course review: part 1 eecs 122: lecture 29 department of electrical engineering and computer...

Course Review: Part 1 EECS 122: Lecture 29

Department of Electrical Engineering and Computer Sciences

University of California

Berkeley

May 1, 2003 Abhay K. Parekh: EECS122 Lecture 28

2

Today

Cover the main points from the following topics Router Lookup and Scheduling QoS Distributed Algorithms Multicast Overlay Networks Error Correction Physical Layer

Goal: Emphasize what is important For topics before the midterm look at the slides of

the earlier review


3

Review: Switch Architectures

Input Queued Faster switch Two congestion points

HOL

Output Queued Slower switch

Backplane speedup is N One point of congestion

input interface output interface

Backplane

CRO


4

How a router manages traffic

Fast routing table lookups Exact and LPM

How to resolve router congestion? Assume output queued Focus on scheduling

Already covered packet dropping

End-to-End peformance yields QoS

input interface output interface

Backplane

CRO


5

LPM in IP Routers“Patricia” trie

Example Prefixesa) 00001b) 00010c) 00011d) 001e) 0101f) 011g) 100h) 1010i) 1100

j) 11110000

e

f g

h i

j Skip 51000

0 1

a b c

d

Nick Mckeown


6

Fairness

Suppose there are N applications sharing the network Each application j can express its satisfaction with network

performance in terms of Uj(r), where r = (r1,…,rN) is the rate allocated to the applications. Most common utility function is Uj(r) =r for all j

If the user can’t use more than a rate of C then Uj(r) =C for all r≥ C.

What is a fair allocation of the throughputs? Fairness is vague so there are many definitions

Maximize the sum of utilities: May penalize some apps Max-Min Fair

Know how to compute this for applications that traverse multiple links (hw)


7

Network Effects on Traffic from a Single Flow

CumulativeBits

Time

delaySource

Router n

bits in thenetwork

Svc function at router 1 is arrival function at

router 2

Router 1


8

How to control delay?

1. Overprovision the network Maybe one day for voice and video but not today

2. Circuit Switch

3. Packet switch with more state than a datagram network

Try to be smart about managing limited resources

Per flow? Per aggregate class?


9

Scheduling

Scheduling in conjunction with packet dropping control performance within a router

Scheduling mechanisms determine how the bandwidth of an output port is shared Mainly used to manage

delay Scheduling is only

effective if packet buffers are large enough to matter

R(f1)

A(t) D(t)

Model of router queues

R(f2)

R(fM)

A1(t)

A2(t)

AM(t)

SchedulingDiscipline


10

Generalized Processor Sharing Each class j has a weight wj If there is traffic to be served from class j at time t

we say that it is Backloged at time t. B(t) is the set of backlogged classes at time t W(t) is the total weight of the backlogged classes at time t

If there is even one backlogged class, the server operates at rate C This is called being work conserving

Serve the classes in proportion to their weights If class j is backlogged at time t, give it service rate

Sj(t) = [wj/W(t)] * C GPS desirable properties


11

Generalized Processor Sharing

Delay =4

Delay =5

B(2)={1,2}B(4)={2}

Class 1 Class 2 w1 = w2


12

Adjust weights to change delay

Delay =5Delay =4

B(2)={1,2}B(4)={1}

Class 1 Class 2 3w1 = w2


13

Weighted Fair Queueing

Work Conserving Tracks GPS

Simulates the arrivals in real-time in a GPS system.

Serves entire packets (only one class at a time) Let Fp be the time that the last bit of packet p

departs the simulated GPS system WFQ attempts to serve packets in order of Fp

Also called “Packetized Generalized Processor Sharing” (PGPS)


14

QoS

QoS Building Blocks: Application describes required level of

QoS (Token Buckets) Resource Reservation (RSVP) Per Hop (Router) Mechanisms:

What efforts are underway to move towards QoS in the internet? IntSrv DiffSrv


15

The Token Bucket

Tokensat rate,

Token bucketsize,

Packet buffer

Packets Packets

One byte (or packet) per

token

ρ: average rateσ: burstiness

Nick Mckeown

A(s,t) ≤σ+ρ(t-s)


16

How the user/flow can conform to the Token Buckets

Tokensat rate,

Token bucketsize

Variable bit-ratecompression

To network

time

bytes

time

bytes

time

bytes

C

A(s,t) ≤σ+ρ(t-s)

Nick Mckeown


17

Result

If all of the routers do per-flow WFQ then explicit worst case end to end delay bounds are possible for a variety of different weighting schemes.

