destination returns the RM cell to the source
❑ The switches adjust the rate down
❑ Source adjusts to the specified rate
Explicit RateExplicit RateCurrent Cell RateCurrent Cell Rate
![Page 12: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/12.jpg)
Raj JainThe Ohio State University
20
ERICA Switch AlgorithmERICA Switch Algorithm
❑ Each manufacturer will have its own explicit rate switchalgorithm
❑ Explicit Rate Indication for Congestion Avoidance (ERICA)is the most thoroughly analyzed algorithm among disclosedalgorithms
❑ Shown to be efficient, fair, fast transient response, able tohandle bursty TCP traffic
❑ ERICA+ allows low delay even at 100% utilization andprovides stability in the presence of high frequency VBRbackground traffic
❑ Being implemented by several vendors
![Page 13: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/13.jpg)
Raj JainThe Ohio State University
21
Segment-by-Segment ControlSegment-by-Segment Control
❑ Virtual source/virtual destinations follow allnotification/control rules
❑ Can be hop-by-hop
❑ Virtual dest/sources maintain per-VC queues.
SS DD
EFCI
RM
EFCI
RM
EFCI
RM
EFCI
RM
SS DD
EFCI
RM
EFCI
RM
![Page 14: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/14.jpg)
Raj JainThe Ohio State University
22
MulticastMulticast
RM Cellmin(B,C)
RM CellA
RM C
ell ARM
Cell B
RM Cell
ARM Cell
c
![Page 15: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/15.jpg)
Raj JainThe Ohio State University
23
Outstanding IssuesOutstanding Issues❑ Bursty sources: Client server, transactions, WWW
❑ Effect of parameters: Optimal parameter values
❑ Priority service for RM cells
❑ Multicast
❑ Connection admission control (CAC)
❑ TCP/IP over UBR
❑ Non-conforming sources
❑ Optimal Source Strategy: Parameter + Out-of-rate cells
❑ Virtual Source/destination
❑ Implicit feedback schemes: Heterogeneous Networks
TCP/IPTCP/IP ATMATM IPXIPX
![Page 16: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/16.jpg)
Raj JainThe Ohio State University
24
Congestion:Congestion:SummarySummary
❑ Traffic Management is key tosuccess of ATM
❑ Several different methods: CAC,Shaping, UPC, Scheduling, ...
❑ Service categories:CBR, VBR,ABR, UBR
❑ Binary feedback too slow for ratecontrol. Especially for satellites.
❑ ER switches provide much betterperformance than EFCI.
![Page 17: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/17.jpg)
Raj JainThe Ohio State University
25
ReferencesReferences❑ R.Jain, "Congestion Control and Traffic Management in
ATM Networks: Recent Advances and A Survey", Invitedsubmission to Computer Networks and ISDN Systems,February 1995.Available at http://www.cis.ohio-state.edu/~jain/
❑ User-Network Interface Specifications, V3.0, Prentice-Hall,September 10, 1993., (515)-284-6751V3.1 released by ATM Forum, January 1995.
❑ K. Siu and R. Jain, "A Brief Overview of ATM: ProtocolLayers, LAN Emulation, and Traffic Management,"Computer Communications Review (ACM SIGCOMM),April 1995.
❑ “ATM Forum Traffic Management Specification, Version4.0,” ATM Forum/95-0013R13, April 1996.
![Page 18: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/18.jpg)
Raj JainThe Ohio State University
26
❑ Craig Partridge, Gigabit Networking, Addison-Wesley, 1993.
❑ S. Akhtar, Congestion Control in a Fast Packet SwitchingNetwork, Master's Thesis, Washington University, 1987.[Originated the concept of leaky bucket.]
❑ Allyn Romanow, Packet Discard Strategies for Controlling AMCongestion, ATM Forum Traffic Management SubworkingGroup submission #94-0107, January 17-24, 1994.
❑ S. Kalyanaraman, R. Jain, S. Fahmy, R. Goyal, F. Lu and S.Srinidhi, ``Performance of TCP/IP over ABR,'' Submitted toGlobecom'96.
❑ R. Jain, S. Kalyanaraman, R. Viswanathan, "The OSUScheme for Congestion Avoidance in ATM networks UsingExplicit Rate Indication," Proceedings WATM'95, Paris,December 1995, Also available an extended OSU technicalreport and an extended conference version.
![Page 19: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/19.jpg)
Raj JainThe Ohio State University
27
❑ "Killer App? It's the high-speed backbone," An interviewwith Raj Jain. Government Computer News, October 16,1995, pp. 24.
❑ R. Jain, "ABR Service on ATM Networks: What is it?"Network World, June 24, 1995.
