Advanced Topics in Congestion Control

1
Advanced Topics inCongestion Control

• Slides originally developed by S. Kalyanaraman
  (RPI) based in part upon slides of Prof. Raj Jain
  (OSU), Srini Seshan (CMU), J. Kurose (U Mass),
  I.Stoica (UCB)

2
Overview

• Queue Management Schemes
• RED, ARED, FRED, BLUE, REM
• TCP Congestion Control (CC) Modeling,
• TCP Friendly CC
• Accumulation-based Schemes TCP Vegas, Monaco
• Static Optimization Framework Model for
  Congestion Control
• Explicit Rate Feedback Schemes (ATM ABR ERICA)

3
Readings

• Refs Chap 13.21, 13.22 in Comer textbook
• Floyd and Jacobson "Random Early Detection
  gateways for Congestion Avoidance"
• Ramakrishnan and Jain, A Binary Feedback Scheme
  for Congestion Avoidance in Computer Networks
  with a Connectionless Network Layer,
• Padhye et al, "Modeling TCP Throughput A Simple
  Model and its Empirical Validation"
• Low, Lapsley "Optimization Flow Control, I
  Basic Algorithm and Convergence"
• Kalyanaraman et al "The ERICA Switch Algorithm
  for ABR Traffic Management in ATM Networks"
• Harrison et al "An Edge-based Framework for Flow
  Control"

4
Queuing Disciplines

• Each router must implement some queuing
  discipline
• Queuing allocates bandwidth and buffer space
• Bandwidth Which packet to serve next
  (scheduling)
• Buffer Which packet to drop next (buffer mgmt)
• Queuing also affects latency

Traffic Sources
Traffic Classes
Class A
Class B
Class C
Drop
Scheduling
Buffer Management
5
Typical Internet Queuing

• FIFO and Drop-tail
• Simplest choice
• Used widely in the Internet
• FIFO (first-in-first-out)
• Implies single class of traffic
• Drop-tail
• Arriving packets get dropped when queue is full
  regardless of flow or importance
• Important distinction
• FIFO Scheduling discipline
• Drop-tail Buffer management policy

6
FIFO and Drop-tail Problems

• FIFO Issues In a FIFO discipline, the service
  seen by a flow is convoluted with the arrivals of
  packets from all other flows!
• No isolation between flows full burden on e2e
  control
• No policing send more packets ? get more service
• Drop-tail Issues
• Routers are forced to have have large queues to
  maintain high utilizations
• Larger buffers gt larger steady state
  queues/delays
• Synchronization End hosts react to same events
  because packets tend to be lost in bursts
• Lock-out A side effect of burstiness and
  synchronization is that a few flows can
  monopolize queue space

7
Design Objectives

• Keep throughput high and delay low (i.e. knee)
• Accommodate bursts
• Queue size should reflect ability to accept
  bursts rather than steady-state queuing
• Improve TCP performance with minimal hardware
  changes

8
Queue Management Ideas

• Synchronization, lock-out
• Random drop Drop a randomly chosen packet
• Drop front Drop packet from head of queue
• High steady-state queuing vs. burstiness
• Early drop Drop packets before queue full
• Do not drop packets too early because queue may
  reflect only burstiness and not true overload
• Misbehaving vs. Fragile flows
• Drop packets proportional to queue occupancy of
  flow
• Try to protect fragile flows from packet loss
  (e.g. color them or classify them on the fly)
• Drop packets vs. Mark packets
• Dropping packets interacts w/ reliability
  mechanisms
• Mark packets Need to trust end-systems to
  respond!

9
Packet Drop Dimensions
Aggregation
Per-connection state
Single class
Class-based queuing
Drop position
Head
Tail
Random location
Overflow drop
Early drop
10
Random Early Detection (RED)
Min thresh
Max thresh
Running Average Queue Length
P(drop)
1.0
maxP
minth
maxth
Running Avg Q Length
11
Random Early Detection (RED)

• Maintain running average of queue length avgQ
• Low pass filtering
• If avgQ lt minth do nothing
• Low queuing, send packets through
• If avgQ gt maxth, drop packet
• Protection from misbehaving sources
• Else mark (or drop) packet in a manner
  proportional to queue length bias to protect
  against synchronization
• Pb maxp(avgQ - minth) / (maxth - minth)
• Further, bias Pb by history of unmarked packets
• Pdrop Pb/(1 - countPb)

