CSIT560 by M. Hamdi 1 QoS in The Internet: Scheduling Algorithms and Active Queue Management.

1CSIT560 by M. Hamdi

QoS in The Internet: Scheduling Algorithms and Active

Queue Management


Principles for QOS Guarantees• Consider a phone application at 1Mbps and an FTP application sharing

a 1.5 Mbps link. – bursts of FTP can congest the router and cause audio packets to be dropped.

– want to give priority to audio over FTP

• PRINCIPLE 1: Marking of packets is needed for router to distinguish between different classes; and new router policy to treat packets accordingly


Principles for QOS Guarantees (more)

• Applications misbehave (audio sends packets at a rate higher than 1Mbps assumed above);

• PRINCIPLE 2: provide protection (isolation) for one class from other classes (Fairness)


Ba

nd

wid

thB

an

dw

idth

Ba

nd

wid

thB

an

dw

idth

DelayDelayDelayDelay

The path as perceived by a packet!The path as perceived by a packet!AA BB

QoS MetricsWhat are we trying to control?

• Four metrics are used to describe a packet’s transmission through a network – Bandwidth, Delay, Jitter, and Loss

• Using a pipe analogy, then for each packet: Bandwidth is the perceived width of the pipe

Delay is the perceived length of the pipe

Jitter is the perceived variation in the length of the pipe

Loss is the perceived leakiness if the pipe


Internet QoS Overview• Integrated services

• Differentiated Services

• MPLS

• Traffic Engineering


QoS: State information

• No State Vs. Soft State Vs. Hard State

No State

IPCircuit

SwitchedATMIntserv/RSVPdiffserv

DedicatedCircuit

HardState

SoftState

No State inside the networkFlow information at the edges

Packet Switched


QoS Router

Policer

Policer

Classifier

Policer

Policer

Classifier

Per-flow Queue

Scheduler

Per-flow Queue

Per-flow Queue

Scheduler

Per-flow Queue

shaper

shaper

Queue management


Queuing DisciplinesFirst come first serve

Class 1

Class 2

Class 3

Class 4

Class based scheduling

Scheduler

flow 1

flow 2

flow n

Classifier

Buffer management


DiffServDiffServ Domain

PremiumPremium GoldGold SilverSilver BronzeBronze

PHBLLQ/WRED

Classification / Conditioning


Functionality at DiffServ Routers


Differentiated Service (DS) Field

Version HLen TOS Length

Identification Fragment offsetFlags

Source address

Destination address

TTL Protocol Header checksum

0 4 8 16 19 31

Data

IPheader

• DS filed reuse the first 6 bits from the former Type of Service (TOS) byte to determine the PHB

DS Field0 5 6 7


R2AR1

R3

B

R4

A RESVA RESV

A RESV

A RESV

Integrated Services RSVP and Traffic Flow example

PATH message will leave the IP address of the previous hop node in each router. Contains Sender Tspec, Sender Temp, Adspec.

Admission/policy control determines if the node has sufficient available resources to handle the request. If request is granted, bandwidth and buffer is allocated.

A RESV message containing a flowspec and a filterspec must be sent in the exact reverse path.The flowspec (T-spec/R-spec) defines the QoS and the traffic characteristics being requested.

Reserved buffer and bw



BPATH

Data B

Data

RSVP maintains soft state information (DstAddr, Protocol, DstPort) in the routers. All packets will get MF classification treatment and put in the appropriate queue.

Routers enforce MF classification and put packets in the appropriate queue.The scheduler will then serve these queues.

Phop = A

Phop = R1

Phop = R2BPATH

BPATH

BPATH


IntServ: Per-flow classification

SenderReceiver


Per-flow buffer management

SenderReceiver


Per-flow scheduling

SenderReceiver


Round Robin (RR)

• RR avoids starvation

• All sessions have the same weight and the same packet length:

A: B: C:

Round #2

…

Round #1


RR with variable packet length

A: B: C:

Round #1 Round #2

…

But the Weights are equal !!!


