Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 1 Review of Networking and Design Concepts (I) Based in part upon slides of Prof. Raj Jain (OSU), S. Keshav (Cornell), L. Peterson (Princeton), J. Kurose (U Mass) Or Shivkumar Kalyanaraman Rensselaer Polytechnic Institute

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 2 q Connectivity: q direct (pt-pt, N-users), q indirect (switched, inter-networked) q Concepts: Topologies, Framing, Multiplexing, Flow/Error Control, Reliability, Multiple-access, Circuit/Packet- switching, Addressing/routing, Congestion control q Data link/MAC layer: q SLIP, PPP, LAN technologies … q Interconnection Devices q Chapter 1,2,11 in Doug Comer book q Reading: Saltzer, Reed, Clark: "End-to-End arguments in System Design""End-to-End arguments in System Design" q Reading: Clark: "The Design Philosophy of the DARPA Internet Protocols":"The Design Philosophy of the DARPA Internet Protocols": q Reading: RFC 2775: Internet Transparency: In HTMLIn HTML Overview

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 3 Connectivity... q Building Blocks q links: coax cable, optical fiber... q nodes: general-purpose workstations... q Direct connectivity: q point-to-point q multiple access

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 4 Connectivity… (Continued) q Indirect Connectivity q switched networks => switches q inter-networks => routers

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 5 What is “Connectivity” ? q Direct or indirect access to every other node in the network q Connectivity is the magic needed to communicate if you do not have a link. q Tradeoff: Performance characteristics worse!

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 6 Connectivity … q Internet: q Best-effort (no performance guarantees) q Packet-by-packet q A pt-pt link: q Always-connected q Fixed bandwidth q Fixed delay q Zero-jitter

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 7 Point-to-Point Connectivity Issues q Physical layer: coding, modulation etc q Link layer needed if the link is shared bet’n apps; is unreliable; and is used sporadically q No need for protocol concepts like addressing, names, routers, hubs, forwarding, filtering … AB

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 8 Link Layer: Serial IP (SLIP) q Simple: only framing = Flags + byte-stuffing q Compressed headers (CSLIP) for efficiency on low speed links for interactive traffic. q Problems: q Need other end’s IP address a priori (can’t dynamically assign IP addresses) q No “type” field => no multi-protocol encapsulation q No checksum => all errors detected/corrected by higher layer. q RFCs: 1055, 1144

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 9 Link Layer: PPP q Point-to-point protocol q Frame format similar to HDLC q Multi-protocol encapsulation, CRC, dynamic address allocation possible q key fields: flags, protocol, CRC q Asynchronous and synchronous communications possible

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 10 Link Layer: PPP (Continued) q Link and Network Control Protocols (LCP, NCP) for flexible control & peer-peer negotiation q Can be mapped onto low speed (9.6Kbps) and high speed channels (SONET) q RFCs: 1548, 1332

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 11 Reliability Mechanisms q Mechanisms: q Checksum: detects corruption in pkts & acks q ACK: “packet correctly received” q Duplicate ACK: “packet incorrectly received” q Sequence number: identifies packet or ack q 1-bit sequence number used both in forward & reverse channel q Timeout only at sender q Reliability capabilities achieved: q An error-free channel q A forward & reverse channel with bit-errors q Detects duplicates of packets/acks q NAKs eliminated q A forward & reverse channel with packet-errors (loss)

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 12 Stop and Wait Flow Control Data Ack Data t frame t prop  = t prop t frame = Distance/Speed of Signal Frame size /Bit rate = Distance  Bit rate Frame size  Speed of Signal = 1 2  + 1 U= 2t prop +t frame t frame U  Light in vacuum = 300 m/  s Light in fiber = 200 m/  s Electricity = 250 m/  s

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 13 Sliding Window Protocols Data Ack t frame t prop U= Nt frame 2t prop +t frame = N 2  +1 1 if N>2  +1

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 14 List of Issues q Connectivity (direct/indirect) q Pt-Pt connectivity: q Framing q Error control/Reliability (optional) q Flow control (optional)

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 15 Connecting N users: Directly… q Pt-pt: connects only two users directly… q How to connect N users directly ? q What are the costs of each option? q Does this method of connectivity scale ? AB... Full mesh Bus

