understanding networked applications: a first course chapter 18 by david g. messerschmitt

Understanding Networked Applications:A First Course

Chapter 18

by

David G. Messerschmitt

Understanding Networked Applications A First Course2

Objectives

• Major functions of the network

• Impact of network on application performance

• Collective issues in networking


Impact of network on applications

• Communication service provided the application– What does application have to do for itself?

• Impact on application performance– Message (packet) latency– Message (packet) loss– Message (packet) corruption


Internetworking (IP)

Subnetworks

Datagram (UDP) Bytestream session (TCP)

Remote method invocation (RMI)

Application

Middleware

OS

Network

Internet protocol architecture


Hosts

Switches

Backbonelinks

Access links

Network topology


Properties of a communication link

010101110101010 0000101101101100110110

Bits waiting to be sent

Bits that have arrived

Bits in transit

Bitrate = number transmitted per second(Bitrate is sometimes called “bandwidth”)Propagation delay = input-output delay experienced by each bit


Sending packets on a link

0101101011010011011010110101010101011010

Packet 1 Packet 2 Packet 3

There must be some way (bit patterns) for the receiver to distinguish packets

Bitrate > # packets per sec x average size of a packet


Messages on threeincoming links

Singleoutgoinglink

Queues

Statistical multiplexing

What are some other examples of statistical sharing?


Congestion

• Cause: irregular packet arrivals, and irregular packet length

• Direct effect: sometimes more packets have to wait in queues for space on output link– In severe cases, packets discarded

• Indirect effect: waiting time in queues contributes to packet (and hence message) latency


Packet structure

• Header– Indicate beginning of packet– Destination address for forwarding– Other information specified by network

protocols

• Payload– Unstructured data to be delivered to application

Header Payload


Packet switch

Addr OutLink

Output link

Routing table

Packet

Packet forwarding


Two basic functions in packet switching

• Packet forwarding– Transmitting each packet on the appropriate

output link– Based on routing table

• Routing– Updating the routing table– Objective: each packet gets closer to destination

via less congested links


Routing

• Full route not written down• Each packet switch has >1 output link• Routing table :

• Reducing table size– use wildcards: 141.211.* next hop is X

Final Destination Next hop141.211.203.32 X207.75.186.1 Y


Routing function

• Switches (routers) talk to each other– “I’m now accepting traffic for 141.211.*”

• Called “advertising a route”

– “Is anyone accepting traffic for 141.211.*?”

• Update entries in own routing table

• Sophistication is in deciding whether to accept traffic


Routing Dangers

• Long routes• Circular routes

– HopLimit limits damage

– Decrement HopLimit at each router

– Discard packet if HopLimit=0

• Hijacking routes– Advertise a route, but don’t deliver

• Route flapping– Frequent updates to routing table


Packet delivery not guaranteed

• Intermediate switch not responding– Temporary malfunction– Queues full (congestion)

• Bad routing– Too long or circular

• Corruption of packet, especially header• Network does not inform sender

– Network may not know!


Why packets?

• Fairness: short message doesn’t have to wait for long message– Allows messages to be sent concurrently– Reduces statistical waiting time

• Store and forward delay reduced

• Data garbling may necessitate resending only a packet, not a whole message


IP Addresses• Every host gets a distinct address

– Can be dynamically assigned

• IPv4 (current standard)– Each address 32 bits– Divide into 8 bit segments– Example: 141.211.203.32– 4 billion addresses

• IPv6 (future standard)– Each address 128 bits– ~1500 addresses per square foot

Recall three ways of locating something:

• Address

• Name

• Reference


Domain Name Service

• IP addresses are inconvenient for people– 32 bits hard to remember– 128 bits very hard to remember

• Domain names– e.g. www.sims.berkeley.edu

• Domain Name Service (DNS)– get an IP address from a domain name


Hierarchy in location

• Addresses hierarchical in topology– Maximize “wild cards” and distribute address

administration

• Names hierarchical in administration– Single administered organizations often

distributed topologically (e.g. ibm.com)

• DNS decouples these two issues


Simulcast Multicast

Forms of broadcast

Multicast requires more sophisticated addressing and routing within the network


Producer Consumer

Queue

Streamof messages

Flow control

Flow control

Normally the producer determines what information is sent, but consumer has to have a way to slow down producer


Network congestion

• Fluctuations of traffic can result in overloads in given network links– Failure of statistical multiplexing

• Analogous to processing congestion of a server, except resent packets can make the problem worse

• Congestion must be limited in some fashion


Carriedtraffic

Offeredtraffic

Network “capacity”

Increasing portion ofnetwork traffic is resentpackets

Socialoptimum

Congestion instability


Questions to address

• What social objectives should a congestion control method achieve?

