COMP 6: NETWORKED & DISTRIBUTED SYSTEMS 8//6 COMP 6: NETWORKED & DISTRIBUTED SYSTEMS Packet Switching Jasleen Kaur Fall 6 Packet Switching: Issues Point-to-point links don t scale to large networks Ø Limited geographical coverage, limited number of hosts length constraints, limited ability to resolve access contention Ø Too expensive to connect everyone in a clique Switches: Ø Enable communication between nodes not directly connected Help create a star topology Ø Forward packets from one link to another Provide high aggregate throughput Issues we will look at: Input Ports Ø Forwarding approaches How switch decides on which outgoing port to forward packet Ø Selecting frame size T T STS- Switch Output Ports T T STS- Copyright by Jasleen Kaur
COMP 6: NETWORKED & DISTRIBUTED SYSTEMS 8//6 FORWARDING APPROACHES Virtual Circuits and Datagrams Virtual Circuit Switching A connection-oriented approach Done in two stages Ø Ø Connection set-up (in all switches): [VCI-in, Port-in, VCI-out, Port-out] VCI: Virtual Circuit Identifier (link-local scope) Data transfer (all packets follow same circuit) Cons: set- up delay heavy failure- recovery Pros: Guaranteed resources Small per- packet overhead 5 Switch Host A Vcin VCout PortOut 5 6 8 Circuit Table (switch, port ) Switch 7 Switch 4 Host B 4 Copyright by Jasleen Kaur
COMP 6: NETWORKED & DISTRIBUTED SYSTEMS 8//6 Source Forwarding A connection-less approach Source specifies route to be taken (using headers) Switch Switch Host A Cons: Not scalable Source needs to know complete topology Header length can be infinite Switch Host B 5 Datagram Switching A connection-less approach Every packet has full destination address Ø Switches maintain mapping of (destination outgoing port) Host D Switch Host E Host F Pros: No setup delay Failure resistance Host C Cons: High header overhead No service guarantee Reordering Host A Address Port A C F G Forwarding Table (switch ) Switch Host G Switch Host B Host H 6 Copyright by Jasleen Kaur
COMP 6: NETWORKED & DISTRIBUTED SYSTEMS 8//6 Forwarding Approaches: Summary Issues that differentiate Ø Signaling overhead Ø Robustness to failure Ideas used: Ø Less state Ø Connection-less model What categories do the following fall in? Ø Postal system? Ø Phone system? 7 PACKET SIZES Fixed vs Variable, Small vs Large 8 Copyright by Jasleen Kaur 4
COMP 6: NETWORKED & DISTRIBUTED SYSTEMS 8//6 Packet Size: ATM as a Case Study Variable length or fixed length? Ø Variable length: since no optimal size for fixed length If too large, low utilization for small messages (need padding) If too small, large header overhead high processing cost (per-packet) Ø Fixed length: facilitate fast, scalable hardware implementations Simpler Enables parallel processing implementations ATM networks: use fixed cell sizes 9 Packet Size: But What is the Right Size? Large cells: Ø Better utilization J: smaller header-to-payload ratio L: Wastage due to cell padding Small cells: Ø Improve queuing behavior Fine-grained preemption for high-priority/latency-sensitive traffic e.g.: 4 KB vs. 5 B on a Mbps link (7.68 µs vs. 4.4 µs) Queues tend to be smaller When larger cells arrive simultaneously, time-averaged queue larger Ø Improve packetization latency at source Larger cells è wait longer before constructing & sending cell ATM uses 48 B cells (compromise between B & 64B)! Copyright by Jasleen Kaur 5
COMP 6: NETWORKED & DISTRIBUTED SYSTEMS 8//6 Packet Sizes in the Internet Internet allows variable packet sizes Ø Too much diversity in link-layer technologies (each with different frame sizes) Ø Selecting a universal MSS might prohibit some link layers Optimality may not guide practical decisions Basic Idea: Ø No upper or lower limit on packet sizes Ø If too large for a downstream link, break into smaller chunks & reassemble Fragmentation and Reassembly (more later) Good example of: simplicity of the service model è generality / greater interoperability Copyright by Jasleen Kaur 6