Solution…

A: B: C:

#1 #2 #3

…

#4


Weighted Round Robin (WRR)

WA=3 WB=1 WC=4

#1

round length = 8

…

#2


WRR – non Integer weights

WA=1.4 WB=0.2 WC=0.8

WA=7 WB=1 WC=4

Normalize

round length = 13

…


Weighted round robin

• Serve a packet from each non-empty queue in turn– Can provide protection against starvation

– It is easy to implement in hardware

• Unfair if packets are of different length or weights are not equal

• What is the Solution?

• Different weights, fixed packet size– serve more than one packet per visit, after normalizing

to obtain integer weights


Problems with Weighted Round Robin

• Different weights, variable size packets– normalize weights by mean packet size

• e.g. weights {0.5, 0.75, 1.0}, mean packet sizes {50, 500, 1500}

• normalize weights: {0.5/50, 0.75/500, 1.0/1500} = { 0.01, 0.0015, 0.000666}, normalize again {60, 9, 4}

• With variable size packets, need to know mean packet size in advance

• Fairness is only provided at time scales larger than the schedule


Max-Min Fairness

• An allocation is fair if it satisfies max-min fairness– each connection gets no more than what it wants

– the excess, if any, is equally shared

Transfer half of excess

Unsatisfied demand

Transfer half of excess

Unsatisfied demand


Max-Min FairnessA common way to allocate flows

N flows share a link of rate C. Flow f wishes to send at rate W(f), and is allocated rate R(f).

1. Pick the flow, f, with the smallest requested rate.

2. If W(f) < C/N, then set R(f) = W(f).

3. If W(f) > C/N, then set R(f) = C/N.

4. Set N = N – 1. C = C – R(f).

5. If N>0 goto 1.


1W(f1) = 0.1

W(f3) = 10R1

C

W(f4) = 5

W(f2) = 0.5

Max-Min FairnessAn example

Round 1: Set R(f1) = 0.1

Round 2: Set R(f2) = 0.9/3 = 0.3

Round 3: Set R(f4) = 0.6/2 = 0.3

Round 4: Set R(f3) = 0.3/1 = 0.3


Fair Queueing

1. Packets belonging to a flow are placed in a FIFO. This is called “per-flow queueing”.

2. FIFOs are scheduled one bit at a time, in a round-robin fashion.

3. This is called Bit-by-Bit Fair Queueing.

Flow 1

Flow NClassification Scheduling

Bit-by-bit round robin


Weighted Bit-by-Bit Fair Queueing

Likewise, flows can be allocated different rates by servicing a different number of bits for each flow

during each round.

1R(f1) = 0.1

R(f3) = 0.3R1

C

R(f4) = 0.3

R(f2) = 0.3

Order of service for the four queues:… f1, f2, f2, f2, f3, f3, f3, f4, f4, f4, f1,…

Also called “Generalized Processor Sharing (GPS)”


Understanding bit by bit WFQ 4 queues, sharing 4 bits/sec of bandwidth, Weights 3:2:2:1

Weights : 3:2:2:1

3

2

2

1

6 5 4 3 2 1 0

B1 = 3

A1 = 4

D2 = 2 D1 = 1

C2 = 1 C1 = 1

Time

3

2

2

1

6 5 4 3 2 1 0

B1 = 3

A1 = 4

D2 = 2 D1 = 1

C2 = 1 C1 = 1

A1A1A1B1

A2 = 2

C3 = 2

Time

Weights : 3:2:2:1

Round 1

3

2

2

1

6 5 4 3 2 1 0

B1 = 3

A1 = 4

D2 = 2 D1 = 1

C2 = 1 C1 = 1

A1A1A1B1

A2 = 2

C3 = 2

D1, C2, C1 Depart at R=1Time

B1C1C2D1

Weights : 3:2:2:1

Round 1


Understanding bit by bit WFQ 4 queues, sharing 4 bits/sec of bandwidth, Weights 3:2:2:1

