Survey
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
TransiNet 2002 Workshop Heinrich-Hertz-Institut Berlin 08. 10. 2002 Switching Optical Packets Motivation and Prospects Dipl.-Ing. Thomas Fischer TU München - Institute of Communication Networks Dipl.-Ing. Martin Maier Dipl.-Inform. Hagen Woesner TU Berlin - Telecommunication Networks Group Outline • Motivation • Difficulties with OPS • IP traffic characteristics • Overhead • Implications • OPS core node • Architecture, complexity & scalability • OPS metro node (distributed router concept) • Architecture, medium access, fairness in the TCP context • Conclusion Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Why use OPS: Motivation • Efficiency • Statistical multiplexing - within the optical network layer • Improved utilization of wavelength channels • Flexibility • Variable bandwidth share - packet duration and/or rate • Bitrate transparency - variable line bitrate for legacy support • Inherent support • Dynamic traffic • Bitrate asymmetry - bidirectional unicast/multicast Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] IP traffic characteristics • Packet arrivals • long-range dependent - correlated across distant periods ⇒ persisting deviations from mean arrival rate (burstiness) • self-similar - similar burstiness across time scales ⇒ compromises statistical multiplexing gain Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] IP traffic characteristics • Packet lengths • IP protocol accounts for vast majority of Internet traffic • Small packets, for the most part from Ethernet LANs, are predominant • 40% of packets ~40 Bytes long • 40 Bytes translates into 32 ns @ 10 Gb/s ⇒ multitude of switching operations Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Overhead • A matter of principle & chromatic dispersion • Destination information • Header synchronization/recognition • Header-payload alignment (slow & fast jitter) • Synchronization trade-off (slotted systems) Time Slot Guard Time Payload Workshop 2002 Berlin Sync Optical Guard Sync Pattern Label Time Pattern Hagen Woesner [email protected] Optical Label Sync Guard Pattern Time Martin Maier Thomas Fischer [email protected] [email protected] Implications (1) • OPS is statistically constrained ⇒ pure loss systems won‘t do Loss System Bursty Packet Arrivals Packet Loss Buffering Statistical Multiplexing < 1% Source Traffic Control Multicasting Aggregation Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Implications (2) • Without further measures, OPS is inherently inefficient ⇒ increase utilization by aggregation Switching Principle Jitter Overhead Short Packet Lenghts High Bitrate Efficiency Loss Many Switching Ops Aggregation Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Scalable OPS core node (1) • Complexity reduction • Scalability improvement O E OE Switch Switch Control Control before after F·W F Taps F·W - Delay Lines (Tunable) - F Delay Lines (Fixed) F 2·F WDM Demultiplexers F·W F·W Receivers (Fixed) F·W F Transmitters (Fixed) F·W - 2x1 Couplers F F WDM Multiplexers W - FxF Switches - W (F-1)x(F-1) Switches Workshop 2002 Berlin E O EO Hagen Woesner [email protected] O E OE E O EO Martin Maier Thomas Fischer [email protected] [email protected] Scalable OPS core node (2) • Unslotted instead of slotted transmission O E OE Switch Switch Control Control • Reduces throughput • Renders packet alignment unnecessary • E O EO O E OE E O EO Ingress components: before after F·W - Delay Lines (Tunable) - F·W Delay Lines (Fixed) Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Scalable OPS core node (3) • Simplification within ingress stage (cost trade-off) O E OE Switch Switch Control Control • Doubles WDM demultiplexer count • Yields „one-for-many“ power tapping & packet retardation E O EO O E OE E O EO • Ingress components : before after F 2·F WDM Demultiplexers F·W F Taps F·W F Delay Lines (Fixed) Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Scalable OPS core node (4) • Using out-band headers • Decrements wavelength channel count O E OE Switch Switch Control Control • Saves many transmitters • Egress components: E O EO O E OE E O EO before after F·W - 2x1 Couplers F·W F Transmitters (Fixed) • Switch matrix: W - FxF Switches - W (F-1)x(F-1) Switches Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Distributed packet switching: Sharing the medium The distributed router concept: •Use the network as the backplane of a big virtual router. •Reduced complexity of the nodes in the network •Shared medium •Medium access protocols •Basically two approaches •Single-hop Networks •Multihop Networks Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Packet Switching in Single-Hop WDM Networks • • • TX and/or RX tunable Sources of inefficiency: • Tuning latency • Chromatic dispersion D • Most MAC protocols for single-hop WDM networks are slotted • Due to different wavelength travel velocities padding required at beginning (or end) of each slot • Padding = D • ∆λ • L = 59.5 ns for 17 ps/nm km, 35 nm, 100 km ⇒ Inefficient packet switching in single-hop WDM networks Multihop WDM networks with fixed-tuned transceivers ⇒ No inefficiency due to tuning latency and chromatic dispersion Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Spatial Reuse Protocol - SRP: Overview • • • • SRP uses a bi-directional dual ring topology Both rings transport data and SRP control packets Insertion buffer MAC protocol Control packets for topology discovery, protection switching and bandwidth control. Outer Ring Data Layer 3 Switch/Router Outer Ring Control Control Packets Inner Ring Data Inner Ring Control Low Pri Rx Data Queue Queues SRP MAC Transit Queue Receive Fiber Transmit Fiber Low Pri second ring Workshop 2002 Berlin Hi Pri High Pri Queue Fairness signaling Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Fibre delay line - node architecture • Only one transit buffer of length 1 packet. • Fixed packet size of 9216 byte (= 7 µs @ 10Gbit/s = 1.4 km FDL) • Aggregation of incoming IP packets (VOQ is an option) • Thresholds for transmit queues. • Definition of fairness algorithm depending on thresholds and counters. From the ring 10% Header extraction Fibre delay line Low priority transmit data queue To the ring LP HP Hi priority transmitReceive data queue queue • Calculation of counters for my_usage and fd_rate in MAC layer. Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] How the protocol works • A node is only allowed to transmit when its FDL is empty. • CSMA ring access • Data rate is controlled by token bucket algorithm • Variable parameters r (fill rate) and b (# of tokens) which are function of my_usage and allowed_usage • Threshold in the low priority transmit queue installed and observed. • If the LP-threshold is reached, • the MAC layer has to tell his upstream node to reduce the traffic sourced onto the ring. • Send the long term average of my_usage counter • Upstream node reduces its rate to the value received Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Why is Fairness needed? • Given a 5- node dual ring architecture • Only nodes 0 and 4 get their data through • Exponential On/Off traffic with 50% 40 byte, 30% 576 byte and 20 % 1500 byte long packet Outer Ring Data 1 2 Outer Ring Control 3 Inner Ring Data 0 4 Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Fairness for unidirectional traffic Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Unfairness for TCP! Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Reasons and consequences • TCP is a dynamic system! • Slow start algorithm initially transmits only 1 (or 2) fragments • Node will never be “congested”, since the threshold will not be reached • Move towards a timer based solution • HOL (head-of-line) timer decides when the node is “congested” • Timer setting depending on the own average traffic load HOL_timer= 1/(lp_my_usage+ε) -1 Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Fairness for TCP Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected] Conclusions • Reduction of switching complexity • Discarding non-vital - and smart use of the remaining functionality, or • Distributing functionality over the network • Switch & protocol design with underlying traffic in mind • OPS nodes alone will not suffice • Development of link layer protocols may lead to design errors without bearing TCP in mind • Destination stripping networks offer spatial reuse • E.g. bi-directional rings • But: Fairness becomes an issue • Global and local fairness Workshop 2002 Berlin Hagen Woesner [email protected] Martin Maier Thomas Fischer [email protected] [email protected]