Special Case of small packets (GPS): There are N flows and they are where the ith flow has token bucket (σj,ρj). Set the weight of flow j to ρj at each router. Then the worst case delay for flow j is less than σj/ρj The Backlog Clearing Rate is > ρj at each router


18

Resource Reservation Protocol: RSVP Establishes end-to-end reservations over a datagram network. Sources

Instead of sending to one receiver send to a GROUP RSVP was designed with multicast in mind

Send TSPEC to the group Receivers

Join a “group” associated with the sender Initiate reservations of the network

Network Establishes the group Helps determine the path Learns the TSPEC Responds to receiver-based reservations


19

RSVP Basic Operations

Two message types: PATH and RESV Sender sends PATH message via the data delivery path

TSPEC – use token bucket Set up the path state on each router including the address of

previous hop Receiver sends RESV message on the reverse path

Specify the reservation style, QoS desired Queueing delay and bandwidth requirements

Source traffic characteristics (from PATH) Set up the reservation state at each router


20

Intserv Node Architecture

Admission Control

Data InData Out

Con

trol P

lan

eD

ata

Pla

ne

Scheduler

Routing Routing

MessagesRSVP

messages

Classifier

RSVP

Route Lookup

Forwarding Table Per Flow QoS Table

Kinds of Service:1. Guaranteed2. Controlled Load3. Best Effort


21

Diffserv Architecture

Ingress routers Police/shape traffic Set Differentiated Service Code Point (DSCP) in Diffserv (DS) field

Core routers Implement Per Hop Behavior (PHB) for each DSCP Process packets based on DSCP

IngressEgressEgress

IngressEgressEgress

DS-1 DS-2

Edge router Core router

Its all about the domain

Kinds of Service:1. Premium2. Assured3. Best Effort


22

Comparison

Best-Effort Diffserv Intserv

Service Connectivity

No isolation

No guarantees

Per aggregate isolation

Per aggregate guarantee

Per flow isolation

Per flow guarantee

Service scope

End-to-end Domain End-to-end

Complexity No setup Long term setup Per flow steup

Scalability Highly scalable (nodes maintain only routing state)

Scalable

(edge routers maintains per aggregate state; core routers per class state)

Not scalable (each router maintains per flow state)


23

Distributed Algorithms

Focus on the algorithms behind protocols How to move from Centralized to Distributed Alg. Synchronous and Asynchronous computation

Why does the Asynchronous Bellman Ford converge? What are the effects of a changing topology on

algorithm design? How can protocols be designed to protect against

dishonest nodes


24

Synchronous v/s Asynchronous Algorithms Synchronous algorithms can be described in terms

of global iterations. The time taken for a given iteration is the time taken for the slowest processor to complete that iteration: time driven E.g. TDM or SONET

Asynchronous algorithms execute at a processor based on received messages and internal state: event driven E.g. IP protocols which must run over heterogeneous

systems Study how this applies to Bellman Ford


25

Synchronization Penalty1

34

6

2

5

1

4

1 1

4

2

3

1

1 32

1 32

idle idle idle

Node 1

Node 6

idleidleidle

1 32 Node 5

Slot size is affected by the slow (1,6) link

Penalty can be huge!


26

Local Synchronization1

34

6

2

5

1

4

1 1

4

2

3

1

Send update k after you’ve heard update k-1 from all neighbors.

1 32

1 32

idle idle idle

idleidleidle

Node 3

Node 4

1 32 Node 5


27

Asynchronous computation1

34

6

2

5

1

4

1 1

4

2

3

1

1 2

1 32

idleidleidle

Node 1

Node 6

1 32 Node 5

No notion of “slot size”

What paths will this compute??

3 4 5 6 7 8 9 10 11 12 13 14 15 16


28

Soft State

State with Time-Out Example: A host joins a group by sending a “join” message to a

“host manager”. The manager adds the host to the group for the next T seconds. If the host wants to stay in the group it must send a refresh message within T seconds to the manager. Otherwise it is dropped.