Weights : 3:2:2:1

3

2

2

1

6 5 4 3 2 1 0

B1 = 3

A1 = 4

D2 = 2 D1 = 1

C2 = 1 C1 = 1

A2 = 2

C3 = 2

B1, A2 A1 Depart at R=2Time

A1A1A1B1B1C1C2D1A1A2A2B1

Round 1Round 2

Weights : 3:2:2:1

3

2

2

1

6 5 4 3 2 1 0

B1 = 3

A1 = 4

D2 = 2 D1 = 1

C2 = 1 C1 = 1

A2 = 2

C3 = 2

D2, C3 Depart at R=2Time

A1A1A1B1B1C1C2D1A1A2A2B1C3C3D2D2

Round 1Round 23

Weights : 1:1:1:1

Weights : 3:2:2:1

3

2

2

1

6 5 4 3 2 1 0

B1 = 3

A1 = 4

D2 = 2 D1 = 1

C2 = 1C3 = 2 C1 = 1

C1C2D1A1A1A1A1A2A2B1B 1B1

A2 = 2

C3C3D2D2

Departure order for packet by packet WFQ: Sort by finish time of packetsTime

Sort packets


Packetized Weighted Fair Queueing (WFQ)

Problem: We need to serve a whole packet at a time.

Solution:

1. Determine what time a packet, p, would complete if we served it flows bit-by-bit. Call this the packet’s finishing time, Fp.

2. Serve packets in the order of increasing finishing time.

Also called “Packetized Generalized Processor Sharing (PGPS)”


WFQ is complex

• There may be hundreds to millions of flows; the linecard needs to manage a FIFO queue per each flow.

• The finishing time must be calculated for each arriving packet,

• Packets must be sorted by their departure time.

• Most efforts in QoS scheduling algorithms is to come up with practical algorithms that can approximate WFQ!

1

2

3

N

Packets arriving to egress linecard

CalculateFp

Find Smallest Fp

Departing packet

Egress linecard


When can we Guarantee Delays?

• Theorem

If flows are leaky bucket constrained and all nodes employ GPS (WFQ), then the network can guarantee worst-case delay bounds to sessions.


Traffic Managers: Active Queue

Management Algorithms


Queuing Disciplines

• Each router must implement some queuing discipline

• Queuing allocates both bandwidth and buffer space:– Bandwidth: which packet to serve (transmit) next - This

is scheduling

– Buffer space: which packet to drop next (when required) – this buffer management

• Queuing affects the delay of a packet (QoS)


Queuing Disciplines

Traffic Sources

Class C

Class B

Class A

Traffic Classes

DropScheduling Buffer Management


Active Queue Management

Queue

SinkOutbound LinkRouterInbound Link

Sink

TCP

TCP

ACK…

ACK…

Queue


Sink

TCP

TCP

ACK…

ACK…

Queue


Sink

TCP

TCP

ACK…

Drop!!!

Queue


Sink

TCP

TCP

Queue


Sink

TCP

TCP AQM

Congestion

Congestion Notification…

ACK…

Queue


Sink

TCP

TCP AQM

Advantages• Reduce packet losses (due to queue overflow)• Reduce queuing delay


QoS Router

Policer

Policer

Classifier

Policer

Policer

Classifier

Per-flow Queue

Scheduler

Per-flow Queue

Per-flow Queue

Scheduler

Per-flow Queue

shaper

shaper

Queue management


Packet Drop Dimensions

AggregationPer-connection state Single class

Drop positionHead Tail

Random location

Class-based queuing

Early drop Overflow drop


Typical Internet Queuing• FIFO + 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: drop policy (buffer management)


FIFO + Drop-tail Problems

• FIFO Issues: (irrespective of the aggregation level)– No isolation between flows: full burden on e2e control

(e..g., TCP)

– 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 => larger steady state queues/delays

– Synchronization: end hosts react to the 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


Synchronization Problem

• Because of Congestion Avoidance in TCP

cwnd

TimeRTT

1

2

4

Slow Start

W*

W W+1

RTT

Congestion Avoidance

W*/2


Synchronization Problem

Queue Size

Time

Total Queue

All TCP connections reduce their transmission rate on crossing over the maximum queue size.

