1

IP Quality of Service (QoS)Muhammad JaseemuddinDept. of Electrical & Computer EngineeringRyerson UniversityToronto, Canada

References1. Larry L. Peterson, Bruce S. Davie, Computer Networks: A Systems

Approach, 2nd edition, Morgan-Kaufmann, 2002.2. Paul P. White, RSVP and Integrated Services in the Internet: A Tutorial,

IEEE Communications Magazine, May 1997.3. David Durham, Raj Yavatkar, Inside the Internet’s Resource

Reservation Protocol, John Wiley,1999.4. A. Leon-Garcia and I. Widjaja, Communication Networks: Fundamental

Concepts and Key Architectures, McGraw Hill, 2000.5. D. D. Clark and W. Fang, Explicit Allocation of Best-Effort Packet

Delivery Service, IEEE/ACM Transactions on Networking, Vol. 6, No.4, August 1998.

6. H. J. Chao and X. Guo, Quality of Service Control in High-Speed Networks, John Wiley & Sons.

2

Queuing Discipline – FIFO❒ Scheduler:A scheduler S serves the queues in the

order defined by its type❒ Max-Min Fair Share: if scheduler allocates link

capacity such that it maximizes the minimum share of a user whose demand is not fully satisfied

❒ Protection: A misbehavior user (exceeding its rate limit) does not affect the performance of other users

❒ Work Conserving: A scheduler is idle only when the queue is empty

❒ Non-work Conserving: A scheduler may be idle even if there is packet in the queue

❒ First-In-First-Out (FIFO)❍ Packets are scheduled in the order they arrive❍ does not discriminate between traffic sources❍ Jitter tends to increase with the number of hops, since

the delays at different hops are uncorrelated❍ Jitter of all flows are shared by all of them❍ It does not provide max-min fair share nor protection

Queue 1

Queue 2

S

SQueue 1

Queuing Discipline – RR Flow 1

Flow 2

Flow 3

Flow 4

Round-robinservice

❒ Round Robin (RR)❍ The scheduler polls each queue in cyclic order and

serves the packet in a queue if the queue is non-empty❍ It provides protection as a misbehaving user overflows

its own queue ❍ It gives equal share of bandwidth among users

❒ Weighted Round Robin (WRR)❍ The scheduler performs RR scheduling but instead of

serving one packet at a time serves packets in proportion to the weight of the queue

❍ It shares the link according to weight, e.g. Wa=3 and Wb=7 indicates 30% capacity is allocated to Queue A and 70% to B

❒ Deficit Round Robin (DRR)❍ Modifies WRR to handle variable packet size in fair manner❍ A queue i is allocated its quantum number of bits Qi for a round❍ Any left over bits, Deficit Di, are carried in another round❍ In every round Di is updated as Di = Di + Qi and that many bits are served❍ If a packet is large, then it is deferred for the next round when enough bits are available for the service

3

Queuing Discipline – WFQ❒ Fair Queuing (FQ)

❍ Explicitly segregates traffic based on flows❍ Ensures no flow captures more than its share of capacity

• it prevents a flow from hogging the queue resource❍ Each flow is guaranteed at least C/n bps, if C is the link capacity in bps

and n is the number of flows• Equitable sharing of link bandwidth• View the queue as a fluid flow system• A flow can get more bandwidth if the queue is empty – work

conserving❍ Variation: weighted fair queuing (WFQ)

❒ Problem of FQ?❍ No preemption❍ Short packets may miss its deadline by large packets under transmission❍ Jitter is experienced by the flow causing burstiness – it is not shared

• Jitter may be too high

Queue 1@ t=0

Queue 2@ t=0

1

t1 2

Fluid-flow system:both packets served at rate 1/2

Both packetscomplete serviceat t=2

0

Packet fromqueue 1 beingserved

1

t1 2

Packet-by-packet system:queue 1 served first at rate 1;then queue 2 served at rate 1.