Advantage: Manager robust to host failure Disadvantage: Too many messages Most internet protocols use this way of communicating Trades of simplicity of correctness with complexity of

communication


29

Trustworthiness

Three levels Honest: Always in conformance of the protocol Selfish: May lie to get better performance out of

the protocol (BGP) Malicious: Unpredictable

Internet Protocols (for the most part) assume Honest protocol agents Unreliable infrastructure

Infrastructure has gotten more reliable, and agents have gotten less honest…


30

The Multicast service Model

R0 joins G

R1 joins G

Rn-1 joins G

S

R0

R1

.

.

.

[G, data]

[G, data]

[G, data]

[G, data]

Rn-1

Net


31

Multicast and Layering

Multicast can be implemented at different layers

data link layer e.g. Ethernet multicast

network layer e.g. IP multicast

application layer e.g. as an overlay network like Kazaa

Which layer is best?


32

Routing: Approaches

Kinds of Trees Shared Tree Source Specific Trees

Tree Computation Methods Intradomain Update methods

Build on unicast Link State: MOSPF Build on unicast Distance Vector: DVMRP Protocol Independent: PIM

Interdomain routing: BGMP This is still evolving…


33

What to do if most of the routers in the internet are not multicast enabled?

Tunnel between multicast enabled routers

Creates an “overlay” network but both operate at Level 3…

This is how multicast was first deployed

MBONE

IP

IPM

IPM


34

Core Based Trees (CBT)

Pick a “rendevouz point” for the group called the core.

Shared tree Unicast packet to core and bounce it back to

multicast group Tree construction is receiver-based

Joins can be tunneled if required Only nodes on One tree per group tree involved

Reduce routing table state from O(S x G) to O(G)


35

PIM

Popular intradomain method UUNET streaming using this

Recognizes that most groups are very sparse Why have all of the routers participate in keeping

state? Two modes

Dense mode: flood and prune Sparse mode: Core-based shared tree approach

with a twist


36

Scalable Reliable Multicast (SRM) Randomize NACKs (request repairs) All traffic including request repairs and repairs are

multicast A repair can be sent by any node that heard the

request A node suppresses its request repair if another node

has just sent a request repair for the same data item

A node suppresses a repair if another node has just sent the repair


37

Application Layer Multicast

Provide multicast functionality above the IP unicast Gateway nodes could be the hosts or multicast

gateways in the network Advantages

No multicast dial-tone needed Performance can be optimized to application

Loss, priorities etc. More control over the topology of the tree Easier to monitor and control groups

Disadvantages Scale Performance if just implemented on the hosts (not

gateways)


38

A network defined over another set of networks

The overlay addresses its own nodes Links on one layer are network

segments of lower layers Requires lower layer routing to be

utilized Overlaying mechanism is called

tunneling

A’

Overlay network

A

A


39

Three kinds of Overlays1. Only Hosts: Peer to Peer Networks (P2P)

Example: Gnutella, Napster

2. Only Gateway nodes: Infrastructure Overlays Content Distribution Networks (CDNs)

Example: Akamai

3. Host and Gateway Nodes: Virtual Private Networks

Overlay node structure Regular: Chord, Pastry Adhoc: Gnutella

Kinds of Overlay Networks


40

Two kinds of overlays functions Overlay provides access to distributed resources Overlay facilitates communication among other client applications

Two kinds of client connectivty Direct: P2P Not direct: Akamai

Overlay Network Operations Select Virtual Edges (fast or slow timescales) Overlay Routing Protocol Edge Mapping Resource Location


41

Content Addressable Network (CAN) Associate to each node and item a unique id

in an d-dimensional space Properties

Routing table size O(d) Guarantee that a file is found in at most d*n1/d

steps, where n is the total number of nodes Virtual Topology related to actual topology

using “binning”


42

Bottom line on Overlays

Overlays are an irreversible trend in network Overlays add new functions to the network

infrastructure much faster than by trying to integrate them in the router relying on a infrastructure service provider on deploy the

function Disadvantages

Overlay nodes can create performance bottlenecks New end-to-end protocols may not work since the overlay

nodes don’t understand them Generally better to improve performance by building

an “underlay” and add functionality by building an overlay


43

Link Functions

Functions1. Construct Frame with Error Detection Code2. Encode bit sequence into analog signal3. Transmit bit sequence on a physical medium (Modulation)4. Receive analog signal5. Convert Analog Signal to Bit Sequence6. Recover errors through error correction and/or ARQ

AdaptorAdaptor AdaptorAdaptorSignal

Adaptor: convert bits into physical signal and physical signal back into bits


44

Link Components

NRZI


45

Framing

Goal: send a block of bits (frames) between nodes connected on the same physical media

This service is provided by the data link layer Use a special byte (bit sequence) to mark the

beginning (and the end) of the frame Problem: what happens if this sequence

appears in the data payload?