The TCP connections increase their tx rate using the slow start and congestion avoidance.

The TCP connections reduce their tx rate again.It makes the network traffic fluctuate.


Global Synchronization Problem

Can result in very low throughput during periods of Can result in very low throughput during periods of congestioncongestion

Max Queue Length


Global Synchronization Problem

TCP Congestion control Synchronization: leads to bandwidth under-utilization

Persistently full queues: leads to large queueing delays

Cannot provide (weighted) fairness to traffic flows – inherently proposed for responsive flows

Flow 1Rate

Time

Flow 2

Aggregate load

bottleneck rate


Lock-out Problem

• Lock-Out:Lock-Out: In some situations tail drop allows a In some situations tail drop allows a single connection or a few flows (misbehaving single connection or a few flows (misbehaving flows: UDP) to monopolize queue space, preventing flows: UDP) to monopolize queue space, preventing other connections from getting room in the queue. other connections from getting room in the queue. This "lock-out" phenomenon is often the result of This "lock-out" phenomenon is often the result of synchronization. synchronization.

Max Queue Length


Bias Against Bursty Traffic

During dropping, bursty traffic will be dropped in benchs – which is During dropping, bursty traffic will be dropped in benchs – which is not fair for bursty connectionsnot fair for bursty connections

Max Queue Length


Active Queue ManagementGoals

• Solve lock-out and full-queue problems

– No lock-out behavior

– No global synchronization

– No bias against bursty flow

• Provide better QoS at a router

– Low steady-state delay

– Lower packet dropping


RED (Random Early Detection)

• FIFO scheduling

• Buffer management: – Probabilistically discard packets

– Probability is computed as a function of average queue length

Discard Probability

AverageQueue Length

0

1

min_th max_th queue_len


Random Early Detection(RED)


RED operationMin threshMax thresh

Average queuelength

minthresh maxthresh

MaxP

1.0

Avg length

P(drop)


Define Two Threshold Values

RED (Random Early Detection)

• FIFO scheduling

Min threshMax thresh

Average queuelength

Make Use of Average Queue LengthCase 1:

Average Queue Length < Min. Thresh ValueAdmit the New Packet


RED (Cont’d)


Average queuelength

Case 2: Average Queue Length betweenMin. and Max. Threshold Value

p

1-p

Admit the New Packet With Probability p…

p

1-p

Or Drop the New Packet With Probability 1-p


Random Early Detection Algorithm

• ave = (1 – wq)ave + wqq

• P = max_P*(avg_len – min_th)/(max_th – min_th)

for each packet arrival: calculate the average queue size ave if ave ≤ minth

do nothing else if minth ≤ ave ≤ maxth

calculate drop probability p drop arriving packet with probability p else if maxth ≤ ave drop the arriving packetarriving packet


Random early detection (RED) packet drop

MaxMaxthresholdthreshold

MinMinthresholdthreshold

Average queue lengthAverage queue length

Forced dropForced drop

ProbabilisticProbabilisticearly dropearly drop

No dropNo drop

TimeTime

Drop probabilityDrop probabilityMaxMax

queue lengthqueue length


Time

Max Queue Size

Active Queue ManagementRandom Early Detection (RED)

• Weighted average accommodates bursty traffic

Max Threshold

Min Threshold

Forced drop

Probabilistic drops

No drops

Drop probability

Average queue length

Probabilistic drops» avoid consecutive drops» drops proportional to bandwidth utilization

– (drop rate equal for all flows)


