Switching

1
Switching

• Information Switching Technology in Networks

2
What is it?

• Switching --- a process by which a network
  element forwards information from its inputs to
  its outputs according to predetermined rules.
• Switch --- an element used to fulfill switching
  task.

3
Types of switches

• Telephone switches
• mostly switching voice information in voice
  samples format
• in telephone networks
• also called exchangers(e.g., PBX).
• Datagram switches
• (Datagram a data packet with its address
  information inside.)
• switching datagram in Internet, called routers
• switching Ethernet datagram(MAC based packet) ,
  called Ethernet switch
• ATM/FR/MPLS switches
• (FR frame relay MPLS MultiProtocol Label
  Switching)
• switching ATM/FR/MPLS datagram
• in ATM, cells, fixed-size packets
• FR/MPLS, variable-size packets

4
Classification

• Switches can be categorized by following ways
• Packet switches vs. circuit switches
• A packet switch --- switching packets
• A circuit switch --- switched data are the pure
  data to be transmitted.
• connectionless switches vs. connection-oriented
  switches
• A connectionless switch --- reads the destination
  address from the incoming packet and decides its
  output port(interface) by looking up an existing
  switching/routing table in the switch.
• A connectionless switch is sure a packet switch,
  but the reverse is not true.
• A connection-oriented switch --- many compose a
  connection-oriented switching system, in which
  data are transferred along a pre-established
  path, the path usually has an identifier.

5
Classification

• Connection-oriented switches(contd)
• The path is denoted by identifiers, which are
  assigned ahead of the data transmission
• in ATM/FR, the path identifier is VCI or VPI/VCI
  pair
• In MPLS, the identifier is a MPLS label
• In exchangers, the identifier is time
  slot TS
• Path Ids are swapped, so changed along the path
• When transmission ends, switch recycles path Ids
• Can be separated to 2 parts switch fabric and
  switch controller
• The switch fabric switches the data to the output
  interface by the path ID(So also a switching
  table/translation table is associated)
• The switch controller assigns a path ID to a data
  stream according to its destination address
  (signaling process)

6
Classification(contd)

• Conclude with following table
• ???

7
Other functions of switches

• Participate in routing algorithms
• to build routing tables, Chp11
• Resolve contention for output links
• Scheduling, Chp9
• Admission control
• to guarantee resources to certain streams,
  Chp13,14
• This Chapter, focus on pure data movement and
  architecture

8
Requirements

• Capacity of switch is the maximum rate at which
  it can move information, assuming all data paths
  are simultaneously active
• Primary goal maximize capacity
• subject to cost and reliability constraints
• A circuit switch will reject a call if it reaches
  its capacity, called call blocking. (In this
  case, the caller usually is told by a busy tone.)
  
• goal minimize call blocking
• If a packet switch reaches its capacity, it will
  try to store data in buffer first. If the buffer
  then becomes overflowed, switch had to discard
  the data, called packet loss.
• goal minimize packet loss

9
Requirements(contd)

• Reordering packets is harmful
• Costly to order back packets
• Increase end-end delay greatly
• Connection-oriented switches generally do not
  reorder packets,but connectionless switches (such
  as routers ) often do.
• Goal minimize packet reordering.
• ???

10
A generic switch

• A generic switch has 4 parts
• input buffers
• a port mapper
• a switch fabric
• output buffers.
• But realistic switches often distribute, combine,
  or omit one or more of the upper parts

11
A generic switch(contd)

• Input buffers temporarily store data arriving on
  the input ports.
• The buffer space differs from different kind of
  switches.
• very small, buffers are at output side
• very large, almost all of the switch buffers at
  input side.
• a port mapper reads destination address or path
  identifier of the packet and uses it as an index
  to look up the switching/translation/routing
  table to find out the output port
• only for packet switches.
• A circuit switch does not need a port mapper,
  every time slot (TS) is statically associated to
  a output port.
• Does not do switching

