Internet Congestion Control with Active Queue Management (AQM) Seungwan Ryu September 4, 2001

advertisement
Internet Congestion Control with
Active Queue Management (AQM)
September 4, 2001
Seungwan Ryu
(sryu@eng.buffalo.edu)
PhD Student of IE Department
University at Buffalo
Contents




Internet Congestion Control
Mathematical Modeling and Analysis
Adaptive AQM and User Response
Further studies
2
I. Internet Congestion Control







Internet Traffic Engineering
What is Congestion ?
Congestion Control and Avoidance
Implicit vs. Explicit feedback
TCP Congestion Control
Active Queue management (AQM)
Explicit Congestion Notification (ECN)
3
Internet Traffic Engineering



Measurement: for reality check
Experiment: for Implementation Issues
Analysis:



Bring fundamental understanding of systems
May loose important facts because of simplification
Simulation:


Complementary to analysis: Correctness, exploring
complicate model
May share similar model to analysis
4
What is congestion ?

What is congestion ?


The aggregate demand for bandwidth exceeds the
available capacity of a link.
What will be occur ?

Performance Degradation
•
•
•
•
Multiple packet losses
Low link utilization (low Throughput)
High queueing delay
Congestion collapse
5
What is congestion ? - 2
Congestion Control
Open-loop control
 Mainly used in circuit
switched network (GMPLS)
Implicit feedback control
 End-to-end congestion control
 Examples:
TCP Tahoe, TCP Reno, TCP Vegas, etc.
Closed-loop control
 Mainly used in packet switched network
 Use feedback information: global & local
Explicit feedback control
 Network-assisted congestion control
 Examples:
IBM SNA, DECbit, ATM ABR, ICMP source
quench, RED, ECN
6
Congestion Control and Avoidance

Two approaches of handling Congestion

Congestion Control (Reactive)
• Play after the network is overloaded

Congestion Avoidance (Proactive)
• Play before the network becomes overloaded
7
Implicit vs. Explicit feedback

Implicit feedback Congestion Control


Network drops packets when congestion occur
Source infers congestion implicitly
• time-out, duplicated ACKs, etc.


Example: end-to-end TCP congestion Control
Simple to implement but inaccurate
• implemented only at transport layer (e.g., TCP)
8
Implicit vs. Explicit feedback - 2

Explicit feedback Congestion Control

Network component (e.g., router) provides
congestion indication explicitly to sources
• use packet marking, or RM cells (in ATM ABR control)



Examples: DECbit, ECN, ATM ABR CC, etc.
Provide more accurate information to sources
But is more complicate to implement
• Need to change both source and network algorithm
• Need cooperation between sources and network
component
9
TCP Congestion Control

Uses end-to-end congestion control

uses implicit feedback
• e.g., time-out, triple duplicated ACKs, etc.

uses window based flow control
• cwnd = min (pipe size, rwnd)
• self-clocking
• slow-start and congestion avoidance

Examples:
• TCP Tahoe, TCP Reno, TCP Vegas, etc.
10
TCP Congestion Control - 2

cwnd
Slow-start and Congestion Avoidance
Slow Start
W*
Congestion Avoidance
W
4
W+1
W*/2
2
1
RTT
RTT
Time
11
TCP Congestion Control - 3

TCP Tahoe


Use slow start/congestion avoidance
Fast retransmit: an enhancement



detect packet (segments) drop by three duplicate ACKs
W = W/2, and enter congestion avoidance
TCP Reno (fast recovery)

Upon receiving three duplicate ACKs




ssthresh = W/2, and retransmit missing packets
W = ssthresh +3
Upon receiving next ACK: W = ssthresh
Allow the window size grow fast to keep the pipeline full
12
TCP Congestion Control - 3

TCP SACK (Selected Acknowledgement)


TCP (Thaoe) sender can only know about a single lost per
RTT
SACK option provides better recovery from multiple losses



The sender can transmit all lost packets
But those packets may have already been received
Operation



Add SACK option into TCP header
The receiver sends back SACK to sender to inform the
reception of the packet
Then, the sender can retransmit only the missing packet
13
Active Queue Management (AQM) - 1

Performance Degradation in current TCP
Congestion Control




Multiple packet loss
Low link utilization
Congestion collapse
The role of the router becomes important


Control congestion effectively in networks
Allocate bandwidth fairly
14
AQM - 2

Problems with current router algorithm


Use FIFO based tail-drop (TD) queue management
Two drawbacks with TD: lock-out, full-queue



Lock-out: a small number of flows monopolize usage of buffer capacity
Full-queue: The buffer is always full (high queueing delay)
Possible solution: AQM