RED Vulnerable to Misbehaving Flows

0 10 20 30 40 50 60 70 80 90 1000

200

400

600

800

1,000

1,200

1,400

FIFORED

UDP blast

TC

P T

hro

ughp

ut

(Kby

tes/

Sec

)

Time (seconds)


Effectiveness of RED- Lock-Out & Global Synchronization

• Packets are randomly dropped

• Each flow has the same probability of being discarded


Effectiveness of RED- Full-Queue & Bias against bursty traffic

• Drop packets probabilistically in anticipation of congestion– Not when queue is full

• Use qavg to decide packet dropping probability : allow instantaneous bursts


What QoS does RED Provide?

• Lower buffer delay: good interactive service– qavg is controlled to be small

• Given responsive flows: packet dropping is reduced– Early congestion indication allows traffic to throttle back

before congestion

• RED provide small delay, small packet loss, and high throughput (when it has responsive flows).


Weighted RED (WRED)

• WRED provides separate thresholds and weights for different IP precedences, allowing us to provide different quality of service to different traffic

• Lower priority class traffic may be dropped more frequently than higher priority traffic during periods of congestion


Random

Dropping

WRED (Cont..)High Priority traffic

Medium Priority traffic

Low Priority traffic


AverageQueue Depth

StandardMinimumThreshold

PremiumMinimumThreshold

Std and PreMaximumThreshold

Adds Per-Class Queue Thresholds for Differential Treatment

Two Classes are Shown; Any number of classesCan Be Defined

Congestion Avoidance: Weighted Random Early Detection (WRED)

Probabilityof Packet

Discard


Problems with (W)RED – unresponsive flows


Vulnerability to Misbehaving Flows

• TCP performance on a 10 Mbps link under RED in the face of a “UDP” blast



• Try to look at the following example:

• Assume there is a network which is set up as:

UDP sources

R1 R2S(m)

S(1)

S(m+1)

S(m+n)

S(m)

S(1)

S(m+1)

S(m+n)

10Mbps

100Mbps 100MbpsTCP

sourcesTCP

sources

UDP sources


Vulnerability to Misbehaving FlowsThroughput Analysis

0

0.2

0.4

0.6

0.8

1

1.2

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20Flow number

Th

rou

gh

pu

t (M

bp

s)

Ideal

RED



• Queue Size versus Time

0 5 10 15 20 25 30 35 40 45 500

50

100

150

200

250CHOKe: Queue Size

Time (Seconds)

Siz

e of

Que

ue (

No.

of

Pac

kets

)

Average Queue SizeCurrent Queue Size

Delay is bounded

Delay is bounded

Global Synchronization solvedGlobal Synchronization solved

RED: Queue Size


Unfairness of RED

0 5 10 15 20 25 30 350

100

200

300

400

500

600

700

800

900

1000RED's Throughput

Flow Number

Thr

ough

put

(Kbp

s)Idea Fair ShareRED's Throughput

Unresponsive Flow (such

as UDP)

Unresponsive Flow (such

as UDP)

32 TCP Flows1 UDP Flow

32 TCP Flows1 UDP Flow

An unresponsiveflow occupies over 95% of bandwidth

An unresponsiveflow occupies over 95% of bandwidth


Scheduling & Queue Management

• What routers want to do?

– Isolate unresponsive flows (e.g., UDP)

– Provide Quality of Service to all users

• Two ways to do it

– Scheduling algorithms:

e.g., WFQ, WRR

– Queue management algorithms:

e.g., RED, FRED, SRED


The setup and problems

• In a congested network with many users

• QoS requirements are different

• Problem:

Allocate bandwidth fairly


• Network node: Weighted Fair Queueing (WFQ)

• User traffic: any type

Problem: complex implementation

lots of work per flow

Approach 1: Network-Centric


Approach 2: User-Centric

• Network node: simple FIFO buffer;

active queue management (AQM): RED

• User traffic: congestion-aware (e.g. TCP)