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 16 Multiplexing vs Have it all q Multiplexing = sharing q Allows system to achieve “economies of scale” q Cost: waiting time (delay), buffer space & loss q Gain: Money ($$) => Overall system costs less Full MeshBus

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 17 Virtualization q The multiplexed shared resource with a level of indirection will seem like a unshared virtual resource! q I.e. Multiplexing + indirection = virtualization q We can “refer” to the virtual resource as if it were the physical resource. q Eg: virtual memory, virtual circuits… q Connectivity: a virtualization created by the Internet! q Indirection requires binding and unbinding… q Eg: use of packets, slots, tokens etc... Physical Bus = AB AB Virtual Pt-Pt Link

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 18 Statistical Multiplexing q Reduce resource requirements (eg: bus capacity) by exploiting statistical knowledge of the system. q Eg: average rate <= service rate <= peak rate q If service rate < average rate, then system becomes unstable!! q First design to ensure system stability!! q Then, for a stable multiplexed system: q Gain = peak rate/service rate. q Cost: buffering, queuing delays, losses. q Useful only if peak rate differs significantly from average rate. q Eg: if traffic is smooth, fixed rate, no need to play games with capacity sizing…

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 19 Stability of a Multiplexed System Average Input Rate > Average Output Rate => system is unstable! How to ensure stability ? 1.Reserve enough capacity so that demand is less than reserved capacity 2.Dynamically detect overload and adapt either the demand or capacity to resolve overload

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 20 What’s a performance tradeoff ? q R=link bandwidth (bps) q L=packet length (bits) q a=average packet arrival rate Traffic intensity = La/R A situation where you cannot get something for nothing! Also known as a zero-sum game.

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 21 What’s a performance tradeoff ? q La/R ~ 0: average queuing delay small q La/R -> 1: delays become large q La/R > 1: average delay infinite (service degrades unboundedly => instability)! Summary: Multiplexing using bus topologies has both direct resource costs and intangible costs like potential instability, buffer/queuing delay.

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 22 Connecting N users: Directly... q Bus: Low cost vs broadcast/collisions, MAC complexity q Full mesh: High cost vs simplicity q New concept: q Address to identify nodes. q Needed if we want the receiver alone to consume the packet!... Full mesh Bus q Problem: Direct connectivity does not “scale”….

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 23 How to build Scalable Networks? q Scaling: system allows the increase of a key parameter. Eg: let N increase… q Inefficiency limits scaling … q Direct connectivity is inefficient & hence does not scale q Mesh: inefficient in terms of # of links q Bus architecture: 1 expensive link, N cheap links. Inefficient in bandwidth use

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 24 Filtering, forwarding … q Filtering: choose a subset of elements from a set q Don’t let information go where its not supposed to… q Filtering => More efficient => more scalable Filtering is the key to efficiency & scaling q Forwarding: actually sending packets to a filtered subset of link/node(s) q Packet sent to one link/node => efficient q Solution: Build nodes which focus on filtering/forwarding and achieve indirect connectivity “switches” & “routers”

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 25 Connecting N users: Indirectly q Star: One-hop path to any node, reliability, forwarding function q “Switch” S can filter and forward! q Switch may forward multiple pkts in parallel for additional efficiency! Star S

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 26 Connecting N users: Indirectly … q Ring: Reliability to link failure, near-minimal links q All nodes need “forwarding” and “filtering” q Sophistication of forward/filter lesser than switch Ring

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 27 Ring Star S Tree Topologies: Indirect Connectivity

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 28 Multi-Access LANs q Hybrid topologies: q Uses directly connected topologies (eg: bus), or q Indirectly connected with simple filtering components (switches, hubs). q Limited scalability due to limited filtering q Medium Access Protocols: q ALOHA, CSMA/CD (Ethernet), Token Ring … q Key: Use a single protocol in network q Concepts: address, forwarding (and forwarding table), bridge, switch, hub, token, medium access control (MAC) protocols

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 29 MAC Protocols: a taxonomy Three broad classes: q Channel Partitioning q divide channel into smaller “pieces” (time slots, frequency) q allocate piece to node for exclusive use q Random Access q allow collisions q “recover” from collisions q “Taking turns”: Token-based q tightly coordinate shared access to avoid collisions Goal: efficient, fair, simple, decentralized

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 30 Channel Partitioning MAC protocols. Eg: TDMA TDMA: time division multiple access q Access to channel in "rounds" q Each station gets fixed length slot (length = pkt trans time) in each round q Unused slots go idle q Example: 6-station LAN, 1,3,4 have pkt, slots 2,5,6 idle

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 31 Review: Multiple Access Protocols q Aloha at University of Hawaii: Transmit whenever you like Worst case utilization = 1/(2e) =18% q CSMA: Carrier Sense Multiple Access Listen before you transmit q CSMA/CD: CSMA with Collision Detection Listen while transmitting. Stop if you hear someone else. q Ethernet uses CSMA/CD. Standardized by IEEE committee.