Packet from queue 2being served

Packet fromqueue 2 waiting

0

Packet-by-Packet Fair Queuing

4

FQ Algorithm❒ Suppose clock ticks each time a bit is transmitted

❍ Suppose each queue is served one bit in a round❍ A round consists of number of clock ticks to serve one bit from all n

queues (flows)❍ R(t) is the number of rounds at time t, which is a function of number of

active queues❍ dR(t)/dt = C/nactive(t) is a piece-wise linear function

❒ Let P(i,k) denotes the length of packet k of flow I❒ Let ti

k is the arrival time and R(tik) be the arrival round of packet P(i,k)

❒ Let S(i,k) denotes the round when start transmitting packet k❒ Let F(i,k) denotes the round when finish transmitting packet P(i,k)❒ F(i,k) = S(i,k) + P(i,k) ❒ When does router start transmitting packet P(i,k)

❍ if queue is not empty, then from the next round after transmitting P(i,k-1)❍ If queue is empty, then start transmitting from R(ti

k)❒ Thus: F(i,k) = MAX (F(i,k-1) , R(ti

k) ) + P(i,k) ❒ Finished round number can be used as a priority tag

Queue 1@ t=0

Queue 2@ t=0

2

1

t3

Fluid-flow system:both packets served at rate 1/2

Packet from queues served at rate 1

0

2

1

t1 2

Packet-by-packet fair queuing:queue 2 served at rate 1

Packet fromqueue 1 beingserved at rate 1

Packet fromqueue 2 waiting

0 3

Packet-by-Packet Fair Queuing

5

FQ Algorithm - Example❒ For multiple flows

❍ calculate Fi for each packet that arrives on each flow❍ treat all Fi’s as timestamps❍ next packet to transmit is one with lowest timestamp

❒ Not perfect: can’t preempt current packet❒ Example

Flow 1 Flow 2

(a) (b)

Output Output

F = 8 F = 10F = 5

F = 10

F = 2

Flow 1(arriving)

Flow 2(transmitting)

Weighted Fair Queuing (WFQ)❒ Since different users have different requirements

❍ Non-equitable fair sharing❍ Giving each flow certain proportion of the bandwidth

❒ Each flow is assigned a weight that determines its relative share of the bandwidth

❒ Fi = MAX (Fi - 1, Ai) + Pi /wi❍ Pi/wi shows the depletion of flow i queue at a rate proportional to

its bandwidth share determined by its weight

❒ WFQ ensures that a flow i with weight wi will get at least (wi/W)C bandwidth

❍ W = ∑ wi

❍ C is the link capacity in Mbps

6

Queue 1@ t=0

Queue 2@ t=0

1

t1 2

Packet-by-packet weighted fair queuing:Flow 1 will receive 1 bit/round (w1=1) and Flow 2 will receive 3 bits/round (w2=3)queue 2 served first at rate 1

F(2,1) = 0 + 1/3 = 1/3then queue 1 served at rate 1

F(1,1) = 0 + 1/1 = 1

Packet from queue 1 being served

Packet fromqueue 2 beingserved

Packet fromqueue 1 waiting

0

1

t1 2

Fluid-flow system:packet from queue 1served at rate 1/4;

Packet from queue 1 served at rate 1

Packet from queue 2served at rate 3/4

0

Packet-by-Packet WFQ

Real time Applications

❒ Require “deliver on time” assurances❍ must come from inside the network

❒ Example application (audio)❍ sample voice once every 125us❍ each sample has a playback time❍ packets experience variable delay in network❍ add constant factor to playback time: playback point

Microphone

Speaker

Sampler,A D

converterBuffer,D A

7

Playback Buffer

Seq

uenc

e nu

mbe

r

Packetgeneration

Networkdelay

Buffer

Playback

Time

Packetarrival

Example Distribution of Delays

1

2

3

Pack

ets

(%)

90% 97% 98% 99%

150 20010050Delay (milliseconds)

• One way delay measured on certain path in Internet on one day• 97% of packets experience delay 100ms or less

8

TaxonomyApplications

Real time

Tolerant

Adaptive Nonadaptive

Delay-adaptive

Rate-adaptive

Intolerant

Rate-adaptive Nonadaptive

Interactive Interactivebulk

Asynchronous

Elastic

• Real time applications• Tolerant: tolerant of losses• Adaptive: adaptable to delay

• Delay Adaptive: Applications that can adjust playback point• Rate Adaptive: Applications that trade off coding with bit rate

Integrated Services❒ Service Classes

❍ Guaranteed (RFC 2212)• Provide bound on delay• Applications set the playback time adjusted to the pre-specified delay

bound❍ controlled-load (RFC 2211)

• provide the same level of service as unloaded best-effort it controls delay and losses under load condition

• For adaptive real-time applications that can tolerate certain losses and delays

❒ Mechanisms❍ signaling protocol

E.g. RSVP❍ admission control

• Parameter-Based Admission Control (PBAC)• Measurement-Based Admission Control (MBAC)