12
A generic switch(contd)

• A switch fabric switches packets (or data) to the
  specific output port buffer according to the
  output port number assigned by the port mapper.
• In packet switches, switch fabric is usually a
  processor or a complex multiprocessor, at
  high-speed.
• In circuit switches, the switch fabric is usually
  an ASIC or multi ASICs
• Output buffers store data, waiting for a turn to
  put on output link.
• A scheduler is usually used to manage the output
  buffers and decide the access right to the output
  link.
• buffer space may differ from different kind of
  switches
• very small, with most buffers at input side.
• very large, with almost all of the switch
  buffers at the output side.
• Usually, if the input buffers are large, the
  output buffers are small, and vice versa.
• ???

13
Outline

• (Circuit switching)
• Packet switching
• Switch generations
• Switch fabrics
• Buffer placement
• Multicast switches

14
Packet switching

• Discuss two types of packet switches
• virtual circuit ATM switches
• datagram switches --- routers
• Other than that, very similar
• ATM switches
• Handle fixed-size packets
• port mapper decides the output port by VCI or VPI
  or VPI/VCI in every incoming packet header
• switching table is a VCI-to-Port table ---
  translation table

15
Packet switching

• Routers --- a kind of switches
• Handle variable-size packets.
• port mapper decides the output port by the entire
  destination address included in every incoming
  packet header
• The switching table is a entire address to
  Port table --- routing table.
• ???

16
Repeaters, bridges, routers, and gateways

• According to the different working position in
  the protocol stack, datagram switches have
  different names
• Repeaters
• operating at physical level(L1)
• repeating input signal on one or more of its
  outputs, only with the signal reinforced at
  physical level (such as regulating the waveform,
  restoring timing, etc).
• A Hub is a kind of repeater.
• Bridges
• at datalink level (L2,)
• connecting different segments of a LAN together
  by forwarding packets directly to the segments
  according to the packets datalink layer
  destination address(in Ethernet, the MAC
  address), lowing the amount of traffic on the
  LAN.
• In Ethernet, is an Ethernet LAN switch(based on
  MAC addresses).
• discover attached segments by listening
• Difference between Ethernet switches and Hubs?

17
Repeaters, bridges, routers, and gateways(contd)

• Routers
• at network level (L3)
• forwarding packets to one of the output ports
  according to the packets network layer address(
  in Internet, the IP address).
• participate in routing protocols
• Differences between bridges and routers
• bridges using datalink layer address, connecting
  different segments of one same network(such as
  LAN) together.
• routers using network layer address, connecting
  different packet based networks together, forming
  an Internet.
• Application level gateways
• a datagram switch working at application
  layer(L7)
• interconnecting between different networks and
  even different systems, treating an entire
  network as a single hop
• E.g
• an IP phone gateway connects PSTN networks into
  Internet, so that a telephone call can be made
  via Internet.
• mail gateways and transcoders

18
Repeaters, bridges, routers, and gateways(contd)

• Summary 
• repeater -gt bridge -gt router -gt gateway
• layer in protocol stack low level -gt upper
  level
• functions simple -gt complex
• forwarding speed fast -gt slow
• Gain functionality at the expense of forwarding
  speed
• for best performance, push functionality as low
  as possible
• ???

19
On port mappers in routers

• Look up output port based on destination address
• Easy for ATM just use a table, with one
  field(VCI or VPI or VPI/VCI) to match
• Harder for IP routers
• need to match more fields
• e.g. packet with address 128.32.1.20
• routing tables may have entries (128.32.),
  (128.32.1.), (128.32.1.20)
• routing table is usually quite large, sometimes
  with tens of thousands of entries, data structure
  and match algorithm become very important for the
  router to reach high performance, should be
  quick, efficient, and reliable.
• An often used data structure for routing table
  trie

20
Trie

• Is a commonly used data structure for routing
  tables
• basically a tree data structure, with the address
  elements as its tree nodes.
• A node is also associated with an output
  interface number, meaning the address the match
  ends at this node should be mapped to that output
  interface.
• A match process starts at the root node, and
  goes as far down as possible.
• The output interface of the farthest node the
  match can go is just the port mappers result.