❑ A. Charny, D. Clark, R. Jain, "Congestion Control withExplicit Rate Indication," Proc. ICC'95, June 1995, 10 pp.
![Page 20: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/20.jpg)
Raj JainThe Ohio State University
28
ERICA+: Full UtilizationERICA+: Full Utilization❑ Allows operation at any point between the knee and the cliff
❑ The queue time can be set to any desired value.
❑ Allows utilization to be 100%
LinkUtilization
TimeQueueLength
50
Thr
ough
put
Load
Del
ay
Load
100%
![Page 21: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/21.jpg)
Raj JainThe Ohio State University
29
❑ Problem: Flows with large packets get more bandwidth.
❑ Bit-by-bit Round Robin
❑ Fair Queueing: Compute end-times of packets under bit-by-bit round robin.
❑ Weighted Fair Queueing: Different flows are alloweddifferent bandwidth.
Fair QueueingFair Queueing❑ Goal: Isolation of flows.
❑ Round Robin:
![Page 22: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/22.jpg)
Raj JainThe Ohio State University
30
Selective Cell DiscardingSelective Cell Discarding
❑ Selective Cell Discard: May discard CLP =1 or CLP =0cells of non-complient VCs.
❑ Goal: Discard cells of VCs not following contract
❑ Normal Discard: CLP=1 cells are discarded if buffer is full.
![Page 23: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/23.jpg)
Raj JainThe Ohio State University
31
Early Packet DiscardEarly Packet Discard
Problem: Unfair to late-comers.
❑ Random Discard: Randomly select a cell in the queue anddiscard it.Problem: Need to look inside the queues.
❑ Early Packet Discard: When a queue threshold exceeds,
– First cell of the next arriving packet is discarded.
– Drop all cells of discarded packets.
❑ Goal: To minimize packet loss rate
❑ Normal Discard: Cells are discarded if buffer is full.
![Page 24: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/24.jpg)
Raj JainThe Ohio State University
32
Cell Delay VariationCell Delay Variation❑ Cell Transfer Delay (CTD): First bit out to last bit in
❑ Cell Delay Variation (CDV) = CTDmax - CTDmin
– Peak-to-peak CDV
– Instantaneous CDV
![Page 25: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/25.jpg)
Raj JainThe Ohio State University
33
Instantaneous CDVInstantaneous CDV❑ I-CDV= Actual-Expected arrival time
❑ Expected = Emission + Nominal delay
❑ Cell Delay Variation Window (CDV-W)CDV-W = |I-CDV(Max)| + |I-CDV(Min)|
❑ Cells arriving outside window are considered lost
❑ Large CDV ⇒ Large buffers ⇒ Higher cost
Expected Actual
![Page 26: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/26.jpg)
Raj JainThe Ohio State University
34
Generic Flow Control (GFC)Generic Flow Control (GFC)❑ ATM header format designed from DQDB cell headers.
The first byte in DQDB is access control for contention.
❑ Some countries wanted to keep this in ATM forcompatibility and others saw no use for it.
❑ Compromise ⇒ One half of the octet was reserved for“Generic Flow Control” to be defined at some future point.
❑ British Telecom adapted multiservice flow control (MSFC)from Orwell ring and proposed to use it for GFC
![Page 27: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/27.jpg)
Raj JainThe Ohio State University
35
Fast Resource ManagementFast Resource Management❑ User requests to send a burst.
❑ Network reserves the resources andgrants the request
❑ User sends the bursts
❑ If network rejects, ???
![Page 28: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/28.jpg)
Raj JainThe Ohio State University
36
ERICA FeaturesERICA Features❑ Measured overload/load at switch
❑ Small queue lengths during steady state
❑ Fast response due to optimistic design
❑ Parameters: Few, insensitive, easy
❑ Insensitive to source not using their allocated rates
❑ Several options: Backward Explicit CongestionNotification
❑ Simplified switch algorithm
❑ Optimized all steps. Eliminated unncessary steps.Eliminated many parameters
![Page 29: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/29.jpg)
Raj JainThe Ohio State University
37
ERICA+: Switch AlgorithmERICA+: Switch Algorithm
❑ Target cell rate = Target Utilization × Link Capacity
❑ Target Utilization= fn(Current load, Queue length, Queue drain time goal)
❑ Rest is similar to ERICA
❑ Features:
– Queue length is bounded during overload
– No queue underflow ⇒ Switches keep ABR cells waiting tobe transmitted as soon as the bandwidth becomes available.