Problem: requires user cooperation


Current Trend

• Network node:

simple FIFO buffer

AQM schemes with enhancement to provide fairness: preferential dropping packets

• User traffic: any type


Packet Dropping Schemes

• Size-based Schemes drop decision based on the size of FIFO queue

e.g. RED

• Content-based Schemes drop decision based on the current content of the FIFO queue

e.g. CHOKe

• History-based Schemes keep a history of packet arrivals/drops to guide drop

decision

e.g. SRED, RED with penalty box, AFD


CHOKe (no state information)


Random Sampling from Queue

• A randomly chosen packet more likely from the unresponsive flow

• Unresponsive flows can’t fool the system


Comparison of Flow ID

• Compare the flow id with the incoming packet

– More accurate

– Reduce the chance of dropping packets from a TCP-friendly flows


Dropping Mechanism

• Drop packets (both incoming and matching samples)

– More arrival More Drop

– Give users a disincentive to send more


CHOKe (Cont’d)


Average queuelength

Case 1: Average Queue Length < Min. Thresh Value

Admit the New Packet


CHOKe (Cont’d)


Average queuelength

p

1-p

Case 2: Avg. Queue Length is between Min. and Max. Threshold Values

A packet is randomly chosen from the queue to compare with the new arrival packet

If they are from different flows, the samelogic in RED applies

If they are from the same flow, both packets will be dropped


CHOKe (Cont’d)


Average queuelength

Case 3:Avg. Queue Length > Max. Threshold Value

A random packet will be chosen forcomparison

If they are from different flows, the new packet will be droppedIf they are from the same flow,

both packets will be dropped


Simulation Setup


Network Setup Parameters

• 32 TCP flows, 1 UDP flow

• All TCP’s maximum window size = 300

• All links have a propagation delay of 1ms

• FIFO buffer size = 300 packets

• All packets sizes = 1KByte

• RED: (minth, maxth) = (100,200) packets


32 TCP, 1 UDP (one sample)


32 TCP, 5 UDP (5 samples)


How Many Samples to Take?

• Different samples for different Qlenavg

– # samples decrease when Qlenavg close to minth

– # samples increase when Qlenavg close to maxth


32 TCP, 5 UDP (self-adjusting)


Two Problems of CHOKe

• Problem I: – Unfairness among UDP flows of different rates

• Problem II:– Difficulty in choosing automatically how many to drop


SAC (Self Adjustable CHOKe

Tries to Solve the previously mentioned two problems


SAC• Problem 1: Unfairness among UDP flows of different rates (e.g.,

when k =1, the UDP flow 31 (6 Mbps) has 1/3 throughput of UDP flow 32 (1 Mbps), and when k =10 , throughput of UDP flow 31 is almost 0).

0 5 10 15 20 25 30 350

50

100

150

200

250

300

350

400

Flow number

Tth

roughput(

Kbps)

Throughput per flow (30 tcp flows and 2 udp misbehaving flows)

CHOKe 1 CHOKe 10 Ideal Fair Share


SAC• Problem 2: Difficulty in choosing automatically how many to drop

(when k = 4, UDPs occupy most of the BW. When k =10, relatively good fair sharing, and when k = 20, TCPs get most of the BW).

0 5 10 15 20 25 30 350

20

40

60

80

100

120

140

160

180

Flow number

Thro

ughput(

Kbps)

Throughput per flow (30 tcp flows and 4 udp misbehaving flows)

CHOKe 4 CHOKe 10CHOKe 20


SAC• Solutions:

1. Search from the tail of the queue for a packet with the same flow number and drop this packet instead of random dropping – because the higher a flow rate is, the more likely its packets will gather at the rear of the queue.

The queue occupancy will be more evenly distributed among the flows.

2. Automate the process of determining k according to traffic status (number of active flows and number of UDP flows)


SAC• Once an incoming UDP is compared with a randomly selected

packet, if they are of the same flow, P is updated in this way:

P (1-wp) P + wp.