• What technical approaches are available to achieve those objectives?

• What is the cost of those approaches?

• Ultimately, what will the customer pay for congestion control, and how does that relate to the value received?


Congestion control methods

• Over-provisioning of facilities (mitigation, not control)

• Network initiated– Network source flow control, or

– Source notification and policies, policing, or pricing incentives, or

– Admission control for sessions

• Sources initiated– Source detects congestion (necessary resent packets is one

method), and

– Voluntary or mandatory policiesExamples of each?

Understanding Networked Applications:A First Course

Quality of service (QoS)

by

David G. Messerschmitt


QoS attributes of a packet

• Packet latency– Time until packet delivered at destination

– Transmission time, propagation time, queuing delay, processing time

• Packet loss• Packet corruption

– Payload only

– Normally network will not deliver corrupt packet


Transport services

• “Raw” packet service is not what is needed by most applications

• Transport services “condition” packet service by adding layers– Reliable delivery

– Message service

– Session

– Time stamps

– etc


Audiocoder

Audiodecoder

End-to-enddelay

Stream ofpackets

Packetlatency

Packet latency affects transport service QoS


Reliable delivery

• Add acknowledgement for each packet

• Lost packet can be detected by missing ACK

• Lost packet can be retransmitted

• Tradeoff:– Reliable delivery for greater latency– Latency-sensitive applications must abandon

reliable delivery (e.g. remote conferencing)


QoS Guarantees

• Source and network enter “session contract”:– Source promises not to exceed specified traffic

parameters for that session• Rate and burstiness

– Network promises to limit impairments such as latency, loss, and corruption


Achieving QoS

• TCP/IP offers only best-effort– Every connection gets “best-effort” service

• Achieving maximum latency guarantees– Reserve resources

– Or attach priorities to packets• Contract may allow network to delay or discard low-priority

packets when necessary

– Application may guarantee traffic “shape”• e.g., steady flow rather than bursts


Pricing and Accounting

• What’s the incentive for using low quality service?– Why should user accept greater latency if less is

an option?– Why should application try to minimize

bandwidth, or shape its traffic?– Answer is good citizenship, or pricing

incentives


Pricing Today

• Commercial services: usually flat rate plus connect time (but no per-bit charges)– e.g., CompuServe, Prodigy, America On-Line

• Internet: flat rate, unlimited usage– Resellers can charge for connect time– Many people have unlimited use through a

university or company


Pricing options

Price component Cost recovery

Fixed Capital/operational costs of accessnetwork

Usage Shared costs of backbone network

QoS Cost of reduced statisticalmultiplexing advantage

Congestion Impact on other users, cost ofupgrading facilities


Congestion Pricing Rationale• The fixed cost of building the network is high• The marginal cost of accepting new session is

nearly zero (assuming not congested)• Economic efficiency: use whenever marginal

benefit > 0– But capacity is fixed (in the short run)

– Person A’s use may reduce B’s quality of service


Congestion Pricing

• If network is underutilized, charge nothing• If network is congested

– Charge person A the amount of B’s lost value

– A will drop out unless his value of use is greater than B’s lost value

• Theorem: this raises enough money to expand the network by the socially optimal amount

• Monitoring and billing overhead


Big advantage of pricing

• Congestion pricing uses incentives rather than forced control or policies to affect user/application behavior– Market mechanism like other goods and services

• User/application can determine freely and independently whether use of network during periods of congestion is warranted

• Major objection is cost of monitoring and billing

understanding networked applications: a first course chapter 18 by david g. messerschmitt

Documents

packet slide

x slide

irregular packet arrivals

packet destination address

course chapter

networking slide

messerschmitt slide

packet structure header