❍ policing❍ packet scheduling

• WFQ, PQ, CB

9

Flowspec

❒ Rspec: describes service requested from network❍ controlled-load: none❍ guaranteed: delay target

❒ Tspec: describes flow’s traffic characteristics❍ average bandwidth + burstiness: token bucket filter❍ token rate r

• must have n tokens to send n bytes❍ bucket depth B

• start with no tokens• accumulate tokens at rate of r per second• can accumulate no more than B tokens

Router Mechanisms

❒ Admission Control❍ decide if a new flow can be supported❍ answer depends on service class❍ not the same as policing

❒ Packet Processing❍ classification: associate each packet with the appropriate reservation❍ scheduling: manage queues so each packet receives the requested service

RSVP Process

Routing

PacketClassifier

PacketScheduler

AdmissionControl

Policy Control

Data Flow in

RSVP Signaling

RSVP Signaling

Data Flow out

10

Guaranteed Service❒ It provides:

❍ Assured level of bandwidth❍ Maximal end-to-end delay bound

• It does not control minimal and average delays❍ No queuing loss for conforming packets

❒ It is for playback applications with strict delay bound requirement❍ the maximal delay bound is normally very large, packets usually arrive

much earlier❍ The application must be capable of storing packets that arrive early

❒ In a fluid flow model the delay bound is b/R provided R >= r❒ To allow deviation, two error terms C and D are defined

❍ C is rate dependent error factor measured in bytes• For example time taken in the network for serializing IP datagrams into ATM

cells at a frequency 1/r❍ D is rate independent error factor measured in micro-seconds

• For example, in some slotted network specific slots are reserved for guaranteed bandwidth traffic. The packets have to wait for a pre-determined amount of time to grab the slot

❍ The delay bound will be: b/R + C/R + D

Guaranteed Service Specification❒ Tspec parameters are:

❍ P = peak rate of flow (bytes/sec)❍ B = bucket depth (bytes)❍ r = token bucket rate (bytes/sec)❍ m = minimum policed unit (bytes)

• Packets of smaller size will be counted by the policer as of size m❍ M = maximum datagram size (bytes)

❒ Rspec parameters are:❍ R = service rate (bytes/sec)

• The rate that will be reserved at every router❍ S = slack term (micro seconds)

• The maximum delay tolerance• a router if does not have enough available bandwidth as requested by

the receiver in R, then it reserves R’ that will cause increase in delay di such that di < S.

11

Guaranteed Service❒ Each router reserves bandwidth R and buffer space B❒ The traffic is policed at the network entry point

❍ Non-conforming packets are treated as best-effort datagram❍ Marked to get the same treatment in the downstream routers❍ Datagrams larger than MTU are policed as non-conformant

❒ The traffic is shaped at every heterogeneous source branch points and all source merge points

❍ Heterogeneous branch point is the point in the multicast distribution tree where the traffic from a source replicates and get forwarded on distinct paths

• The Tspec reservation is not the same on all the links• Reshaping is required when the Tspec of the outgoing link is less than the Tspec of the

immediately upstream link❍ Source merge point is the point in the multicast distribution tree where traffic

from different sources sharing the same reservation merge❒ The amount of buffering required at a reshaping point to reshape the

traffic to the token bucket specification is b+Csum+(Dsum*r)❍ Csum and Dsum are the sum of parameters C and D between the last reshaping

point and the current reshaping point

Guaranteed Service❒ The source and any reshaping point when transmits at peak rate p must not

exceed M+pT data transmission in any time T, where M is the maximum packet size

❍ The traffic must not exceed at any time T M+min(pT, rT+b-M)❒ Considering only token bucket traffic descriptor (r,b)

❍ Q = b/R + Ctot/R + Dtot❍ Buffer space = b + Csum + (Dsum*R)

❒ Considering peak rate p and maximum packet size M more accurate bound equations are:

❍ Q = (b-M)(p-R)/R(p-r) + (M+Ctot)/R + Dtot when p>R>=r❍ Q = (M+Ctot)/R + Dtot when R>=p>=r

❒ Peak rate puts a limit on the rate at which burstiness b can be caused❍ Hence considering peak rate in computing buffer space gives less buffer

requirement❒ Considering peak rate, the buffer requirement can be computed as:

