Optically Switched Networking Michael Dales Intel Research Cambridge

Overview Part 1 – Technology overview Part 2 – Example network Optical fibre as a connection medium Optical switching fabrics Optical switches Part 2 – Example network SWIFT Architecture overview Current work Research topics

Recommended reading If optical networks turn you on then the following text book is worth seeking: “Optical Networks, A Practical Perspective” by Rajiv Ramaswami and Kumar N. Sivarajan, Morgan Kaufman

Part 1 – Technology overview

Optical fibre links Optical fibre – yet another wire Advantages: Capacity – long haul links of 160 Gbps over a single fibre Range – signal can travel further without regeneration Noise immunity – does not suffer from EM interference Weight/space – a lot lighter/smaller than copper Power – … Popular in the long haul network

Optical fibre links Not all good – some problems: Polarisation sensitivity Chromatic dispersion Non-linear behaviour Fibre more delicate Can’t be thrown around like copper Minimum coil radius Coupling/splitting costs

Optical fibre links In copper we use TDM to multiplex multiple channels on a single link In fibre can also use Wavelength Division Multiplexing (WDM) Each wavelength (lambda, l) can carry a different channel Free extra wires! Can TDM each wavelength too

Switched optical networks Optical links are common in high speed switched networks: ATM, Infiniband, Fibre-channel But all these networks convert data back to electrons at the switch

Switched optical networks O-E-O switch design makes it easy to design an optical network (just like copper ones!) Disadvantages: Size/power – need to duplicate electronics for each lambda Latency – O-E-O conversion takes time Bandwidth – for really high capacity, electronics can become the bottleneck(?)

Optically switched networks A key focus of the optical network community is to find ways to make all optical networks Packets stay in photons from edge to edge Techniques used depend on traffic type – circuit switching and packet switching have very different requirements Might want to move to different wavelength across switch

Optical switch fabrics Switch fabric design covered later in course Here we look at switching elements for light Need a way to switch light from one port to another Many possible ways with varying loss, switching time, polarisation dependency, etc. Mechanical – moveable mirrors Can uses MEMS devices for compactness (e.g., glimmerglass) Thermo-optical – heat it to change Electro-optical – control by current

Buffering? In an electronic switch we use buffering to: Delay packet whilst we decide what to do with it Resolve contention when multiple packets want to go to the same place at the same time There is no optical equivalent of random access memory Best we have are Fibre Delay Lines Use a long loop of fibre to delay the signal for a while

Optical switches The switching fabric is only half the story – how do we decide where to switch the packet? In electronic switch read header and then route through fabric accordingly In optical switches we have three options: Convert the header to electrons and process electronically Process the header optically using optical logic Forget it all and use some form of reservation

Optical switches Use electronics to route packet: Read header from photons and convert to electrons Use a FDL to buffer packet whilst switch makes decision

Optical switches Alternatively use reservation – signal ahead of time that a packet is coming, typically on a reserved l One popular technique is Optical Burst Switching Packets grouped into a burst at source to amortise overhead Control packet fired into network ahead of time – passes through switches setting up a path A fixed-delay time later the burst is sent through network No guarantee that you’ll get through

Optical switches Alternatively use photonic devices to perform optical header reading No need to convert to electrons Still not a prime time technology – can only handle a couple of addressing bits

Part II - Example

SWIFT optical network SWIFT is a research project between Intel Research, University of Cambridge, Essex University, and Intense Photonics Aim to built a short range, high capacity, wavelength striped, optically switched, packet switched network Aim for 100 Gbps and up Use photonic devices under electronic control

SWIFT motivation Optics traditionally used in long haul, but not in short range, where copper dominates… …but copper will eventually run out (eventually…) SWIFT looks at applying optics to short range: Device interconnects, Cluster/supercomputer interconnects, Storage Area Networks, etc. Want have optical data-path, but still use electronics for control

Architecture overview A short range packet switched network based upon: WDM to increase bandwidth per link An all optical data path A single switch for simplicity (for now) An electronic control plane Use WDM for l striping – use all ls for one channel Create a light bus Reserve one channel for control

Overview

Switch design Optical data-path: packets remain optical throughout the network Light-paths need to be constructed through the switch before packets arrive Asynchronous control signalling used to request switch configuration

Switch fabric Many light switching technologies, ranging from mechanical mirrors to semiconductor solutions Switch response time is important for packet switching We use Semiconductor Optical Amplifiers (SOAs) Turn light on or off based on an electrical input Have a switching time of a few nanoseconds

55us/div – Packet is 94.72us data, 1.28 us guard band Switch fabric Demonstrated switching 10 * 10 Gbps channels through an SOA 55us/div – Packet is 94.72us data, 1.28 us guard band

Host interface Host interface has two main tasks Taking packets and converting them to striped format and vice versa Negotiating with the switch for access When a node wishes to transmit it requests permission over the control channel and waits for a light-path to be setup

Wavelength striping Incoming packet From arbiter grant To arbiter Request Incoming packet To optical switch

Demonstrator Have built a test-bed network Goal is to allow practical evaluation at many levels: Photonics evaluation MAC layer testing Real application performance Used to validate a simulation model for investigation of network scaling

Testbed overview Built a 3 node test-bed Two main components: host interfaces and switch Control electronics on FPGAs 2 data l in 1500nm range 1 control l in 1300 nm range Couplers/AWGs used to combine/split ls

Current demonstrator Current setup seen here Three racks: PCs off shot 1: switch 2: host interface board 3: host interface transceivers PCs off shot Large due to using off the shelf components!

Status Recently got first stage working Switches packets between nodes Data striped over both wavelengths Can run TCP, UDP, ICMP, etc. end to end Currently tuning performance for benchmarking Have simulation model in NS2 ready to correlate against testbed

Future work Looking at several areas, including Switch fabric design Photonic device control Current SOA configuration done manually Want to automate this process using electronics Network scheduling and management Improve on request/grant protocol