12
RED Issues

• Issues
• Breaks synchronization well
• Extremely sensitive to parameter settings
• Wild queue oscillations upon load changes
• Fail to prevent buffer overflow as sources
  increases
• Does not help fragile flows (e.g. small window
  flows or retransmitted packets)
• Does not adequately isolate cooperative flows
  from non-cooperative flows
• Isolation
• Fair queuing achieves isolation using per-flow
  state
• RED penalty box Monitor history for packet
  drops, identify flows that use disproportionate
  bandwidth

13
Variant ARED (Feng, Kandlur, Saha, Shin 1999)

• Adaptive RED
• Motivation RED extremely sensitive to number of
  sources and parameter settings
• Idea Adapt maxp to load
• If avgQ lt minth, decrease maxp
• If avgQ gt maxth, increase maxp
• No per-flow information needed

14
Variant FRED (Ling Morris 1997)

• Flow RED
• Motivation Marking packets in proportion to
  flow rate is unfair (e.g. adaptive vs.
  non-adaptive flows)
• Idea
• A flow can buffer up to minq packets w/o being
  marked
• A flow that frequently buffers more than maxq
  packets gets penalized
• All flows with backlogs in between are marked
  according to RED
• No flow can buffer more than avgcq packets
  persistently
• Where avgcq average per-flow buffer use
• Need per-active-flow accounting

15
Variant BLUE (Feng, Kandlur, Saha, Shin 1999)

• BLUE ??? (2nd 3rd authors at IBM)
• Motivation Wild oscillation of RED leads to
  cyclic overflow underutilization
• Algorithm
• On buffer overflow, increment marking probability
• On link idle, decrement marking probability

16
Variant Stochastic Fair Blue

• Motivation Protection against non-adaptive
  flows
• Algorithm
• L hash functions map a packet to L bins (out of
  NxL )
• Marking probability associated with each bin is
• Incremented if bin occupancy exceeds threshold
• Decremented if bin occupancy is 0
• Packets marked with min p1, , pL

h1
h2
hL
hL-1
nonadaptive
adaptive
17
Stochastic Fair Blue continued

• Idea
• A non-adaptive flow drives marking prob to 1 at
  all L bins it is mapped to
• An adaptive flow may share some of its L bins
  with non-adaptive flows
• Non-adaptive flows can be identified and
  penalized with reasonable state overhead (not
  necessarily per-flow)
• Large numbers of bad flows may cause false
  positives

18
REM Athuraliya Low 2000

• REM Random Exponential Marking
• Decouple congestion performance measure
• Price adjusted to match rate and clear buffer
• Marking probability exponential in price

REM
RED
1
Avg queue
19
Comparison of AQM Performance

• REM
• queue 1.5 pkts
• utilization 92
• g 0.05, a 0.4, f 1.15

DropTail queue 94
20
The DECbit Scheme

• Basic ideas
• Mark packets instead of dropping them
• Special support at routers and e2e
• Scheme
• On congestion, router sets congestion indication
  (CI) bit on packet
• Receiver relays bit to sender
• Sender adjusts sending rate
• Key design questions
• When to set CI bit?
• How does sender respond to CI?

21
Setting CI Bit
Queue length
Current time
Time
Previous cycle
Current cycle
Averaging interval
AVG queue length (previous busyidle)
current interval/averaging interval
22
DECbit Routers

• Router tracks average queue length
• Regeneration Cycle Queue goes from empty to
  non-empty to empty
• Average from start of previous cycle
• If average gt 1 ? router sets bit for flows
  sending more than their share
• If average gt 2 ? router sets bit in every packet
• Threshold is a trade-off between queuing and
  delay
• Optimizes power (throughput / delay)
• Compromise between sensitivity and stability
• Acks carry bit back to source

23
DECbit Source

• Source averages across acks in window
• Congestion if gt 50 of bits set
• Will detect congestion earlier than TCP
• Additive increase, multiplicative decrease
• Decrease factor 0.875
• Increase factor 1 packet
• After change, ignore DECbit for packets in flight
  (vs. TCP ignore other drops in window)
• No slow start

24
Alternate Congestion Control Models

• Loss-based TCP Classic, TCP Reno,
• Accumulation-based schemes TCP Vegas, Monaco
• Use per-flow queue contribution (backlog) as a
  congestion estimate instead of loss rate
• Explicit rate-based feedback
• Controller at bottleneck assigns rates to each
  flow
• Packet Pair congestion control Not covered
• WFQ at bottlenecks isolates flows, and gives fair
  rates
• Packet-pair probing discovers this rate and sets
  source rate to that.