❍ M+(b-M)(p-r)/(p-r)+(Csum/R+Dsum)*r if (b-M)/(p-r) < (Csum/R+Dsum)❍ M+(b-M)(p-R)/(p-r)+(Csum/R+Dsum)*R if (b-M)/(p-r) > (Csum/R+Dsum)

and p >R❍ M+(b-M)(p-p)/(p-r)+(Csum/R+Dsum)*p if (b-M)/(p-r) > (Csum/R+Dsum) and

p < R

12

Guaranteed Service

❒ Slack term indicates the difference between the required delay bound and the bound on delay estimated by the fluid flow model❍ S = Dreq – (b/r + C/r + D), where R>=r❍ It is used by the routers to reserve less than R resource❍ But the overall delay bound advertised by the network

remains unchanged

❒ A router can employ WFQ to implement this service

❒ Applications can set their playback points❍ (b-M)/R*(p-R)/(p-r)+(M+Ctot)+Dtot where R<p

Controlled Load Service❒ The application receiving CL service is assured by the

routers along the path to receive the end-to-end behavior tightly approximated to the best-effort datagram delivery under unloaded network condition

❒ The application then expects:❍ A very high percentage of packets will be successfully delivered

• The percentage of packet drop will be closed to that caused by the basic packet error rate of the transmission medium

❍ A very high percentage of packets delivered will experience delay that does not greatly exceed the minimum transmit delay of any successfully delivered packet

• Min. transmit delay includes medium propagation delay and fixed processing time in the routers and other network elements

❒ It provides no firm quantitative guarantees on loss or delay

❒ It is used by adaptive realtime applications ❒ The application is required to provide an estimation of its

traffic in Tspec – as for guaranteed service

13

Controlled Load❒ If a router accepts a flow for CL, it makes a commitment that the flow

will continue receiving the service equivalent to the best-effort service on a lightly loaded network

❍ CL service does not deteriorate if the network load increases• Regardless of the level of load increase

❍ It essentially puts an expectation on the duration of service disruption• short term disruption is considered due to ‘statistical’ effect• Long term disruption is indicative of insufficient allocation of

resources along the path❍ It employs statistical approach to resource allocation

• must allocation bandwidth higher than the token rate r• employs any scheme to control the burst effect

– E.g. borrow the bandwidth from other flows to clear a burst❍ It can implement the service using priority queue

• Higher priority queue for controlled load• Lower priority queue for best-effort service

❍ It controls the admission of flows into CL queue

Reservation Protocol❒ Proposed Internet standard: RSVP❒ Consistent with robustness of today’s connectionless model❒ Uses soft state (refresh periodically)❒ Designed to support multicast❒ Receiver-oriented❒ Two messages: PATH and RESV❒ Source transmits PATH messages every 30 seconds❒ Destination responds with RESV message❒ Merge requirements in case of multicast❒ Can specify number of speakers

14

RSVP – PATH Message❒ Path Setup

❍ PATH Message is generated by the source(sender) and traverse along the route to the destination (receiver)

❍ It causes all the RSVP capable router to install path state – path setup❒ It contains

❍ Phop• The address of the last RSVP capable router• Updated by every RSVP capable router on the path

❍ Sender template• Filter specification to identify the traffic

❍ Sender Tspec

RSVP Path Message Content❍ Adspec - One Pass With Advertisement (OPWA)

• Updated by every RSVP capable router• Contains General Parameter Segment and service specific segment.

GPS contains:– min. path latency: sum of link propagation delays along the path– min. bandwidth: minimum of individual link bandwidth along

the path– path MTU: min. MTU of individual links along the path– IS hop count: incremented by each RSVP capable router along

the path, indicating their number– Global break bit: This bit is cleared by the sender. It is set by a

router that finds the phop router (address) is not directly connected through the incoming interface on which it receives the path message. When the bit is set it indicates that there are routers on the path that do not support RSVP (IS), hence the adspec content may be invalid.

15

RSVP Path Message Processing❒ It is intercepted by every RSVP capable router along the path

❍ If the router finds any error, it sends PathErr message back to the sender❍ If it finds no error, updates or creates the path state that contains

• Sender template, Phop, Adspec• Sender TSPec: provides a ceiling to clip an erroneously high

reservation request in Rspec• Set cleanup timer to cleanup time interval

❍ Path state is deleted if the refresh Path message doesn’t arrive at least once in the cleanup time interval – RSVP soft state

RSVP – PATH Message Refresh

❒ Router generates Path messages❍ if the routing update causes change in outgoing

interface in the forwarding path❍ if the router detects any change in the path state❍ Refresh the downstream nodes

• Refresh period is typically much less than cleanout time interval to tolerate few Path message losses