Definition:
A group of FIFO based queue management
mechanisms to support end-to-end congestion
control in the Internet
15
AQM - 3

Goals of AQM

Reducing the average queue length:


Reducing packet losses:


More efficient resource allocation
Methods:



Decreasing end-to-end delay
Drop packets before buffer becomes full
Use (exponentially weighted) average queue
length as an congestion indicator
Examples: RED, BLUE, ARED, SRED, FRED,….
16
AQM - 4

Random Early Detection (RED)


use network algorithm to detect incipient
congestion
Design goals:
•
•
•
•

minimize packet loss and queueing delay
avoid global synchronization
maintain high link utilization
removing bias against bursty source
Achieve goals by
• randomized packet drop
• queue length averaging
17
RED
avgQ  (1  WQ )avgQ  WQ Q
P
0


avgQ  min th
Pd   pmax
max th  min th

1

avgQ  min th
min th  avgQ  max th
max th  avgQ
1
maxp
minth
maxth
K
18
AQM - 5 : BLUE

Algorithm

Concept



To avoid drawbacks of RED

Upon packet loss
 if (now - last_update >freeze_t)
 Pm = pm + d1
 last_update = now
upon link idle
 if (now - last_update >freeze_t)
 Pm = pm - d2
 last_update = now





Parameter tuning problem
Actual queue length fluctuation
Decouple congestion control from
queue length
Use only loss and idle event as an
indicator
Maintains a single drop prob., pm
Drawback

Can not avoid some degree of
multiple packet loss and/or low
utilization
19
AQM - 6 : SRED

Algorithm







ith arriving packet is compared with a
randomly selected one from Zombie list
Hit = 1, if they are from same flow
= 0, if NOT

p(i)=hit frequency=(1-)p(i-1)+Hit
p(i)-1: estimator of # of active flows
Packet drop probability
Concept



Drawbacks


Pzap
1
 Psred * min( 1,
)
(256  P(i)) 2


 (1 / 4) pmax

0

pmax
psred
(1 / 3) B  q  B
(1 / 6) B  q  (1 / 3) B
q  (1 / 6) B
stabilize queue occupancy
use actual queue length
Penalize misbehaving flows

P(i)-1 is not a good estimator
for heterogeneous traffic
Parameter tuning problem:
Psred, Pzap, etc.
Stabilize queue occupancy
when traffic load is high.
(When load is low ?)
20
AQM - 7 : ARED

Adapt aggresiveness of RED according to the
traffic load change


adapt maxp based on queue behavior
Operation



Increase maxp when avgQ crosses above maxth
Decrease maxp when avgQ crosses below minth
freeze maxp after changing to prevent oscillation
21
AQM - 8

Problems with existing AQM Proposals




Mismatch between macroscopic and microscopic
behavior of queue length
Insensitivity to the change of input traffic load
Configuration (parameter setting) problem
Reasons:



Queue length averaging
use inappropriate congestion indicator
Use inappropriate control function
22
Explicit Congestion Notification (ECN)

Current congestion indication



Use packet drop to indicate congestion
source infer congestion implicitly
ECN




to give less packet drop and better performance
use packet marking rather than drop
need cooperation between sources and network
need two bits in IP header: ECT-bit, CE-bit
23
ECN - 2
ECT
IP Header
CE
1
ECT
0
CE
1
1
1
TCP Header
0
0
CWR
CWR
2
1
ACK TCP
Header
ECN-Echo
TCP Header
3
1
CWR
Source
4
Router
Destination
24
Contents




Internet Congestion Control
Mathematical Modeling and Analysis
Adaptive AQM and User Response
Further Studies
25
II. Mathematical Modeling and
Analysis

An Overview

Mathematical Modeling of AQM



Window based packet switching and the Internet
Mathematical modeling and analysis of AQM
Problems with existing AQMs


Problems with existing AQMs
Adaptive congestion indicator and control function
26
Overview - 1

Goal of mathematical modeling




See steady state system dynamics
Capture main factors influence to performance
Provide recommendations for design and operation
Two approaches for TCP Congestion Control

Modeling steady state TCP behaviors
• the square root law*, PFTK [Padhye et al., 1998]
• assume TD queue management at the router

Mathematical modeling and analysis of AQM (RED)
c
*: T  RTT p , T: Throughput, p: constant drop rate
27
Overview - 2

AQM modeling and analysis




Analytic modeling and analysis
Control Theoretic Analysis
Window based modeling and Analysis
Assumptions



Poisson assumption for input traffic
Fixed number of persistent TCP traffics
Steady state window size saturation
28
Mathematical Modeling of AQM - 1