– 100% Utilization even with VBR
![Page 30: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/30.jpg)
Raj JainThe Ohio State University
38
Control MechanismsControl Mechanisms❑ nrt-VBR: Open-Loop + Optional closed loop component
– Traffic shaping, CLP
– EFCI optional⇒ No switch or end-system behavior specified
❑ UBR: Local policy. CLP=0 or 1. Not subject to CAC.
❑ ABR:
– CLP=1 data cells not allowed.
– No EFCI in RM cells
– All parameters are negotiated independently for the twodirections
– MCR=0 not subject to CAC due to load
![Page 31: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/31.jpg)
Raj JainThe Ohio State University
39
Rate RepresentationRate Representation
❑ Rate in cells/second = [2e(1+m/512)]*nz
❑ Example:0-1-01010-0 1100 1010 = 218(1+202/512)= 262144 × 1.394523= 365,566 cells/sec = 155 Mbps
Reserved Nonzero Exponent Mantissa
11 5 9 Size in bits
![Page 32: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/32.jpg)
Raj JainThe Ohio State University
40
OSU Congestion PrinciplesOSU Congestion Principles❑ Input rate (and not queue length) is the load measure
❑ Congestion avoidance (and not congestion control) shouldbe the goal
❑ Transient performance (and not the steady stateperformance) is more important
![Page 33: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/33.jpg)
Raj JainThe Ohio State University
41
Which Link is MoreWhich Link is MoreOverloaded?Overloaded?
QueueLength
Time
1000 QueueLength
Time
100
Link A Link B
![Page 34: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/34.jpg)
Raj JainThe Ohio State University
42
Answer: It Depends!Answer: It Depends!❑ Link Speed: OC-12 or T1?
❑ Control: Rate or Window?Q = Window, dQ/dt = Rate
❑ For Rate Control: Monitor input rate
QueueLength
Time
1000 QueueLength
Time
100
Link A Link B
![Page 35: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/35.jpg)
Raj JainThe Ohio State University
43
Conclusions IConclusions I☞ Instantaneous queue length is not a good indicator
of load for a rate controlled system.Q(t) = Q(t-1) + Input rate(t) - Service rate(t)
☞ Using queue length as the load indicator in a ratecontrolled system leads to unnecessaryoscillations.
☞ Input rate monitoring not only correctly tellswhether the system is overloaded, it also tells bywhat factor.
☞ Queue = n is not a good goal. Input rate=servicerate is.
![Page 36: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/36.jpg)
Raj JainThe Ohio State University
44
Congestion AvoidanceCongestion Avoidance❑ High throughput, Low delay
❑ Small queues
❑ Bounded oscillations ⇒ Good for Video traffic
LinkUtilization
TimeQueueLength
Time
5
Thr
ough
put
Load
Del
ay
Load
![Page 37: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/37.jpg)
Raj JainThe Ohio State University
45
Why Worry About Transients?Why Worry About Transients?
On most networks:
❑ There are no infinite sources.
❑ Sources come and go
❑ VCs may stay but are mostly inactive
❑ Traffic is highly bursty
⇒ Networks are operating in the transient region,most of the time.
Time
ThroughputQueue Length
orUtilization Transient
Steady State
![Page 38: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/38.jpg)
Raj JainThe Ohio State University
46
Burst PerformanceBurst Performance
Time
RateArrival
Departure
EffectiveThroughput
Time
Rate
BurstResponseTime
EffectiveThroughput
![Page 39: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/39.jpg)
Raj JainThe Ohio State University
47
Legacy LANs vs ATMLegacy LANs vs ATM❑ Today’s LANs have a very fast transient response. Can get
to the peak rate within a few microseconds
❑ On ATM LANs:Wait for connection setup and then...Everytime, a burst arrives, take several milliseconds to rampup
❑ Q: Given 100 Mbps Switched Ethernet and 155 Mbps ATMat the same price, which one would you buy?
![Page 40: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/40.jpg)
Raj JainThe Ohio State University
48
QuizQuizT F Please check True/False
1. o o Congestion is not a problem in high-speed networks
2. o o User parameter control (UPC) allows a user to set itsparameters
3. o o CDVT measures the cell delay variation caused by the network
4. o o ABR users do not have to specify CDVT
5. o o GCRA allows a network to determine conforming and non-conforming cells
6. o o All non-conforming cells are dropped at the source
7. o o Credit based scheme requires per-VC queueing
8. o o EFCI is better than explicit rate for high-speed networks
9. o o VSVD allows a network operator to use proprietary controlscheme inside its networks.
![Page 41: atm_cong](https://reader035.vdocuments.site/reader035/viewer/2022062423/55cf92e3550346f57b9a496f/html5/thumbnails/41.jpg)
Raj JainThe Ohio State University
49
Thank You!Thank You!