• If they are of different flows, P is updated as follows:

P (1-wp) P .

• If P is small, then there are more competing flows, and we should increase the value of k.

• Once there is an incoming packet, if it is a UDP packet, R is updated in this way:

R (1-wr) R+ wr..

• If it is a TCP packet, R is updated as follows:

R (1-wr) R.

• If R is large, then we have a large amount of UDP traffic, and we should increase k to drop more UDP packets.


SAC simulation• Throughput per flow (30 TCP flows and 2 UDP flows of different

rate)


SAC simulation• Throughput per flow (30 TCP flows and 4 UDP flows of the same

rate).


SAC simulation• Throughput per flow (20 TCP flows and 4 UDP flows of different

rates)


AQM Using “Partial” state

information


Congestion Management and Congestion Management and Avoidance:Avoidance: GoalGoal

Provide fair bandwidth allocation similar to WFQ Be simple to implement like RED

Simplicity

Fa

irn

es

s

WFQ

RED

Ideal


• Objective: achieve fairness close to that of max-min fairness

1. If W(f) < C/N, then set R(f) = W(f).

2. If W(f) > C/N, then set R(f) = C/N.

• Formulation:– Ri: the sending rate of flow i

– Di: the drop probability of flow i

– Ideally, we want» Ri (1 – Di) = Rfair (equal share)

» Di = (1 – Rfair/Ri)+ (That is, drop the excess)

AQM Based on Capture RecaptureAQM Based on Capture Recapture


AQM Based on Capture-Recapture:AQM Based on Capture-Recapture:

Incoming packets

Active Queue Management

The estimation of the sending rate

The estimation of the fair share

The adjustment mechanism

The key question is: how to estimate the sending rate (Ri) and the fair share (Rfair) !!!

Fair allocation of BW


Capture-Recapture ModelsCapture-Recapture Models

• The CR models were originally developed for estimating demographic parameters of animal populations (e.g., population size, number of species, etc.).

– It is an extremely useful method where inspecting the whole state space is infeasible or very costly

– Numerous models have been developed to various situations

• The CR models are being used in many diverse fields ranging from software inspection to epidemiology.

• It is based on several key ideas: animals are captured randomly, marked, released and then recaptured randomly from the population.




Time is then allowed for the marked individuals to mix with the unmarked individuals.



Then another sample is captured.



Capture-Recapture Model

• Unknown number of fish in a lake

• Catch a sample and mark them

• Let them loose

• Recapture a sample and look for marks

• Estimate population size

• n1 = number in first sample 15 n2 = number in second sample 10 n12 = number in both samples 5 N = total population size assume that n1/N = n12/n2 therefore 15/N = 5/10

N = (10 x 15) / 5 = 30


Capture-Recapture ModelsCapture-Recapture Models• Simple model: estimate a homogeneous population of animals (N):

– n1 animals are captured (marked)

– n2 animals were recaptured, and

– m2 of these appeared to be marked.

• Under this simple capture recapture model (M0):

m2/n2 = n1/N

N

n1

N

n2

N

n1n2

m2



• The capture probability refers to the chance that an individual animal get caught.

• M0 implies that the capture probability for all animals are the same.

– ‘0’ refers to constant capture probability

• Using the Mh model, the capture probabilities vary by animal, sometimes for reasons like difference in species, sex, or age, etc..

– ‘h’ refers to heterogeneity.



• Estimation of N under the Mh Model is based on the capture frequency data f1, f2…, and ft (t captures)

– f1 is the number of animals that were caught only once,

– f2 is the number of animals that were caught only twice, … etc.

• The jackknife estimator of N is computed as a linear combination of these capture frequencies, s.t.:

N = a1f1 + a2f2 + … + atft

where ai are coefficients which are a function of t.


AQM Based on Capture-RecaptureAQM Based on Capture-Recapture

• The key question is: how to estimate the sending rate (Ri) and the fair share (Rfair) !!!