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 32 10Base5 Ethernet Cabling Rules q Thick coax q Length of the cable is limited to 2.5 km, no more than 4 repeaters between stations  No more than 500 m per segment  “10Base5” 2.5m 500 m Repeater Terminator Transceiver

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 33 10Base5 Cabling Rules (Continued) q No more than 2.5 m between stations q Transceiver cable limited to 50 m 2.5m 500 m Repeater Terminator Transceiver

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 34 Inter-connection Devices q Repeater: Layer 1 (PHY) device that restores data and collision signals: a digital amplifier q Hub: Multi-port repeater + fault detection q Note: broadcast at layer 1 q Bridge: Layer 2 (Data link) device connecting two or more collision domains. q Key: a bridge attempts to filter packets and forward them from one collision domain to the other. q It snoops on passing packets and learns the interface where different hosts are situated, and builds a L2 forwarding table q MAC multicasts propagated throughout “extended LAN.” q Note: Limited filtering intelligence and forwarding capabilities at layer 2

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 35 Interconnection Devices (Continued) q Router: Network layer device. IP, IPX, AppleTalk. Interconnects broadcast domains. q Does not propagate MAC multicasts. q Switch: q Key: has a switch fabric that allows parallel forwarding paths q Layer 2 switch: Multi-port bridge w/ fabric q Layer 3 switch: Router w/ fabric and per-port ASICs These are functions. Packaging varies.

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 36 Interconnection Devices H H B H H Router Extended LAN =Broadcast domain LAN= Collision Domain Network Datalink Physical Transport Router Bridge/Switch Repeater/Hub Gateway Application Network Datalink Physical Transport Application

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 37 Ethernet (IEEE 802) Address Format q 48-bit flat address => no hierarchy to help forwarding q Hierarchy only for administrative/allocation purposes q Assumes that all destinations are (logically) directly connected. q Address structure does not explicitly acknowledge indirect connectivity q => Sophisticated filtering cannot be done! G/L bit (Global/Local) G/I bit (Group/Individual) OUI (Organizationally Unique ID)

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 38 Ethernet (IEEE 802) Address Format q G/L bit: administrative q Global: unique worldwide; assigned by IEEE q Local: Software assigned q G/I: bit: multicast q I: unicast address q G: multicast address. Eg: “To all bridges on this LAN” G/L bit (Global/Local) G/I bit (Group/Individual) OUI (Organizationally Unique ID)

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 39 Ethernet & Frame Format q Ethernet q IEEE Dest. Address Source Address Type 662 Size in bytes Dest. Address Source Address Length Info 662 IPIPXAppleTalk LLC IPIPXAppleTalk CRC 4 4 Pad Length Info Maximum Transmission Unit (MTU) = 1518 bytes Minimum = 64 bytes (due to CSMA/CD issues)

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 40 “Taking Turns” MAC protocols - 1 Channel partitioning MAC protocols: q share channel efficiently at high load q inefficient at low load: delay in channel access, 1/N bandwidth allocated even if only 1 active node! Random access MAC protocols q efficient at low load: single node can fully utilize channel q high load: collision overhead “Taking turns” protocols look for best of both worlds!

