Design of Routers for Optical Burst Switched Networks

1
Design of Routers for Optical Burst Switched
Networks

• Dissertation Proposal Defense
• Jai Ramamirtham (jai_at_arl.wustl.edu)
• Dissertation Advisor Dr. Jon Turner

2
Optical Burst Switching Concept
WDM links
Concentrator
One or more wavelengths used to transmit control
information (Burst Header Cells).
Optical switches switch data from incoming
wavelength channels to outgoing wavelength
channels (possibly switching the data from one
wavelength to another)
Concentrators transmit user data in the OBS
format. May aggregate user packets to improve
efficiency.

• Uses optics for data transmission
• Uses electronics for control
• Achieves good statistical multiplexing
  performance with little or no buffering

3
Optical Burst Switching Concept
WDM links
Concentrator
4
Optical switch design
Switching
WDM fibers
Contention resolution

• Multiplexing/Demultiplexing
• Passive couplers
• Switching
• High speed optical crossbars based on
  Semiconductor optical amplifiers (SOA) or Lithium
  Niobate switches

5
Contention resolution

• Buffering using Fiber Delay Lines
• Half a second of buffering gt 94,000 miles of
  fiber
• Recirculation through a smaller fiber gt signal
  degradation
• Not random access
• Deflection routing
• Forward packet that loses contention to a random
  next hop
• Effectiveness depends heavily on network topology
• Out of order delivery of packets
• Wavelength conversion
• Doesnt rely on buffers to provide good
  performance
• Technology still remains expensive

Most optical switching solutions are not cost
effective
6
Dissertation outline

• Objective To identify and study switch designs
  that reduce the cost of optical switching systems
  and give good performance
• Wavelength converting switches using Arrayed
  Waveguide Grating Routers (WGRs)
• Time Sliced Optical Burst Switching (TSOBS)
• Aggregation and Load Balancing for TSOBS networks

7
Wavelength Converting Switches
h wavelengths
d input/output fibers

• d input/output fibers (typical values d4,8,16)
• h wavelength channels (typical values
  h64,128,256)
• d8, h256, each wavelength channel carries 10
  Gb/s, agg. system throughput 20.48 Tb/s

8
Wavelength Grating Router(WGR)
TWC

• Passive device, much less complex than crossbars
• Combination of tunable wavelength converters and
  wavelength router provides required switching

9
WGR-based switch
hxh WGR
h/d
h
d
d
hxh WGR
h
h/d

• Optical crossbars are substituted with WGRs
• Complexity of system reduced significantly
• Introduces blocking in the system

10
Modeling as a Game board
Color of a square corresponds to the output fiber
reached using the wavelength (column) at the
input (row)
Each row pattern is a cyclic shift of the
previous row within a block
Each block can have an independently different
pattern

• The configuration can be modeled as a game board
• d hxh blocks
• Rows within a block are cyclic shifts of the
  previous one
• Different blocks can have different patterns

11
Puzzle formulation
Only one token of any color in a column
Tokens placed on squares of the same color

• Setting up connections within switch can be
  modeled as a puzzle
• Place colored (up to d.h) tokens beside the board
• Move the tokens on the board on squares such that
• Token has the same color as the square it is
  placed on
• No two tokens of the same color are in the same
  column
• If puzzle solvable for all possible setups,
  switch is rearrangeably non-blocking

12
Two-Player Game formulation

• Game models a burst switch where bursts arrive
  and leave randomly
• Two players blocker and setter
• Blocker has k tokens of each color

• Blocker removes zero or more tokens and places
  one or more tokens beside the board
• Setter tries to place the tokens on the board
• If setter can keep the game going indefinitely,
  then non-blocking

13
Solving puzzle setups

• Breaks apart into d sub-graphs for each output
• Accommodate new connections by rearranging
  existing ones using augmenting path algorithm

14
Good game boards

• Definition
• A game board is k-solvable if every puzzle setup
  with at most k tokens of any color can be solved
• Definition
• Cover of row i for output j set of columns that
  have a j-colored square in row i
• Cover of a set of rows R Union of the covers of
  the rows contained in R
• Game board is k-solvable iff the cover of all
  sets of r ? k rows contains at least r columns
• A good game board is one that maximizes the
  covers of all possible row sets R of size ? h

15
Finding good game boards

• Spread the squares unevenly across the rows
• Randomizing the positions of the squares can help
  maximize the covers of rows

16
A bad game board
17
Game board results

• Is there a game board configuration that is
  rearrangeably non-blocking? Answer No
• Number of tokens that can be guaranteed to be
  placed ? h-d1
• For d2, a contiguous game board is optimum