❒ Path tear down❍ Explicit PathTear message can be generated by the

sender to remove path state from the routers along the path

• Expedite path tear down process

16

Reservation Model

❒ Reservation Models❍ OPWA Reservation

• Sender includes Adpsec• Receiver determines end-end service resulting from the request

❍ One-pass Reservation• No adspec – receiver reserves the resources without

determining the end-end service

RSVP – RESV Message❒ Receiver sends Resv Message upstream to the sender

❍ It is sent hop by hop to the address a RSVP node learns from the Phop object

❒ Resv Message contains❍ An indication of the reservation style❍ Filter spec

• Identifies sender – similar to the sender template❍ Flow spec

• Rspec– Calculated value of R and slack term to zero

• Tspec– Typically set to sender Tspec with M changed to PathMTU

• Optionally a reservation confirmation object, ResvConf, containing the IP address of the receiver. The router at which the reservation ends up the distribution tree if finds ResvConf sends ResvConf message to the receiver indicating that there is a high probability that the reservation succeeds

17

Reservation Style

❒ The Filterspec and the effective Flowspec at any router is obtained as a result of merge process

❍ Filterspec identifies the packets to which the reservation applies❒ The merging rules depend upon the reservation style❒ The router calculates the filterspec and flowspec to be sent

upstream applying merging rules on the stored state❍ If as a result of merging Resv state changes, then the new Resv

message is sent immediately❍ Otherwise the Resv message is sent following the periodic refresh

cycle❒ Merging rules apply to the packets of the same session

Wildcard-Filter(WF)

Wildcard

Shared-Explicit(SE)

Fixed Filter (FF)Explicit

SharedDistinct

Fixed Filter (FF)❒ The fixed filter indicates a single sender only❒ The router receiving Resv message checks the installed reservation

states on the received interface for that particular sender❒ The effective Flowspec of the reservation installed is the maximum of

all FF reservation requests received on that interface for that particular sender

ReserveS1 {4B}

| S2 {3B}|S3 {5B}ReserveS2 {6B}

| S4 {4B} |S6 {2B}

ReserveS2 {3B}

| S3 {2B} |S5 {4B}

FF(S1{2B},S2{3B},S4{5B})FF(S1{4B},S2{2B})

FF(S4{4B})FF(S2{6B},S4{2B},S6{2B})

FF(S2{3B},S3{2B},S5{4B})

FF(S1{4B},S2{6B})

FF(S3{2B},S4{5B})

FF(S5{4B},S6{2B})

Toward S1,S2

Outgoing requests after merging Incoming reservation request

Toward S3,S4

Toward S5,S6

18

Shared Explicit (SE) Type❒ The Filterspec of each stored SE reservation on the

interface identifies a set of individual hosts (senders)❒ The Filterspec

Reserve

|(S1,S2,S4 {5B})|

Reserve

|(S4,S6 {3B})|

Reserve

|(S2,S3,S5 {4B})|

SE((S2,S4){5B})SE((S1,S2){2B})SE((S1,S2){5B})

Outgoing requests after merging Incoming reservation request

SE((S3,S4){5B})

SE((S5,S6){4B})

Toward S1,S2

Toward S3,S4

Toward S5,S6

SE(S4){3B})SE(S4,S6{2B})

SE(S2,S3,S5){4B})

Wildcard Filter (WF)❒ The Filterspec of WF reservation installed at an

interface matches any sender from upstream❍ Indicates the reservation shared by all upstream senders

❒ The single Resv message sent upstream contains the maximum Flowspec of all WF reservation requests received on that particular interface

Reserve

| (*{5B}) |

Reserve

| (* {3B}) |

Reserve

| (* {4B})

WF(*{5B})WF(*{2B})WF(*{5B})

Outgoing requests after merging Incoming reservation request

WF(*{5B})

WF(*{5B})

Toward S1,S2

Toward S3,S4

Toward S5,S6

WF(*{3B})WF(*{2B})

WF(*{4B})

19

Slack Term❒ Receivers includes slack term S and R in its Resv message for guaranteed service❒ It represents the amount by which the end to end delay is below the delay bound

requirement of the application❒ It gives flexibility to schedulers

❍ Deadline schedulers can overestimate delay while reserve R bandwidth❍ Rate based scheduler incur more delay when reserve less than R bandwidth

❒ It must satisfy: Sout + b/Rout + Ctoti/Rout <= Sin + b/Rin + Ctoti/Rin where r <= Rout <= Rin