46

Approaches

Byte/Bit oriented Stuffing v/s frame count Clock-based (Sonet)

Looks at the path not just the link


47

Link Layer Reliable Transmission Prevalent Approach

Detect errors using error detection codes Use retransmission methods (e.g. Go back n)

Alternative Approach Use error correction codes (e.g. perfect parity codes)

Algorithmic challenges: Achieve high link utilization, and low overhead


48

Error detection

The frame consists of a header and payload

If all n-bit strings are valid payload strings, the error detection must occur in the header bits

Break the header into two parts

Error detecting code part contains bits that add redundancy

PayloadHeader

Payload EDCFH

n bits

k bitsn bits


49

Error Detecting Codes

Goals: Reduce overhead, i.e., reduce the number of

redundancy bits Increase the number and the type of bit error

patterns that can be detected Examples:

Even Parity Rectangular Codes Cyclic Redundancy Check (CRC)


50

Hamming Distance

Given codewords A and B, the Hamming distance between them is the number of bits in A that need to be flipped to turn it into B Example H(011101,000000)=4

If all codewords are at least d Hamming distance apart, then up to d-1 bit errors can be detected


51

Error Correction

If all the codewords are at least a hamming distance of 2E+1 apart then up to E bit errors can be corrected Just pick the codeword closest to the one received!

How many bits of EDC are required to correct d errors when there are n bits in the payload?

Example: d=1: Suppose n=3. Then any payload can be transformed into 3 other payload strings (000 into 001, 010 or 100). Need at least two bits in the ECD to differentiate between 4

possibilities In general need at least k ≥ log2(n+1) bits A scheme that is optimal is called a perfect parity code

Perfect Codes


52

Error correction or Retransmission? Error Correction requires a lot of redundancy

Wasteful if errors are unlikely Retransmission strategies are more popular

As links get reliable this is just done at the transport layer

Error correction is useful when retransmission is costly (satellite links, multicast)


53

Noise limits the link rate

Suppose there were no noise and no measurement error E.g. Send s(t) always receive s(t+Δ) Take a message of N bits say b1b2….bN, and send a pulse of

amplitude of size 0.b1b2….bN Can send at an arbitrarily high rate This is true even if the link distorts the signal but in a known way

In practice the signal always gets distorted in an unpredictable (random) way Receiver tries to estimate the effects but this lowers the effective

rate One way to mitigate noise is to jack up the power of the signal Signal to Noise ratio (SNR) measures the extent of the distortion

effects


54

Fundamental Result

The affect of noise on the data is modeled probabilistically. It turns out that there is a maximum possible reliable rate for

most channels called the capacity C: There is a scheme to transmit at C with almost no errors Finding this scheme may be tricky but it exists

For a commonly observed kind of noise called Additive White Gaussian Noise (AWGN) the capacity is given by: C = Wlog2(1 + S/N) bits/sec (Shannon) Example: Voice grade line: S/N = 1000, W=3000, C=30Kbps Technology has improved S/N and W to yield higher speeds such

as 56Kb/s


55

Encoding

Goal: send bits from one node to another node on the same physical media

This service is provided by the physical layer Problem: specify a robust and efficient

encoding scheme to achieve this goal


56

Approaches

Non Return to Zero Non Return to Zero Inverted Manchester 4/5


57

Review Checklist

Router Performance Input/Output Queuing Lookups Scheduling (GPS/WFQ) Fairness

QoS Token Buckets RSVP Intsrv/Diffsrv

Distributed Algorithms Synchronous/Asynchronous

Algorithms Soft State

Multicast MOSPF, DVMRP, RPB, PIM SRM

Overlays Different Types CAN

Link Layer and below Error Detecting Codes Perfect Parity Codes Encoding Framing

course review: part 1 eecs 122: lecture 29 department of electrical engineering and computer...

Documents