• For a random game board, with a randomly selected
  set of rows, r

18
Random game boards
Also upper bound on probability that a given set
of r rows misses one or more columns.

• Shows average number of columns that are not
  covered by a set of rows
• For d8, probability that a set of ? 140 rows do
  not cover a column is one in a million
• easy to solve a random puzzle setup

19
Burst switch simulation

• Switch Size 8 input/output fibers
• No buffering
• Burst inter-arrival times are exponentially
  distributed
• Bursts assigned to a random output
• Arriving bursts assigned to first available
  wavelength to reach the output
• Corresponds to placing the token in the leftmost
  available square
• Burst rejection probability
• Fraction of arriving bursts that must be discarded

20
Simulation results
d 8

• Results for random game boards

21
Game board configurations
b) Interleaved
a) Contiguous

• c) Random configuration
• d) Hand-tuned configuration

22
Different configurations
Config 1 contiguous Config 2
interleaved Config 3 random Config 4
hand-tuned

• Results for various game boards for
• d8, h256

23
Results with rearrangement

• Use the bipartite augmenting path algorithm to
  rearrange connections
• Results matched non-blocking results almost
  exactly
• Reaffirms the fact that for a set of inputs, the
  available output channels is sufficient in most
  cases
• Better token placement strategies for the setter
  can result in better performance

24
Token placement strategies

• Place tokens on the game board such that the
  probability of new bursts getting dropped is
  minimized
• Evaluated strategies
• First available column
• Random available column
• Least affecting column
• Most available column
• 89 of the utilization of a strictly non-blocking
  switch

25
Extension with buffering
TWC
WavelengthRouter
h
TWC
d
input fibers
d
TWC
WavelengthRouter
h
TWC
Few channels at the inputs need to be dedicated
to buffer ports
RCV
TL
WavelengthRouter
h
RCV
TL
Contention among inputs to reach buffer ports
b ports for buffering
b
RCV
TL
WavelengthRouter
h
RCV
TL
tunable lasers
fixed receivers
electronic buffers

• Determine the optimum number of buffering ports
  to get maximum performance

26
Extension using multiple WGRs

• Game board has k.h columns
• Additional constraint that only one token of a
  color can be placed in columns i, hi,, (k-1)hi
• Trivial construction with k d that is strictly
  non-blocking

27
Dissertation outline

• Objective To identify and study switch designs
  that reduce the cost of optical switching systems
  and give good performance
• Wavelength converting switches using Arrayed
  Waveguide Grating Routers (WGRs)
• Time Sliced Optical Burst Switching (TSOBS)
• Aggregation and Load Balancing for TSOBS networks

28
Network architecture
WDM links
Concentrator
Packet from a host
Space-division optical switches switch data from
incoming timeslots to timeslots in the outgoing
link (possibly delaying the data)
Lower bit-rate host interface (e.g. Gig-Ethernet)
Frame of time slots

• Time Sliced Optical Burst Switching (TSOBS) is
  designed to eliminate need for wavelength
  conversion by switching in the time domain
  instead
• Can be done with very little buffering capacity

29
Design Issues Timeslot duration

• Timeslot duration
• each timeslot has a guard time to allow for
  timing uncertainties
• solid-state switches perform switching in 10 ns
  or less
• accuracy of synchronization of timeslots also is
  a determining factor for the guard time
• guard times of 10-100 ns implies that we need to
  have a timeslot of the order of 1 ?s for data
  transmission efficiency
• at transmission rates of 10 Gb/s, 1 ?s timeslot
  corresponds to approximately 1100 bytes of user
  data
• with 90 timeslots per frame, each timeslot
  corresponds to a 100 Mb/s channel

30
Design Issues Timeslots per frame

• For single timeslot bursts, good performance with
  moderate number of timeslots per frame

31
Effect of burst length on performance

• Performance reduces if bursts are longer than a
  single timeslot
• We expect most packets to be contained within a
  single timeslot

32
Switch architecture
The controller used the information in the BHCs
to make switching decisions and generates the
corresponding control signals
Optical Time Slot Interchangers provide the
required time domain switching
Crossbars perform the required space switching
SYNC blocks synchronize incoming frame boundaries
to local timing reference using variable delay
lines, with feedback control from controller
33
Design Issues Signal degradation

• Optical signals degrade when traveling through
  multiple hops requiring regeneration midway
• Equip each switch with few ports of regenerators
• BHC of burst carries information on the number of
  hops, distance traveled
• Information in BHC used to regenerate bursts as
  necessary
• If bursts travel through ten or more routers
  before regeneration, TSOBS has a decisive cost
  advantage
• Minimizing number of switching operations within
  a switch becomes very important