❒ It increases the probability of reservation success❒ For example, for a traffic with r=1.5Mbps, the receiver computes (R=2.5Mbps, S=0)

reservation request❍ R3 has only 2Mbps unused bandwidth, hence it denies the reservation request and sends ResvErr

message to the receiver❒ The receiver computes new reservation (R1=3Mbps, S1>0)

❍ R3 uses the margin in delay provided by S1 and reserves 2Mbps < R1, which causes an increase in delay di < S1

❍ It computes and send upstream new reservation request (R2,S2=S1-di)

RSVP Example

R

R

R

R

R

Sender 1

Sender 2

PATH

PATH

RESV(merged)

RESV

RESV

Receiver B

Receiver A

20

Integrated Services Implementation

WFQ

PQ

GS Flows

CLS Flows

BE Flows

Class 1

Class 2

Packet Scheduler

Existing Flows

Admission ControlNew Flows

Reject

Accept

Router

RSVP

IntServ Issues

❒ Issues with IntServ❍ Scalability

• Every router needs to keep per flow RSVP state

❍ It does not recognize internet domain architecture• Autonomously administered domains• It is a gross assumption that every domain will provide the

same bounded service

❍ Service definition is the forte of service providers• It is business oriented

21

Allocated-Capacity Framework (Clark & Feng)

❒ Problem with IntServ: scalability❒ Idea: support two classes of packets

❍ Distinguish between packets conforming to the profile and non-conforming packets

❍ Can be used to offer different service class• premium• best-effort

P(drop)

1.0

MaxP

Minin MaxinMaxoutMinout

AvgLen

❒ Predecessor of Differentiated Services

❒ Mechanisms❍ packets: ‘in’ and ‘out’ bit❍ edge routers: tag packets❍ core routers: RIO

(RED with In and Out)

• Service Allocation Profile• Traffic Specification

• what exactly is provided? (e.g. 5 Mbps average throughput)• Geographic Scope

• To where the service is provided? (e.g. a group of destinations, or all in a particular network, or every where etc)

• Probability of Assurance• With what level of assurance the service is provided? (level of performance uncertainty the user is expected to tolerate)

• Profile Meter• meters the traffic and test it for conformance to the service allocation profile• marks the conformant packets in (in-profile) and the packets exceeding the profile as out (out-profile)

• Preferential Dropping Scheme• Every router must be capable of preferentially dropping tagged (marked) packets• Some sort of RED

Framework Implementation

22

Implementation Scenario

RD

M4 M5ISP2

RD

M2 M3ISP1

RD

M6ISP3

M1H1

H2• Location of Profile Meters

• Inside the source host • policy meter – it marks the packets based on policy relationship with the ISP

• At the network entrance point• checks the incoming traffic and marks the packets out if they exceed the profile

• At the network exit point• policy meter – it chooses the packets to mark based on administrative policy

• Edge Routers• at the access networks interfacing source hosts meter the flows• at the network edges interfacing other networks operating under peering relationship with the other network meter flow aggregates (aggregation of flows)

• Core Routers• perform preferential dropping on flow aggregates

Implementation Algorithm• Differential Dropping

• RIO: RED routers with in/out bit• Twin Algorithm

• two sets of RED parameters one for in packets one for out packets• upon packet arrival router determines whether packet is in or out• For in packets

• router computes avg_in, average number of in packets in the queue• three RED parameters are min_in, max_in, and Pmax_in• [0, min_in) normal operation• [min_in, max_in) congestion avoidance phase• [max_in, ∝) congestion control phase

• For out packets• router computes avg_total, average number of total (in+out) packets in the queue• three RED parameters are min_out, max_out, and Pmax_out• [0, min_out) normal operation• [min_out, max_out) congestion avoidance phase• [max_out, ∝) congestion control phase

23

RIO – Characteristics

• RIO router drops out packets more aggressively than in packets• it drops out packets much earlier than in packets, by choosing min_out smaller than min_in• it drops out packets with a higher probability than in packets during congestion avoidance phase by setting Pmax_out higher than Pmax_in• it goes into congestion control phase for out packets much earlier by choosing max_out smaller than max_in

it drops out packets first when it detect incipient congestion, and drops all out packets if the congestion persists

• It drops in packets only as a last resort when the router is flooded with in packets• this should not happen in a well-provisioned network• if the router consistently operates in congestion control phase indicated by in packet drop rate, it shows that the router is under-provisioned

Pmax_out

P(drop)

1.0

MaxoutMinout

Avg_total

P(drop)

