Switch Datapath in the Stanford Phictious Optical Router SPOR

1
Switch Datapath in the Stanford Phictious
Optical Router (SPOR)

• H. Volkan Demir, Micah Yairi, Vijit Sabnis
• Arpan Shah, Azita Emami, Hossein Kakavand,
  Kyoungsik Yu, Paulina Kuo, Uma Srinivasan
• Optics and Routing Seminar (ORS) June 12, 2001
• Stanford University, Stanford, CA

2
Outline

• What are we going to be working on?
• What is the bigger picture (SPOR)?
• How does optics fit into SPOR?
• What are the optics related limitations?
• What are the flexible points in the design?
• Proposed routing schemes
• Conclusions

3
SPOR Specifications

• IP router (by 2005)
• is a box with a certain number of input and
  output ports that can map (route) any input to
  any output
• sits in a digital, on/off keyed, optical network
• performs packet switching (bits come in packets
  with destination addresses)
• with an aggregate data bit rate of 100Tb/s
• with a reasonable latency between the input and
  output ports

4
When does optics come to play?

• Tasks at the required minimum level
• receive packet (with minimal IP packet
  processing)
• detect destination address
• schedule and configure connection for packet
• transmit packet
• Physical components
• ingress line cards (RXE) ILC
• egress line cards (ETX) ELC
• arbiter (E) A
• interconnect (E vs. O) XC
• Optics
• We would like to use optics for interconnect (TX
  on ILC OXC RX on ELC)
• Number of components (hence number of
  connections) increases with increasing aggregate
  bit rate
• space problem
• distance problem (100s of feet)
• When does it make sense to use optics as
  switching medium? (Electrically controlled
  optical routing)
• If so, what do we win?

5
SPOR Further specifications

• Delay is OK
• Make packets wait (buffer packets) until
  scheduling is done, and switch datapath is
  configured
• How much routing delay is reasonable? ms?
• Buffering ns
• Arbitration ns-ms
• 100 Tb/s aggregate bit rate
• Linecards
• 625 of them each with one fiber at 160Gb/s, or
  each with 4 fibers at 40 Gb/s
• Fibers
• TDM
• In a time slot of 6.25 ps, put 16
  time-multiplexed channel, each at 10Gb/s
• In a time slot of 25 ps, put 4 time-multiplexed
  channel, each at 10Gb/s
• WDM
• put 16 different wavelengths, each at 10Gb/s
• put 4 different wavelengths, each at 10Gb/s
• Presently 625 linecards each with one fiber at
  160Gb/s, each fiber with 16 ? each at 10Gb/s
• In future possibly 625 linecards each with one
  fiber at 160Gb/s, each fiber with 16 ? each at
  40Gb/s

6
Optics related limitations -I

• TX on ILC OXC RX on ELC
• cost
• fabrication, packaging, integration
• How much does it cost with electronics (assuming
  XC is done)?
• power budget
• What is reasonable total electrical power
  consumption?
• space
• Will attempt to quantify these parameters

7
Optics related limitations-II

• TX on ILC
• Feasible bit rate is limited
• externally modulated laser diode presently at
  10Gb/s, likely at 40Gb/s by 2005, perhaps
  160Gb/s in far future
• directly modulated (gain switched) laser diode
  slower
• mode locked laser diode difficult
• From each ILC, use 16 wavelengths each at 10Gb/s
• Number of available wavelengths is limited
• tunable laser 100 (with 50GHz spacing)
• separate laser diodes not easy
• If more required, use multiple (serial or
  parallel) stages

8
Optics related limitations-III

• OXC
• Number of (passive or active) optical channels
  may be limited by
• loss
• crosstalk
• Configuration time of (active) optical channels
  is finite
• MEMS ?s-ms
• Electroholography ns
• Tunable laser ns-ms
• Other means?
• Compare against electronics ns
• If faster configuration required, use multi
  switching layers or units

