WavelengthAllocation Strategies in Optically Switched Networks for Avionics

1
Wavelength-Allocation Strategies in Optically
Switched Networks for Avionics

• Casey B. Reardon, John D. Profumo,
• and Alan D. George
• HCS Research Laboratory
• University of Florida

2
Outline

• Introduction
• Background Information
• LION Overview
• OSMOSIS Switch Architecture
• OSMOSIS Case Study
• Network Layout
• Experimental Configuration
• List of Experiments
• Simulation Results
• Latency Statistics
• Analysis
• Conclusions

3
Introduction

• Optical network technology is rapidly maturing
• Increased reliability, performance,
  cost-effectiveness of optical components
• Emergence of optical switch technology offers
  further augments to performance, flexibility and
  scalability to WDM networks
• WDM an attractive option for next-generation
  avionics network
• Huge bandwidth capacities, better weight
  characteristics, protocol transparency, etc.
• Challenge to provide packet-switched performance
  using connection-based components
• Optical switching technologies present very
  promising solution to meeting challenge in LAN
  environment
• One optical switching architecture is evaluated
  and analyzed via simulative experiments
• Two competing wavelength allocation approaches
  compared within switch architecture
• This approach represents just one of many
  potential WDM architectures for avionics networks

4
LION Library Overview

• MLDesigner selected as simulation modeling tool
• Discrete-event simulation environment, developed
  by MLDesign Technologies Inc.
• Advantages offered by MLD
• Models are fully extendible and user-definable
• Inherent hierarchical design facilitates modeling
  at multiple levels
• LION Library for Integrated Optical Networking
  (UF)
• Bridge gap between optic-centric and
  network-centric modeling and simulation analysis
  tools

• LION currently contains 39 optical component
  modules
• Components include couplers, splitters, lasers,
  receivers, etc.
• Parameters model key timing and physical
  component effects
• Low-level components used to realize any number
  of higher-level modules

Example Optical Receiver Model in LION
5
Optical Switching Architecture

• Optical switches based upon OSMOSIS architecture
  developed by IBM for HPC systems1
• Each connection includes both an optical link and
  electronic link
• Optical link is reserved for data transmission
  between ONICs
• Nodes make transmission requests to switch
  through control path
• An arbiter inside switch reserves optical paths
  as needed, and responds to transmitter when
  optical path is available
• Data transmissions are separated into timeslots,
  which are allocated by the arbiter(s)
• A broadcast-and-select approach is used for
  optical switching
• Each input is split and distributed to all
  outputs
• Each output chooses desired input among fibers,
  then wavelengths
• Use of smaller switch modules and slower
  switching technologies can make devices suitable
  for avionics

1 R. Hemenway, and R. Grzybowski, Optical
Packet-Switched Interconnect for Supercomputer
Applications, Journal of Optical Networking,
Vol. 3, No. 1, Dec. 2004.
6
OSMOSIS Case Study Network Layout

• Switches connected in a Clos2 topology
• Clos networks are highly connected, offering
  multiple paths between end points and limited
  fault-tolerance
• Additional nodes may be added by increasing
  perimeter switch count
• Additional bandwidth provided by increasing
  backbone switch count
• LAN consists of 8 perimeter switches, 3 backbone
  switches
• Each end node is connected to a perimeter
  switch
• Each perimeter switch can accommodate up to 28
  end nodes
• Backbone switches used solely to interconnect
  perimeter switches

Proposed CLOS LAN Topology
2 Clos, Charles, A Study of Non-Blocking
Switching Networks, Bell System Technical
Journal, March 1953, pp. 406-424.
7
OSMOSIS Case Study Experimental Configuration

• Experimental setup created to represent a
  centralized military avionics platform
• Platform includes of 97 nodes in eight different
  subsystems
• Data produced from each subsystem passes through
  central processing
• Limited communication between outer subsystems

• Two seconds of network traffic simulated in each
  experiment
• Aggregate traffic generated averages 300 MB/s
• Message sizes uniformly distributed between 1,000
  and 30,000 bits
• Sources generate either continuous, bursty, or
  random traffic