25
TCP Reno (Jacobson 1990)
window
time
CA
SS
SS Slow Start CA Congestion Avoidance
26
TCP Vegas (Brakmo Peterson 1994)
window
time
CA
SS

• Accumulation-based scheme
• Converges, no retransmission
• provided buffer is large enough for no loss.

27
Accumulation Single Queue

• Flow i at router j
• Arrival curve Aij(t)
• Service curve Sij(t)
• Cumulative
• Continuous
• Non-decreasing
• If no loss, then

Aij(t)
bit
delay
Sij(t)
b2
qij(t1)
b1
time
t2
t1
27
28
Accumulation Series of Queues
Model as a single queue
1
j
j1
J
ingress
egress
dj
fi
µij
?i,j1
µi
?i
28
29
Queue vs. Accumulation Behavior

• Queue qij(t) -- info of flow i queued in a FIFO
  router j

• Accumulation ai(t) -- info of flow i queued in a
  set of FIFO routers 1, , J

• The collective queuing behavior of a set of FIFO
  routers looks similar to that of one single FIFO
  router

29
30
Accumulation Distributed, Time-shifted Sum
31
Control Policy

• Control objective keep
• If , no way to probe increase of available
  bw

• Control algorithm

31
32
Two Accumulation-Based Schemes

• Monaco
• Accumulation estimation out-of-band / in-band
• Congestion response
• Additive Increase/Additive Decrease (AIAD)
• Vegas
• Accumulation estimation in-band
• Congestion response
• Additive Increase/Additive Decrease (AIAD)

32
33
Accumulation vs. Monaco Estimator


time
1
j
j1
J
34
Accumulation vs. Monaco Estimator
1
jf
Jf
jf1
djf
fi
data
µij
?i,j1
µi
?i
ctrl
Jb
jb1
jb
1
djb
ctrl
out-of-bd ctrl
classifier
FIFO
in-band ctrl, data pkt
34
35
Monaco

• Congestion estimation
• Out-of-band and in-band control packets
• Congestion response
• If qm lt a, cwnd(k1) cwnd(k) 1
• If qm gt ß, cwnd(k1) cwnd(k) 1 1 a lt
  ß 3

35
36
TCP Vegas

• Congestion estimation
• Define qv ( cwnd / rttp cwnd / rtt ) rttp
• where rttp is estimate of round trip
  propagation delay
• Congestion response
• if qv lt a, cwnd(k1) cwnd(k) 1
• if qv gt ß, cwnd(k1) cwnd(k) 1 1 a lt ß
  3

36
37
Vegas Accumulation Estimator

• The physical meaning of qv
• rtt rttp rttq rttq is queuing time
• qv ( cwnd / rttp cwnd / rtt ) rttp
• ( cwnd / rtt ) ( rtt rttp )
• ( cwnd / rtt ) rttq if rtt is typical
  
• sending rate rttq Littles Law
• packets backlogged Littles Law again

• So vegas maintains a ß number of packets queued
  inside the network
• It adjusts sending rate additively to achieve this

37
38
Accumulation vs. Vegas Estimator

• Backlogv

38
39
Vegas vs. Monaco Estimators

• Vegas accumulation estimator
• Ingress-based
• Round trip (forward data path and backward ack
  path)
• Sensitive to ack path queuing delay
• Sensitive to round trip propagation delay
  measurement error
• Monaco accumulation estimator
• Egress-based
• One way (only forward data path)
• Insensitive to ack path queuing delay
• No need to explicitly know one way propagation
  delay

40
Queue, Utilization w/ basertt Errors
40
41
TCP Modeling

• Given the congestion behavior of TCP can we
  predict what type of performance we should get?
• What are the important factors
• Loss rate
• Affects how often window is reduced
• RTT
• Affects increase rate and relates BW to window
• RTO
• Affects performance during loss recovery
• MSS
• Affects increase rate