9
Optics related flexible points-I

• TX on ILC OXC RX on ELC
• optical power management
• dispersion management
• comparatively easier to design
• both ends of the link are internally operated
• link is short distance (100s of feet)

10
Optics related flexible points-II

• Routing speed
• Bursty routing
• Slower XC reconfiguration time can be compensated
  for by
• faster bit rate than minimally required
• higher aggregate bit rate (e.g., through multi
  wavelength channel) than minimal
• larger chunks of data to be routed at a time
  (longer packets, more than one packet)
• To handle 160Gb/s out of a linecard, we need to
  route
  20 B every 1ns, 20 kB
  every 1us, or 20 MB every 1ms
• Parallel routing (load balancing) through
  multiple routing layers (Prof McKeown)
• Decompose routing into many routing layers, each
  of which is reconfigured separately by different
  arbiters
• requires bit rate to be at least twice as fast
• reduces required routing speed by 1/N (N
  number of routing layers)
• Multiple planes (or units) in one routing layer
• Decompose one routing layer into multiple planes,
  one of which is actively used for routing at a
  time while the others are being reconfigured
• reduces routing speed to package rate
• a pair of MEMS planes
• several (tunable) laser diodes (each with
  different wavelength range)
• Combinations

11
Optics related flexible points-III

• Routing scheme
• We can choose where the decision (which node is
  the packet to be routed to) is introduced to the
  link
• spatially reconfigurable link XC is reconfigured
• wavelength sensitive link TX is reconfigured
• time sensitive link RX is reconfigured
• Make use of decomposition in the domains of
• Space (in 2D, or several 2D planes in 3D)
• Optical wavelength (carrier frequency)
• Time
• Or use multiple dimensions?
• Make use of uniting in
• Wavelength
• Time
• Space

12
Proposed routing schemes

• One extreme
• Electrical switching fabric with optical
  interconnect
• sets a baseline
• The other extreme
• Electrically controlled optical switching fabric
• There is a cross-over from electrical to optical
  routing as bit rate demand increases
• Cross-over point moves as both electrical and
  optical technologies improve
• Martin Zirngibl from Lucent Technologies
• Presently the cross-over point is 1 Tb/s
• If optical interconnect is required, electrical
  switching fabric has cost disadvantage (3 pairs
  of RX-TX needed)
• Hybrid scheme?
• Can we combine optical and electrical switching
  fabrics to use advantages of each?
• A central optical fabric (switched at a slow
  rate) that connects the surrounding electrical
  switching fabrics (Prof Horowitz)

13
Proposed optical routing scheme

• Wavelength routing
• TX on ILC
• tunable laser diode or array of laser diodes
• sample grated DBR-LD
• array of monolithically integrated DFB-LD
• array of VCSELs ...
• coupler or wavelength multiplexer
• external modulator to code data
• OXC
• wavelength sensitive
• array waveguide grating (AWG)
• electroholographic switches (by Trellis
  Photonics) ...
• RX on ELC
• may need wavelength sensitive RX
• wavelength demultiplexer RX for each wavelength
• resonator photodiodes
• photodiodes fiber grating circulator ...

14
Trellis Photonics Electroholographic Switch Matrix

• Bragg mirror (sensitive to a specific wavelength)
  is holographically written
• Bragg mirror is turned-on with voltage
• hundreds of volts
• in ns

15
Proposed optical routing scheme (conted)

• Space routing
• Micro electrical mechanical routing
• Prof Solgard
• TX(VCSEL) and RX directly-attached (e.g., through
  flip-chip bonding)
• Large number of interconnects?

16
Conclusions

• Need to improve our understanding
• Optical Internet-Next Generation by SNRC
• Next Generation Internet-Supernet by DARPA
• None of the optical routing schemes seems easy
• Need to work out backbone of each design to get
  an idea on which one of them are realistic
  approaches
• Need to compare against electrical switch