• We use an arrival buffer to store the recently arrived packet headers (we can have control over how large the buffer is, and is a better representation of the nature of the flows when compared to the sending buffer):

1. We estimate Ri using the M0 capture-recapture model

2. We estimate Rfair using the Mh capture-recapture model (by estimating the number of active flows).


AQM Based on Capture-RecaptureAQM Based on Capture-Recapture

Ri is estimated for every arriving packet (we can increase the accuracy by having multiple captures, or decrease it by capturing packets periodically)

If the arrival buffer is of size B, and the number of captured packets is Ci, then Ri = R Ci/B where R is the aggregate arrival rate

Rfair may not change every single time slot (as a result, the capturing and the calculation of the number of active flows could be done independently of the arrival of each incoming packet)

Rfair = R/(number of active flows)

The capture-recapture model gives us a lot of flexibility in terms of accuracy vs. complexity

The same capture-recapture can be used for calculating both Ri and Rfair


AQM Based on Capture-AQM Based on Capture-RecaptureRecapture

Incoming packets

Active Queue Management(Capture-Recapture)

The estimation of Ri by the M0 model

The estimation of Rfair by the Mh CR model

Di = (1 – Rfair/Ri)+

Fair allocation of BW


Performance evaluationPerformance evaluation• This is a classical setup that researchers use to evaluate AQM

schemes (we can vary many parameters, responsive vs. non-responsive connections, the nature of responsiveness, link delays, etc.)

UDP sources

R1 R2S(m)

S(1)

S(m+1)

S(m+n)

S(m)

S(1)

S(m+1)

S(m+n)

10Mbps

100Mbps 100MbpsTCP

sourcesTCP

sources

UDP sources


Performance evaluationPerformance evaluation• Estimation of the number of flows

Estimation of numnber of flows (varying number of flows)

0

10

20

30

40

50

60

Time

Est

imat

ed n

umbe

r of

flow

s

SRED CAP Ideal


Performance evaluationPerformance evaluation• Bandwidth allocation comparison between CAP and RED

0

0.2

0.4

0.6

0.8

1

1.2

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25Flow number

Thro

ughp

ut (M

bits

/s)

Ideal RED CAP


Performance evaluationPerformance evaluation• Bandwidth allocation comparison between CAP and SRED

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25Flow number

Thro

ughp

ut (M

bits

/s)

Ideal SRED CAP


Performance evaluationPerformance evaluation• Bandwidth allocation comparison between CAP and RED-PD

0

0.2

0.4

0.6

0.8

1

1.2

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25Flow number

Thro

ughp

ut (M

bits

/s)

Ideal RED RED-PD CAP


Performance evaluationPerformance evaluation• Bandwidth allocation comparison between CAP and SFB

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25Flow number

Thro

ughp

ut (M

bits/s

)Ideal SFB CAP


Normalized Measure of PerformanceNormalized Measure of Performance

• A single comparison of the fairness using a normalized value, where norm is defined as:

where bi is ideal fair share, bj is the bandwidth received by each flow

• Thus, ||BW|| = 0 for the ideal fair sharing


Normalized Measure of Normalized Measure of PerformancePerformance

0

200

400

600

800

1000

1200

25 30 25 40 45 50 55 60 65 70number of flows

||BW

||

Ideal RED SRED SFB RED-PD CAP


Performance Evaluation: Variable Performance Evaluation: Variable amount of unresponsivenessamount of unresponsiveness

0

500

1000

1500

2000

2500

3000

1 5 10 15 20 25 30UDP load (% of bottleneck link)

||BW

||

Ideal RED RED-PD SFB SRED CAP

CSIT560 by M. Hamdi 1 QoS in The Internet: Scheduling Algorithms and Active Queue Management.

Documents

pipe slide

flow queue scheduler

active queue management

classes fairness slide

qos guarantees

state ip circuit

b b qos metrics

scheduling scheduler