Experimental Configuration Diagram
8
OSMOSIS Case Study List of Experiments

• Two wavelength-allocation strategies compared in
  this study
• Fixed-destination Each node is assigned a fixed
  wavelength for receiving data, transmitters must
  tune to match
• Fixed-transmitter Each transmitter uses the same
  wavelength, receivers tune to match the senders
  wavelength
• Two additional parameters varied in simulative
  experiments
• Timeslot period Length of each timeslot for data
  transmission
• Timeslot period varied from 300 to 2000 ns
• Maximum slot allotment Maximum number of
  consecutive timeslots an arbiter may assign to
  one transmitter at a time, before resuming
  round-robin service
• Values of 7, 10, and 15 used for maximum slot
  allotment
• Simulation constants
• All optical transmitters and receivers operate at
  2.5 Gbps
• 100ns reserved at the end of each timeslot to
  perform optical switching
• Results accumulated for 2 seconds of network
  traffic
• 1 us tuning delay in tunable lasers and receivers

9
Simulation Results
Fixed-Destination Mean Latency (µs)
Fixed-Transmitter Mean Latency (µs)
Network System Model
10
Analysis of Results

• Fixed-destination protocol consistently offers
  better performance that fixed-transmitter
  allocation
• Average difference in overall latency is
  approximately 1 µs, same as the optical
  transmitter/receiver tuning delay
• This difference can attributed to ability to
  overlap scheduling and tuning in
  fixed-destination scheme
• Optical transmitters can tune to receivers
  wavelength while negotiating with arbiter for
  data transmission
• These results only considered unicast traffic
• Fixed-transmitter allocation may be optimal when
  there is significant multicast traffic, since
  multiple receivers could tune to same transmitter
  at once

Fixed-transmitter Allocation
Fixed-destination Allocation
11
Analysis of Results

• Variation of timeslot period caused most
  significant change in performance
• Packet latencies dropped as timeslot period
  increased
• With longer timeslots, a smaller fraction of each
  timeslot is needed for optical switching, which
  is constant at 100 ns
• When timeslot period is increased to 2000 ns,
  performance begins to decrease
• Timeslots are underutilized when handling small
  messages, which only occupy part of a single
  timeslot
• Ideal timeslot periods dependent upon
  distribution of message sizes
• A minimum of two timeslots is required for
  scheduling a transmission, longer timeslots
  increase the minimum scheduling delay seen by all
  nodes
• These scheduling delays are increased further for
  packets traversing multiple switches, which must
  go through scheduling once for each switch
• Small average gains seen by increasing the
  maximum timeslot allotment value
• Larger consecutive timeslot allotments increase
  the chance a message can be transmitted without
  interruption
• Increasing this value also increases the maximum
  queueing delay from the round-robin scheduler
• Smaller values guarantee faster delivery for
  small messages

12
Conclusions

• Optical switch architectures are a rapidly
  maturing technology
• Capable of realizing high-speed, flexible, and
  scalable networks
• OSMOSIS switch architecture is one such example
  being targeted for supercomputing environments
• Virtual prototyping is an ideal way to evaluate
  whether a modified architecture can meet the
  needs of avionics platforms
• Fixed-destination wavelength allocation showed
  better performance in unicast network
  configuration
• Results likely opposite in presence of
  broadcast/multicast traffic
• Tuning of timeslot period created largest
  variations in latency
• 1 µs timeslots provided best performance in our
  case study
• Optimal timeslot period will depend upon nature
  of network traffic for each platform
• Modest gains in performance seen by increasing
  maximum timeslot allotment parameter
• Reduction in average latency comes at the price
  of increasing the maximum scheduling/queueing
  delays

13
Future Work and Acknowledgements

• Future Work
• Increase the number of variable parameters
• e.g. varying switch port counts, available
  wavelengths, non-fixed wavelength allocations,
  etc.
• Investigate mechanisms for providing
  connection-oriented services
• Pre-allocation of resources for periodic
  transmissions
• Consideration of multicast traffic
• Implementation of advanced routing techniques and
  fault-tolerance
• Study and evaluation of additional optical
  switching architectures
• Acknowledgement
• This work was made possible by Navy STTR WDM
  Fiber-Optic Network Architecture Analysis,
  Modeling, Optimization, and Demonstration for
  Aerospace Platforms (c/o NAVAIR)
• STTR partner OptoNet Inc.