Agenda (Rev 1)

Week 1: Internet History and Basic ConceptsWeek 2: Routing vs. SwitchingWeek 3: Architecture and Topology TrendsWeek 4: Performance, Congestion ControlWeek 5: Multimedia Support, ATM vs. IPWeek 6: Routing part 1 (Intro, RIP, OSPF)Week 7: Routing part 2 (BGP, state of the Internet)Week 8: Guest lectures: Greg Minshall, and ??Week 9: Failure Modes and Fault DiagnosisWeek 10: Product evaluation criteria

Loose Ends...

• RTP vs. UDP

• Enet framing: postamble byte

• Token Ring vs. Ethernet Reliability

• Repeaters = Hubs = Layer 1 or 2?

Week 3: Architecture & Topology Trends

• Focus on Campus/Enterprise networks

• Use UW network as case study

• Introduce DNS and DHCP

• Continue to examine design issues/choices

Technology/Usage Trends

• TCP/IP

• Switching & point-to-point links

• Multimedia

• Desktop web servers

• Push publishing

• Web caching

• Non-locality of reference

Backbone Design Issues

• Link Technology & Topology• Routers vs. Switches vs. ATM• Single vs. Multiprotocol• Central vs. In-building Routers• Low-Density vs. High-Density Routers• Large vs. Small Subnets• Address Management• Redundancy

Core Network Elementseverything except the end-systems

• Name Resolution

• Host Configuration

• Multimedia Support

• Data Transport

• Data Caching??

• Management

Name Resolution

• DNS = Domain Name System

• Distributed, hierarchical directory service

• Maps host/service names to IP addresses

• Resiliency requires client failover

• Susceptible to bad data in root servers

• Growth of .com domain triggered crisis

• Need: security and dynamic update

Host Configuration

• RARP

• BOOTP (and variants)

• DHCP

Problems with DHCP

• Client bugs leading to duplicate addresses

• Scaling

• Redundancy

• Conflict with desktop server trend

• Conflict with network management needs

• How long should the leases be?

Data (Web) Caching

• Important for improving web performance

• Resiliency requires client failover

• Scalability requires server-server protocol

• ICP = Internet Caching Protocol

• Legal & Economic issues: copying & click-thrus

Data Transport

• Getting bits from A to B

• But how fast? How well?

• Not just unicast

Multimedia Support

• Multicast

• QoS = Quality of Service

• Performance = Speed + QoS

• Is QoS important if you have enough bandwidth?

Performance Elements

• Client Machine/Software– Computer-to-Closet– Closet-to-BDF– BDF-to-Router– Router-to-Router

• Server Machine/Software

Network CoreEnd-EndSystem

High-Speed Technologies• 100 Mbps

– FDDI (MTU=4500)– 100VG (MTU=1500)– 100BaseT (MTU=1500)

• 155 Mbps– PPP over SONET OC3c– ATM over SONET OC3c

• 1000 Mbps– PPP over SONET OC24 or 48– ATM over SONET OC24 or 48– Gigabit Ethernet– HIPPI, Fiber Channel

Ethernet Performance Levels

• 10 Mbps– Shared– Dedicated (= Switched)– Dedicated Full-Duplex

• 100 Mbps– Shared– Dedicated (= Switched)– Dedicated Full-Duplex

• 1000 Mbps– Shared– Dedicated (= Switched)– Dedicated Full-Duplex

ATM Performance Levels

• 25 Mbps

• 155 Mbps (OC3)

• 622 Mbps (OC12)

• 1244 Mbps (OC24)

• 2488 Mbps (OC48)

Next-step Desktop Connectivity

• Switched 10 (Half Duplex)• Shared 100 (Half Duplex)• Switched 100

• Would you rather have switched 10 or shared

100?• What are the implications of each on the

backbone?

Case Study: UW’s Campus Network

• The Problem

• History

• Growth

• Key Decisions

• Topology Evolution

• Future Choices

UW’s Network Problem “Death of the net predicted; film at eleven”

• More users

• More usage

• More demanding applications

• More bad guys

• Apparent slow-downs due to net congestion

• Delays still spotty, but expected to worsen

More Demanding Applications

• Non-interactive: email

• Baseline interactive: telnet, web

• Multimedia: desktop conferencing, VOD

• High-end: Medical imaging, VR

Scaling Considerations

• Where do we feel the pressure from increasing use?

– Performance (Speed + QoS)– Address Management– End-user Support

UW Network History

• 1988: five anti-interoperable campus nets...– 3,000 machines on a bridged Ethernet– A large Micom terminal network– Separate library, hospital, and administrative nets

• 1997: one campus net with...– 12,000 PCs– 6,000 Macs– 4,000 Unix workstations– 3,000 X terminals– 1,000 hubs, routers

UW Node Growth

• By 12/94 we had 17,000 nodes and 650 modems

• By 12/95 we had 22,000 nodes and 1,300 modems

• By 12/97 we had 27,000 nodes and 1,500 modems

• Run-rate had been 3k/year nodes, now flat… > Saturation at last??

UW Backbone Traffic

0

50

100

150

200

250

300

1990 1991 1992 1993 1994 1995 1996

Billions of Bytes eachNovember

UW Key Decisions

• Use Internet standards (Interoperate!)

• Route only IP (Simplify!)

• Use lots of 10BaseT Ethernet (Cheap!)

• Use multiple links (Redundancy, loadsharing)

• Use lots of subnets (Isolate Faults)

• Use lots of switches (Isolate Traffic)

• Use DHCP (Automate!)

UW Topology Evolution

• Epoch 1 (c. 1989): Dual Shared Ethernet Cables

• Epoch 2 (c. 1992 ): Dual Routers

• Epoch 3 (c. 1995): Quad Ethernet Switches

• Epoch 4 (c. 1997): Quad Fast Ethernet Switches

UW Current Backbone Topology

S1 S2 S3 S4

R1 R2 R38 R39 R40

To Building Subnets …

UW Building Infrastructure

To Router Center

UW Future Topology Choices

• Ring?

• Mesh?

• Continue with Hierarchy?

Should we…

• Use conventional routers?

• User “layer 3 switches”?

• Use edge routers, ATM core?

• Use Ipsilon IP switching?

• Use 3Com VLANs & Fast IP architecture?

• Use Cisco Tag switching?

Where to put Layer 3 Functionality?

• Edges, nearest the end-systems

• In each Building Distribution Frame

• Centrally, at/near top of hierarchy

• One arm router between VLANs

Decision Criteria

• Interoperability

• Reliability

• Performance

• Fault Tolerance

• Simplicity/Manageability

• Cost

Conclusions?

• Simplify!

• IP Rules!

• Ethernet simpler/cheaper than ATM

• Adequate Frame-based QoS still a question

• Avoid *having* to upgrade end-systems

• Caching becoming part of the network