Window based packet switching Model (Yang 99)


Determine the steady state window size, Ws, of each
flow sS
If link j is not congested
nsj  0, Q j  0 s  S ( j ), j
C j  s  s

j
If link j is congested
nsj  0, Q j  0 s  S ( j ), j
C j  s  s
j
29
Mathematical Modeling of AQM - 2

Window equation for an individual flow

Since
nsj
Qj

s
Cj
Q j  0
     (1)


Qj
Ws  s Rs   jJ ( s ) nsj   s   j
 Rs  s  S ( j )


Cj



Limitation of this model

Assume infinite buffer size
• No buffer overflow
• No packet drop
• No queue management algorithm at routers
30
Mathematical Modeling of AQM - 3
A simple AQM model
Sources
s1
1
S2
AQM Router
2
Bottleneck
Link
Destination
C
S
K
Min_th
SS
31
Mathematical Modeling of AQM - 4

Extend Yang’s Model to AQM model



Finite buffer capacity K
The router use AQM to control congestion
When congested
• Yang’s Model:
• Our Model:

s

s
s
s
 C,
s  s
 C , s  s (1  pd )
32
Mathematical Modeling of AQM - 5

Case 1: Tail drop

Packet drop probability Pd:
1
pd  
0
if
o.w.
C
and
QK
33
Mathematical Modeling of AQM - 6

Case 2: AQM



Q  Q  min th  s ns
Let
Then since
Q
W   (1  pd )( R  )
C
Packet drop prob. Pd:
W

1


pd    ( R  Q )
C

0

if
  C , Q  min th
O.w.
34
Mathematical Modeling of AQM - 7

Congestion Indicator


Input traffic load  should be the congestion
Indicator
Current AQMs
• Use queue length Q as an alternative
• Assume that the input traffic load  is fixed in equilibrium

Reason
• can not measure(or estimate)  exactly for on line
implementation of packet drop function
35
Mathematical Modeling of AQM - 8

Packet drop function
p d  f ( )

Reason
• The traffic load

fluctuate, NOT stay in equilibrium
• queue length is a function of input traffic

Alternatively:
pd  f ( , Q)
36
Problems with existing AQMs

Mismatch between macroscopic and
microscopic behavior of queue length

Insensitivity to the input traffic load variation

parameter configuration problem
37
Problems with existing AQMs - 2

Mismatch problem
Internet Traffic Generation
40
35
25
20
15
10
5
time
31
28
25
22
19
16
13
10
7
4
0
1
Window size
30
38
Problems with existing AQMs - 3
Mismatch between macroscopic and
microscopic behavior of queue length
Rho
Queue Length

Actual
Wq=0.02
Wq=0.1
25

20
2.0
15
1.5
10
1.0
5
0.5
0
0
1
6
11
16
Time
21
26
31
39
Problems with existing AQMs - 4

Insensitivity to the input traffic load variation
: u=0.7
: u=0.45
: u=0.25
: RED
: GRED
: Scheme III
1.00
Packet drop rate
0.80
0.60
0.40
0.20
0.00
0.3
0.5
0.7
0.9
1.1
1.3
Traffic Inte nsitie s (loads)

Schemes: I:RED, II:GRED, III: pd  f ( , Q)
40
Problems with existing AQMs - 5

Parameter configuration problem


Has been a main design issue since 1993
Many modified AQMs has been proposed
• Verified with simple simulation or simple experiment
• good for particular traffic conditions
• Real traffic is totally different.

Need adaptive congestion indicator and control
function
• Adaptive to input traffic load variation
• Avoid congestion NOT based on current state (i,e,. Q)
41
Contents




Internet Congestion Control
Mathematical Modeling and Analysis
Adaptive AQM and User Response
Further Studies
42
III. Adaptive AQM and User Response




Input traffic load Prediction
Adaptive AQM algorithms
Adaptive parameter configuration
Adaptive User response algorithm
43
Input traffic load Prediction

Consider time-slotted model



Time is divided into unit time slots, t, t=0,1,…
calculate parameters at the end of each slot
estimate Qt+1 to detect congestion proactively
Qt 1  (t 1  C )  Qt
• Predict ˆt  1 from measured input traffic t-1, t of past
two time slots
• Then, predict Q̂t  1 of next time slot t
44
Adaptive AQM algorithms

Algorithm I: E-RED and E-GRED

Enhanced-RED

0,

Q̂  minth

p   max p t  1
maxth  minth


1

Q̂t  1  minth
minth  Q̂t  1  maxth
maxth  Q̂t  1
E-GRED: similar to E-RED
45
Adaptive AQM algorithms - 2