42
Overall TCP Behavior

• Lets focus on steady state (congestion
  avoidance) with no slow starts, no timeouts and
  perfect loss recovery
• Some additional assumptions
• Fixed RTT
• No delayed ACKs

Window
Time
43
Derivation
window
t

• Each cycle delivers 2w2/3 packets
• Assume Each cycle delivers 1/p packets 2w2/3
• Delivers 1/p packets followed by a drop
• gt Loss probability p/(1p) p if p is small.
• Hence

44
Alternate Derivation

• Assume Loss is a Bernoulli process with
  probability p
• Assume p is small
• wn is the window size after nth RTT

45
Law

• Equilibrium window size
• Equilibrium rate
• Empirically constant a 1
• Verified extensively through simulations and on
  Internet
• References
• T.J.Ott, J.H.B. Kemperman and M.Mathis (1996)
• M.Mathis, J.Semke, J.Mahdavi, T.Ott (1997)
• T.V.Lakshman and U.Mahdow (1997)
• J.Padhye, V.Firoiu, D.Towsley, J.Kurose (1998)

46
Implications

• Applicability
• Additive increase, multiplicative decrease (Reno)
• Congestion avoidance dominates
• No timeouts, e.g., SACKRH
• Small losses
• Persistent, greedy sources
• Receiver not bottleneck
• Implications
• Reno equalizes window
• Reno discriminates against long connections
• Halving throughput gt quadrupling loss rate!

47
Refinement (Padhye, Firoin, Towsley Kurose 1998)

• Renewal model including
• FR/FR with Delayed ACKs (b packets per ACK)
• Timeouts
• Receiver window limitation
• Source rate
• When p is small and Wr is large, reduces to

48
TCP Friendliness

• What does it mean to be TCP friendly?
• TCP is not going away
• Fairness means equal shares
• Any new congestion control must compete with TCP
  flows
• Should not clobber TCP flows and grab bulk of
  link
• Should also be able to hold its own, i.e. grab
  its fair share, or it will never become popular

49
Binomial Congestion Control

• In AIMD
• Increase Wn1 Wn ?
• Decrease Wn1 (1- ?) Wn
• In Binomial
• Increase Wn1 Wn (? / Wnk)
• Decrease Wn1 Wn - ? Wnl
• k0 l1 ? AIMD
• l lt 1 results in less than multiplicative
  decrease
• Good for multimedia applications

50
Binomial Congestion Control continued

• Rate 1/ (loss rate)1/(kl1)
• If kl1 ? rate 1/p0.5
• TCP friendly
• AIMD (k0, l1) is the most aggressive of this
  class
• SQRT (k1/2,l1/2) and IIAD (k1,l0)
• Good for applications that want to probe quickly
  and can use any available bandwidth

51
Static Optimization Framework
pl(t)
xi(t)
Feedback?

• Duality theory ? equilibrium
• Source rates xi(t) are primal variables (i is a
  flow)
• Congestion measures pl(t) are dual variables,
  where pl(t) prob of loss on link l at time
  t
• Congestion control is optimization process over
  Internet

52
Overview Equilibrium

• Interaction of source rates xs(t) and congestion
  measures pl(t)
• Duality theory
• They are primal and dual variables
• Flow control is optimization process
• Example congestion measure
• Loss (Reno)
• Queueing delay (Vegas)

53
Overview Equilibrium continued

• Congestion control problem

• Primal-dual algorithm

• TCP/AQM protocols (F, G)
• Maximize aggregate source utility
• With different utility functions Us(xs)

54
Model

• Sources s
• L(s) - links used by source s
• Us(xs) - utility if source rate xs
• Network
• Link l has capacity cl

55
Primal Problem

• Assumptions
• Strictly concave increasing Us
• Unique optimal rates xs exist
• Direct solution impractical

56
Duality Approach
57
Gradient Algorithm

• Gradient algorithm

Theorem (Low, Lapsley, 1999) Converges to
optimal rates in an asynchronous environment
58
Example
59
Example continued