1.0

Pmax_in

Minin Maxin

Avg_in

Simulation

1

2

3

4

5

6

7

8

9

10

A

11

12

13

14

15

16

17

18

19

20

B33 Mbps

5 ms

• Using ns-2• Ten TCP connections via a single ISP• All connections share a common link with BW 33 Mbps• The connections have different RTTs from 20 to 100 ms• Their target rates RT are either 1 Mbps or 5 Mbps • All TCP connections run for 20 s

24

Result – TCP connections onlyConnection RTT RED routers RT With RIO# (ms) (Mbps) (Mbps) (Mbps)

1 20 7.04873 1 2.27892 20 6.22214 5 5.76193 40 2.83662 1 1.280114 40 2.28316 5 5.267575 50 2.62307 1 1.219576 50 2.81556 5 5.188237 70 1.61073 1 1.348318 70 1.57837 5 4.127949 100 1.64488 1 0.99632610 100 1.85132 5 4.12563

Total 30.51458 31.588476

• Parameters for RED (10, 30, 0.02)• Parameters for RIO - for in packets (40, 70, 0.02) and for out packets (10, 30, 0.2)• TCP Reno

Result – TCP + CBR connections

Connection RTT Today’s Internet RT With RIO# (ms) (Mbps) (Mbps) (Mbps)

1 20 7.04873 1 1.440032 20 5.36329 5 5.211023 40 1.98478 1 1.071884 40 2.56938 5 5.012375 50 2.443 1 1.238276 50 2.54567 5 4.762367 70 1.28912 1 1.03328 70 1.53377 5 4.476489 100 1.1074 1 0.81777310 100 1.50127 5 4.3094CBR 50 5.85338 0 2.61297

Total 31.59858 31.9857

• Parameters for RIO - for in packets (40, 70, 0.02) and for out packets (10, 30, 0.2)• TCP SACK• CBR is UDP traffic at 6 Mbps

25

Conclusion• Scalable

• edge routers do metering, marking and differential dropping• core routers only perform differential dropping• edge routers in access network potentially perform per flow policing• all other routers perform actions on in and out packets – aggregate flow

• Service profile contains service assurance – qualitative• it delivers service assurance by protecting TCP from non-responsive to congestion flows

• Recognizes ISP domain boundaries• Service delivery requires proper provisioning• RIO provides basis of Differentiated Services Architecture

Differentiated Service Architecture (RFC 2475)• Differentiated Service Architecture defines an architecture for implementing scalable service differentiation in internet

• Service• A "Service" defines some significant characteristics of packet transmission in one direction across a set of one or more paths within a network. These characteristics may be specified in quantitative or statistical terms of throughput, delay, jitter, and/or loss, or may otherwise be specified in terms of some relative priority of access to network resources.

• Architectural Components• Per-hop forwarding behavior• Packet classification functions• Traffic conditioning functions including metering, marking, policing and shaping

• Scalable• introduce complex classification and traffic conditioning at network boundary nodes• network core nodes apply per-hop behavior on packets marked by edge routers with appropriate DS code point

• DS code code point indicates per-hop behavior to be invoked• It provides service differentiation in only one direction – asymmetric

26

Per-Hop Behavior (PHB)• A per-hop behavior is externally observable forwarding behavior at a DS node applied to a DS behavior aggregates

• DS behavior aggregate: a group of packets with the same DS code point crossing the link in a particular direction• observable forwarding behavior: loss, delay and jitter• in the event that only one behavior aggregate occupies a link, the observable forwarding behavior often depends upon relative link loading• when multiple behavior aggregates compete for the link resources, the distinctions in observable behavior can be made• a simple example of a PHB is one that guarantees certain minimal share X% of link bandwidth (over a reasonable time interval) to a behavior aggregate

• The PHB is the means by which a router allocates resources to a behavior aggregate• It can be specified:

• in terms of its resource requirement, e.g. buffer space, bandwidth• in terms of its priority relative to other PHBs• in terms of relative observable traffic characteristics, e.g. delay, loss

• PHBs are implemented at a node by buffer management and scheduling

DiffServ (DS) Code Point❒ 8-bit field in IP Header defined in RFC 2474❒ The DS Field contains the DS Code Point (DSCP)

DS Field

ToS Field

❒ DS Field is not compatible with ToS

27

Class Selector (CS) Code Point – RFC 2474❒ Provides backwards compatibility with legacy routers

❍ Uses 3 IP Precedence bits in ToS Field

❒ 8 Code Points❍ CS7 (Highest priority)❍ ….❍ ….❍ CS0 (Lowest priority)

ToS Field

Packet Treatment at Network EdgePolicing

Scheduling

MarkingDSCP

Shaping

Multi-Field ClassificationFiltering based on:

Source/Destination IP Address,TCP/UDP Port, Protocol ID

PolicingDoes the flow conform to policy?