Algorithm II:

Use both predicted traffic intensity ˆ t  1 and current
buffer utilization t=Qt/K
 ˆ
t  1 represents imminent traffic changes in near future


t represents current status of traffic
Possible algorithms:
 1 ˆ t  1t ,
 2 ˆ t2 1t ,
3 2
ˆ t  1
t
46
Adaptive AQM algorithms - 3

Example:

maintain Qindex to impose appropriate drop rate
adaptively to traffic load change

Then,
 pd * (1  Qindex ), arriving packets
Qt
Pd  
, where Qindex 
existing packets
Qt  Qˆ t 1
 pd * Qindex ,
• If t is low and ˆ t  1 is high: more penalty to incoming
packets
ˆ t  1is low: more penalty on existing
• If t is high and
packets
• Only High penalty for both packets when t and ˆ t  1 are
high
47
Adaptive AQM algorithms - 4

Algorithm III: E-BLUE

BLUE Algorithm
• uses packet drops and link idle for adjusting packet drop
probability
• Can not avoid some degree of performance degradation

Enhancement
• Use Virtual lower/upper bound (VL, VU)
• Combine predicted queue length Q̂t  1 with BLUE
48
Adaptive parameter configuration

Adaptive queue length sampling interval t

Previous recommendations
• In [Firoiu et al.], minimum RTT was recommended
• In [Hollot et al.], static and link speed independent value
was recommended
• However, above recommendations were obtained from
assumptions of persistent and fixed N TCP traffics

Our recommendation
• The amount of incoming traffic fluctuate with time
• Adjust t according to the varying traffic situation
(i.e., adjust t according to the amount of input traffic)
49
Adaptive parameter configuration - 2
Q
(i-1)
i
(i+1)
(i+2)
Time
50
Adaptive parameter configuration - 3

Adaptive filtering weight wq


In RED, wq was recommended with 0.002 for long-term
(macroscopic) performance goal
Fixed small value of wq shows problems
• Parameter setting problem
• Insensitivity of control function to the change of traffic
• Fairness problem: impose penalty to innocent packets


Need to have adaptive wq to the change of traffic load
One possible method:
• Set wq as a function of current queue utilization,
e.g., wq =  Qt/C , 0 <  < 1
51
Adaptive User response algorithm


AQM need work with intelligent source
response for better performance
Enhanced-ECN

If receive ECN feedback in (t-1)
• If No ECN feedback in t



If received ACK > 0 , W= W+M/W + M
Else , W= W+M/W
• Else, Continue usual response to ECN feedback
Else, Continue TCP Congestion Avoidance
52
Contents




Internet Congestion Control
Mathematical Modeling and Analysis
Adaptive AQM and User Response
Further Studies
53
IV. Further Studies

Mathematical Modeling and Analysis




Simulation studies




Stability and Control Dynamics
Alternative Modeling
Control Theoretic Consideration
Traffics
Performance Metrics
Other approaches of congestion control
More about AQM
54
Mathematical Modeling and Analysis

Since p=f(,q) ,

Then find equilibrium point (*,p*)
p
=g(p)
T (  , q )     ( 1  p ) 
C( 1  p )
R p
P=f()
(*,p*)

55
Mathematical Modeling and Analysis - 2

Alternative Modeling:

State dependent service M/M/1/K queueing model


0
1
C




L-1
C
L
C


L-1
(C+p1)

K-1
(C+pK’-1)
K
C+
L=minth, K’=K-minth
56
Mathematical Modeling and Analysis - 3

Service rates

Steady state probabilities
Qi  min th
 C,

S  C  pi, min th  Qi  K
 C  ,
K  Qi




( ) i 0 ,
i  min th

C

 i  min th
  min th


i  
0 ,
min th  i  K
 j1
 ( C )
(

C

p

)

i 

1
min th 
K





  ( )i   ij1min th
( ) min th  ,
i0


C
(

C

p

)

C
 i 1
i  min th 1 
i 

57
Mathematical Modeling and Analysis - 3

Control Theoretic Consideration
Control
Function
S
t
Queue
dynamics
Router
Buffer
t(1-p)
D
ACK (or NACK)
58
Simulation study

Goal of simulation study



See dynamics and performance of our AQM
Compare results with other AQM such as RED
Use realistic traffic


previous studies has been done with simple and
unreal traffic (fixed number of persistent TCPs)
Generate realistic Internet traffic
• Long-lived (FTP) and short-lived (web-like) TCP traffic
• UDP traffic: CBR and/or ON/OFF
59
Performance Metrics