34
Optical Time Slot Interchanger
Signals de-multiplexed before switching and
re-multiplexed onto delay lines. Cost of delay
lines shared by the different wavelengths.
One crossbar per wavelength to switch the signals
onto delay lines

• Cost of crossbars is critical
• Need to minimize the number of delay lines

35
Design issues for OTSIs

• Number of delay lines
• Cost of OTSI grows quadratically with the number
  of delay lines
• Average number of times a burst gets switched
  within the OTSI
• Complexity of scheduling operation
• With 1 ?s timeslot duration and average burst
  lengths of 10 timeslots, 10 ?s to process one
  burst header cell from each input wavelength
  channel
• Each wavelength channel can be handled in parallel

36
Non-blocking OTSIs

• Straightforward design with N delay lines of one
  timeslot each
• Least possible delay line size N
• Large crossbar size (N1)(N1)
• Up to N switching operations
• Reduce switching operations by using delay lines
  of length 1, 2,,N instead.
• Delay lines of length 1,2,, ?N1/2?-1 and ?N1/2?,
  2?N1/2?,,(?N1/2?-1)?N1/2?
• Crossbar size (2?N1/2?-1)(2?N1/2?-1) 3131,
  N256
• Length of fiber N?N1/2?/2 (2048, N 256)
• Maximum number of switching operations 3

37
Blocking OTSIs

• Lower complexity alternative with a small
  non-zero blocking probability
• Natural choice of delays 1,2,,N/2
• Crossbar size log2Nlog2N (88 for N256)
• Length of fiber N-1 (255 for N 256)
• Define a search procedure to find sequence of
  delay lines to switch signals onto without
  creating conflicts
• We show that the number of switching operations ?
  3, under most conditions
• Also, the impact of blocking on the statistical
  multiplexing performance is small

38
Blocking vs. Nonblocking OTSI
non-blocking

• We do not lose much by using blocking OTSIs by
  way of performance

39
Avg. number of switching operations

• For loads up to 70, average remains below 2 and
  for loads up to 90, the average remains below 3

40
Design issues for synchronizers

• SYNC blocks realign data at the input ports that
  arrive at varying phases
• Use the same basic structure as OTSIs
• using space division optical switches with a
  finely calibrated set of delay lines
• Two parameters that affect the cost of a SYNC
• Precision
• The difference between successive delay values
• Range
• Maximum value a timeslot can be delayed
• Ratio of Range to Precision determines number of
  delay values
• Need only to align data on timeslot boundaries
  rather than frame boundaries

41
Dissertation outline

• Objective To identify and study switch designs
  that reduce the cost of optical switching systems
  and give good performance
• Wavelength converting switches using Arrayed
  Waveguide Grating Routers (WGRs)
• Time Sliced Optical Burst Switching (TSOBS)
• Aggregation and Load Balancing for TSOBS networks

42
Aggregation in TSOBS networks

• Why aggregation?
• Reducing the amount of control information
• Single fiber link with 64 wavelength channels
  with 10 Gbps capacity each, 48 byte packets, link
  utilization of 60 gt need to forward 1 billion
  packets per second
• Aggregation increases transmission efficiency
• Minimum packet size constraints imposed by the
  optical switches
• In TSOBS networks, a timeslot is 1 ?s or 1100
  bytes long at 10 Gbps transmission speed

43
Aggregation method

• Assume that there are N interfaces
  (concentrators) that form an OBS network
• At each interface,
• Qi queue of packets destined to interface i
• After receiving a packet when the queue is empty,
  wait for a fixed duration called the burst
  aggregation period
• Collect the packets going to destination
  interface i during this time period
• Form the burst at the end of the time period and
  transmit it

44
Effects of Aggregation

• Aggregation introduces delay into the data path
• Performance penalty due to delay
• Sources with low data rates are affected
• Low bandwidth TCP sources have a lower sending
  rate than without aggregation
• Correlation benefit
• Sources with high data rates benefit the most
• High bandwidth TCP sources have a higher sending
  rate than without aggregation

45
Issues with Aggregation

• Aggregation efficiency depends on the short term
  traffic characteristics
• Consider a network with 1000 interfaces, 1000
  hosts each transmitting at 100 Mbps, assuming
  uniform spreading of traffic
• With a 1 ms aggregation period, the data received
  is 100 Kbits
• If there are 10,000 interfaces instead, the data
  received goes down to 10 Kbits
• Exploit long term traffic behavior to dimension
  the burst aggregation period

46
Adaptive burst aggregation

• Basic idea
• Start with some burst aggregation period
• Measure the burst aggregation efficiency