MarkingMark DS Code Point (DSCP) based on network policy

Congestion Management Drop PrecedenceCongestion Avoidance: RED

ShapingImproves efficiency of link speed

SchedulingEnsures queues get serviced according to priorities

28

Packet Treatment at Network Core

Behavior Aggregate (BA) ClassificationFiltering based on:

DS Code Point (DSCP)

Congestion Management Drop PrecedenceCongestion Avoidance: RED

ShapingImproves efficiency of link speed

SchedulingEnsures queues get serviced according to priorities

Shaping

Scheduling

Policing

MarkingDSCP

Shaping

Scheduling

Assured Forwarding PHB (RFC 2597) • AF PHB provides different level of forwarding assurances – It is designed in response to a demand for some kind of assured forwarding of IP packets

• e.g. a company with several sites desire to receive assurance on packet forwarding at high probability for its intranet traffic as long as its traffic aggregate from each site does not exceed some agreed upon subscribed bit rate

• if the packets exceed the profile (out-profile), then they are not delivered at high probability as the probability of in-profile packet delivery• packets of the same micro-flow are not re-ordered regardless whether they are in or out packets

• AF PHB• Four classes, where each class is allocated certain bandwidth and buffer resources• within each AF class there are three drop precedence

• at time of congestion the drop precedence within a class determines the drop probability of a packet in that class

• congested router protect packets marked with lower drop precedence code point by preferentially dropping packets marked with higher dropprecedence code point

• within a DS router the level of forwarding assurance depends upon:• the amount of bandwidth and buffer space allocated to an AF class• the current load within the AF class• the drop precedence marking of the packets

29

AF PHB• AF PHB implementation

• A DS router must implement at least two loss probability within each AF class• this will be suffice in a network, e.g. enterprise network, where congestion is a rare and brief occurrence

• it can implement three loss probabilities with an AF class in a network where congestion is more frequent and persistent• it must not re-order packets belong to the same AF class regardless of their drop precedence values

• it can delay the packet without any quantified bound on delay and jitter• an AF implementation must avoid long term congestion while respond to short term congestion (packet bursts) by queuing packets

• it can drop packets to minimize long term congestion• it can queue packets to respond short term congestion e.g by employing AQM

• the packet drop algorithm must be insensitive to short term traffic characteristics of micro-flows using an AF class

• micro-flows with same long term rate but different short term burst shapes receive the same loss probability• it should treat packets at same drop precedence within an AF class equally

• this means that the loss probability of a micro-flow depends upon its share of bandwidth within that drop precedence level

AF PHB – Classes

DropPrecedence

Class 1 Class 2 Class 3 Class 4

Low AF11 AF21 AF31 AF41Medium AF12 AF22 AF32 AF42

High AF13 AF23 AF33 AF43Highest Discard

Lowest Discard

Olympic Service

DropPrecedence Gold Bronze Silver

Low AF11 AF21 AF31Medium AF12 AF22 AF32

High AF13 AF23 AF33

30

Expedited Forwarding PHB (RFC 3246)❒ Provide low delay, low jitter, and low loss service

❒ It ensures that an aggregate is served at a certain configured rate

❒ Dominant cause of delay in internet is propagation delay and queuing

delay

❒ Queuing delay introduces delay variation or jitter to packet forwarding

❒ Jitter is defined as variation between maximum and minimum delay

❒ The packets suitably marked to receive EF PHB treatment in the

network encounters short or empty queue

❒ As a corollary effect short or empty queue causes low losses

❒ To ensure short queue encountered by EF packets

1. The service rate at a given output interface offered to the EF packets must exceed their arrival rate at that interface for long and short time intervals

2. Independent of the load of other traffic

EF PHB❒ The EF PHB

❒ ensures that each receives guaranteed service at or above configured rate

❒ provides means to quantify the accuracy with which the configured service rate is delivered over any time interval

❒ also provides a means to quantify the maximum delay and jitter that a packet may experience under bounded operating conditions

❒ Designed to emulate a “leased line”

❍ Packets Egress the node with minimal delay

❒ Preempts Lower Priority Traffic Classes

❒ Uses Priority Queuing

❒ Drops packets that exceed maximum configured rate