TCP traffics

Network-centric: for aggregate traffic
• Throughput (or goodput)
• Packet dropping (marking) probability
• Link utilization (or queueing delay)

User-centric: for Individual traffic
• goodput (or throughput)
• mean response time (RTT)

UDP traffic
• individual packet drop probability and its distribution
60
Other approaches of CC - 1: Pricing

Smart-market [Mackie-Mason 1995]




A price is set for each packet depends on the level of
demand for mandwidth
Admit packets with bid prices that exceed the cut-off value
The cut-off is determined by the marginal cost
Paris metro pricing (PMP) [Odlyzko]



To provide differentiated services
The network is partitioned into several logical separate
channels with different prices
With less traffic in channel with high price, better QoS would
be provided.
61
Other approaches - 2: Optimization

Concept

Network resource
allocation problem:





User problems
Network problems
User problem sends

bandwidth request with
pricie
Network problem allocate
bandwidth to each users
by solving NLP
User problem


Users can be distinguished by a
utility function
A user wants to maximize its
benefit (utility - cost)
Network problem


maximize aggregate utilities
subject to the link capacity
constraints
Then, it can be formulated to a
Non-linear programming (NLP)
problem
62
Other approaches - 3: Fairness

Two fairness issues



Fair bandwidth sharing: network-centric
Fair packet drop (mark): user-centric
Fair bandwidth sharing

Max-min fair [Bertsekas, 1992]:



No rate can be increased without simultaneous decreasing other
rate which is already small
provides equal treatment to all flows
Proportional fair [Kelly 1998]


A feasible set of rates are non-negative and the aggregate rate is
not greater than link capacity and the aggregate of proportional
change is zero or negative
provides different treatment of each flow according to their rates
63
More about AQM

Responsive (TCP) vs. unresponsive flows (UDP)

RED fail to regulate unresponsive flows



UDP do not adjust sending rate upon receiving congestion signal
UDP flows consumes more bandwidth than fair share
FRED [Lin & Morris, 1997]
 Tracks the # of packets in the queue from each flow



Fair share for a flow is calculated dynamically
unresponsive flows are identified and penalized


maintain logical queues for each active flows in a FIFO queue
Drop packets proportional to bandwidth usage
See TCP-friendly website
(http://www.psc.edu/networking/tcp_friendly.html)
64
More about AQM - 2

Providing QoS and DiffServ with AQM




Try to support a multitude of transport protocol (TCP,
UDP, etc.)
Classify several types of services rather than one besteffort service.
Then, apply different AQM control to each services
classes.
Examples:


RIO (RED In and Out) [Clark98]
CBT (Class based Thresholds) [Floyd1995]
65
More about AQM - 3

RIO (RED in and out) [Clark 1998]


Separate flows into two classes: IN and OUT service profile
router maintains two different statistics for each service profiles.



Different parameters and average queue lengths
Avgs: for IN packet: avgIN, for OUT profile: avgTOTAL
When congested, apply different control to each classes
p
1
Pmax_OUT
Pmax_IN
Minth_OUT Maxth_OUT
= Minth_IN
Maxth_IN
avg
66
More about AQM - 4

CBT [Floyd 1995]





packets are classified into
several classes
maintain a single queue
but allocate fraction of
capacity to each class
Apply AQM (RED) based

control to each class
Once a class occupies its
capacity, discard all
arriving packets
Drawbacks

Fairness problem in case of
changing traffic mix


static threshold setting
Total utilization can be fluctuated
Dynamic-CBT [Chung2000]


Track the number of active flows
of each class
dynamically adjust threshold
values of each class
67
More about AQM - 5

Other Issues


AQM vs. Tail Drop(TD)
Congestion Indicator:


Parameter tuning problem:



wq, maxp, static or dynamic sampling
Alternative ways: virtual queue approach


Average queue length vs. Instantaneous queue length
EX: [Gibbens 1998], [Kuniyur2000]
Performance with/without ECN mechanism
Control objective” Router-centric vs. user-centric
68
References






S. Floyd et al. “Random early detection gateways for congestion
avoidance control.” IEEE/ACM TON, 1993.
RED web page, http://www.aciri.org/floyd/red.html
RED for dummies, http://www.magma.ca/~terrim/RedLit.htm
S. Ryu et al. “Advances in Internet congestion control.”
submitted to IEEE comm. Survey & Tutorial, 2001
B. Braden et al. “Recommendations on queue management and
congestion avoidance in the Internet.” IETF RFC2309, 1998.
K. Ramakrishinan et al. “A proposal to add explicit congestion
notification (ECN) to IP.” IETF RFC2481, 1999.
69
Download