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 41 Polling: q Master node “invites” slave nodes to transmit in turn q Request to Send, Clear to Send messages q Concerns: q polling overhead q latency q single point of failure (master) Token passing: q Control token passed from one node to next sequentially. q Token message q Concerns: q token overhead q latency q single point of failure (token) “Taking Turns” MAC protocols - 2

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 42 “Taking Turns” Protocols –3 Reservation-based a.k.a Distributed Polling: q Time divided into slots q Begins with N short reservation slots q reservation slot time equal to channel end-end propagation delay q station with message to send posts reservation q reservation seen by all stations q After reservation slots, message transmissions ordered by known priority

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 43 Additions to List of Issues q Filtering techniques: q Learning, routing q Multiple access q How to share a wire q Partitioning, Random Access, Taking Turns q Switching, bridging, routing q Addressing, Packet Formats

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 44 Inter-Networks: Networks of Networks = Internet … … …… Our goal is to design this black box on the right

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 45 Inter-Networks: Networks of Networks q What is it ? q “Connect many disparate physical networks and make them function as a coordinated unit … ” - Douglas Comer q Many => scale q Disparate => heterogeneity q Result: Universal connectivity! q The inter-network looks like one large switch, q User interface is sub-network independent

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 46 Inter-Networks: Networks of Networks q Internetworking involves two fundamental problems: heterogeneity and scale q Concepts: q Translation, overlays, address & name resolution, fragmentation: to handle heterogeneity q Hierarchical addressing, routing, naming, address allocation, congestion control: to handle scaling q Two broad approaches: circuit-switched and packet- switched

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 47 How to design large inter-networks? Circuit-Switching q Divide link bandwidth into “pieces” q Reserve pieces on successive links and tie them together to form a “circuit” q Map traffic into the reserved circuits q Resources wasted if unused: expensive. – Mapping can be done without “headers”. – Everything inferred from timing.

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 48 How to design large inter-networks? Packet-Switching q Chop up data (not links!) into “packets” q Packets: data + meta- data (header) q “Switch” packets at intermediate nodes q Store-and-forward if bandwidth is not immediately available. Bandwidth division into “pieces” Dedicated allocation Resource reservation

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 49 Packet Switching A B C 10 Mbs Ethernet 1.5 Mbs 45 Mbs D E statistical multiplexing queue of packets waiting for output link  Cost: self-descriptive header per-packet, buffering and delays due to statistical multiplexing at switches.  Need to either reserve resources or dynamically detect and adapt to overload for stability

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 50 Spatial vs Temporal Multiplexing q Spatial multiplexing: Chop up resource into chunks. Eg: bandwidth, cake, circuits… q Temporal multiplexing: resource is shared over time, I.e. queue up jobs and provide access to resource over time. Eg: FIFO queueing, packet switching q Packet switching is designed to exploit both spatial & temporal multiplexing gains, provided performance tradeoffs are acceptable to applications. q Packet switching is potentially more efficient => potentially more scalable than circuit switching !

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 51 Scalable Forwarding, Structured Addresses q Address has structure which aids the forwarding process. q Address assignment is done such that nodes which can be reached without resorting to L3 forwarding have the same prefix (network ID) q A simple comparison of network ID of destination and current network (broadcast domain) identifies whether the destination is “directly” connected q I.e. Reachable through L2 forwarding only q Within L3 forwarding, further structure can aid hierarchical organization of routing domains (because routing algorithms have other scalability issues) Network IDHost ID Demarcator

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 52 Flat vs Structured Addresses q Flat addresses: no structure in them to facilitate scalable routing q Eg: IEEE 802 LAN addresses q Hierarchical addresses: q Network part (prefix) and host part q Helps identify direct or indirectly connected nodes

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 53 i  ii q If information about i, and  is known in a central location where control of i or  can be effected with zero time delays, q the congestion problem is solved! The Congestion Problem 1 n Capacity Demand Problem: demand outstrips available capacity

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 54 The Congestion Problem (Continued) q Problems: q Incomplete information (eg: loss indications) q Distributed solution required q Congestion and control/measurement locations different q Time-varying, heterogeneous time-delay

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 55 Additions to Problem List q Internetworking problems: heterogeneity, scale. q Circuit Switching vs Packet Switching q Heterogeneity: q Overlay model, Translation, Address Resolution, Fragmentation q Scale: q Structured addresses, hierarchical routing q Naming, addressing q Congestion control

Shivkumar Kalyanaraman Rensselaer Polytechnic Institute 56 Summary: Laundry List of Problems q Basics: Direct/indirect connectivity, topologies q Link layer issues: q Framing, Error control, Flow control q Multiple access & Ethernet: q Cabling, Pkt format, Switching, bridging vs routing q Internetworking problems: Naming, addressing, Resolution, fragmentation, congestion control, traffic management, Reliability, Network Management