• xs proportionally fair (Vegas)
• pl Lagrange multiplier, (shadow) price,
  congestion measure
• How to compute (x, p)?
• Gradient algorithms, Newton alg, Primal-dual alg,
  
• Relevance to TCP/AQM ??
• TCP/AQM protocols implement primal-dual
  algorithms over Internet

60
Example continued

• xs proportionally fair (Vegas)
• pl Lagrange multiplier, (shadow) price,
  congestion measure
• How to compute (x, p)?
• Gradient algorithms, Newton alg, Primal-dual alg,
  
• Relevance to TCP/AQM ??
• TCP/AQM protocols implement primal-dual
  algorithms over Internet

61
Active Queue Management (AQM)

• Idea provide congestion information by
  probabilistically marking packets
• Issues
• How to measure congestion (p and G)?
• How to embed congestion measure?
• How to feed back congestion info?

62
RED (Floyd Jacobson 1993)

• Congestion measure average queue length
• pl(t1) pl(t) xl(t) - cl
• Embedding p-linear probability function

marking
1
Avg queue
63
REM (Athuraliya Low 2000)

• Congestion measure price
• pl(t1) pl(t) g(al bl(t) xl (t) - cl
  )
• Embedding exponential probability function

64
Key Features

• Clear buffer and match rate

Theorem (Paganini 2000) Global asymptotic
stability for general utility function (in the
absence of delay)
65
AQM Summary
66
Reno F(p(t), x(t))
for every ack (ca) W 1/W for every
loss W W/2


Primal-dual algorithm
x(t1) F( p(t), x(t) ) p(t1) G( p(t),
x(t) )
67
Reno Implications

• Equilibrium characterization
• Duality
• Congestion measure p loss
• Implications
• Reno equalizes window wi ti xi
• inversely proportional to delay ti
• dependence for small p
• DropTail fills queue, regardless of queue
  capacity

68
Reno Gradient Algorithm

• Gradient algorithm

• TCP approximate version of gradient algorithm

69
Reno Gradient Algorithm

• Gradient algorithm

• TCP approximate version of gradient algorithm

70
Vegas
for every RTT if W/RTTmin W/RTT lt a then
W if W/RTTmin W/RTT gt a then W --
for every loss W W/2
queue size
71
ATM ABR Explicit Rate Feedback
RM Cell
Source
Destination

• Sources regulate transmission using a rate
  parameter
• Feedback scheme
• Every (n1)th cell is an RM (control) cell
  containing current cell rate, allowed cell rate,
  etc
• Switches adjust the rate using rich information
  about congestion to calculate explicit, multi-bit
  feedback
• Destination returns the RM cell to the source
• Control policy Sources adjust to the new rate

72
ERICA Design Goals

• Allows utilization to be 100 (better tracking)
• Allows operation at any point between knee and
  cliff
• The queue length can be set to any desired value
  (tracking)
• Max-min fairness (fairness)

100
Link Utilization
Throughput
Time
Load
QueueLength
Delay
50
Load
73
Efficiency vs. Fairness OSU Scheme

• Efficiency high utilization
• Fairness Equal allocations for contending
  sources
• Worry about fairness after utilization close to
  100 utilization . Target Utilization (U) and
  Target Utilization Band (TUB).

overload region
worry about fairness here
999591
TotalLoad
TUB
U
underload region
Time
74
ERICA Switch Algorithm

• Overload Input rate/Target rate
• Fair Share Target rate/ of active VCs
• This VCs Share VCs rate /Overload
• ER Max(Fair Share, This VCs Share)
• ER in Cell Min(ER in Cell, ER)
• This is the basic algorithm.
• Has more steps for improved fairness, queue
  management, transient spike suppression,
  averaging of metrics.

75
TCP Rate Control

• Step 1 Explicit control of window

Congestion window (CWND)
W
Actual Window Min(Cwnd, Wr)

• Step 2 Control rate of acks (ack-bucket)
  Tradeoff ack queues in reverse path for fewer
  packets in forward path

r
pkts
W
acks
R
Time
76
Summary

• Active Queue Management (AQM) RED, REM, etc
• Alternative models
• Accumulation-based schemes Monaco, Vegas
• Explicit Rate-based Schemes
• TCP stochastic modeling
• Static (Duality) Optimization Framework
• Can manage TCP with very little loss