21
Trie

• An example
• IP address based trie structured routing table.
  Each node represents a partial string of the
  whole IP address.
• If no match can be made for a given IP address,
  the match ends at the root node. This is the
  default routing state for the router, so the
  associated output interface for root node is for
  all default routes.
• All other nodes have specific IP addresses or
  address segments to match with, and all are
  associated with output interface numbers.
• Match for 128.32.5.10 ?

Out1
Out3
Out2
Out4
Out5
Out8
Out5
Outport6
Out7
22
Trie

• What other data structures for routing tables? 
• - linear tables, for example  128.32.1.120
• 128.32.1.100
• 128.32.1.
• 128.32.25.
• 128.32.
• Benefits from trie net
• quick in address match, the larger the routing
  table the quicker. Much quicker than linear table
  structures.
• But for small tables, this is not true
• small in table storing space.
• Some improvement tricks for port mapper
• using a cache to store recently matched IP
  addresses. The cache is first checked before a
  new match begins.
• using a cache to store some often used IP address
  match results. The cache is first checked before
  a new match begins.
• ???

23
Blocking in packet switches

• Can have both internal and output blocking
• Internal
• no path to output, happen at input buffers, port
  mapper, and switch fabric.
• Output
• trunk unavailable, happen at output buffers and
  output ports.
• If packet is blocked, must either buffer or drop
  it
• May first try to buffer, then to drop
• cannot predict if packets will blocked or not,
  unlike a circuit switch
• circuit switches can only block a call during a
  call-setup time. During data transmission phase,
  no block anymore. To block the call is simply to
  reject to setup a path for the call. So blocking
  in circuit switches only affects connection setup
  ability for a network instead of transmission
  quality of a call.
• Packet switches can not predict quantity of
  packets to transfer
• Even in ATM switches, blocking may happen

24
Dealing with blocking

• Speed up
• Speeding up internal switching processing to
  prevent internal blocking, internal links much
  faster than inputs
• Buffers
• Placing more at input or output
• QoS
• Using admission control technology to simulate
  circuit switching.
• Backpressure
• if blocking happens at switch fabric or output
  port, switch sends a signal back to its input
  port telling the blocking and preventing more
  data entering the switch center. (flow control )
• Improve performance of switch fabric
• Sorting and randomization
• Parallel switch fabrics
• increases effective switching capacity
• ???

25
Outline

• Circuit switching
• Packet switching
• Switch generations
• Switch fabrics
• Buffer placement
• Multicast switches

26
Three generations of packet switches

• Along with ever-advancing technologies, switches
  have evolved for three generations, with the
  switching capacity becoming larger and larger.
• Different trade-offs between cost and performance

27
First generation switch

• Switch fabric is realized by software in CPU, the
  same is for port mapper part. Buffers are mainly
  computer memory.
• Most Ethernet switches and cheap packet routers
  are
• Also PC based routers
• Bottleneck can be CPU, host-adaptor or I/O bus,
  depending
• ???

28
Example

• First generation router built with 133 MHz
  Pentium(Instruction cycle1/1337.52
  nanoseconds), assume
• Mean packet size 500 bytes
• Interrupt takes 10 microseconds,
• word access take 50 ns(time to r/w a word from
  memory,I/O bus speed)
• Per-packet processing(port mapper, switch fabric)
  time takes 200 instructions 1.504 µs
• Copy loop(one word)
• register lt- memoryread_ptr
• memory write_ptr lt- register
• read_ptr lt- read_ptr 4
• write_ptr lt- write_ptr 4
• counter lt- counter -1
• if (counter not 0) branch to top of loop
• Copy one word takes 4 instructions 2 memory
  accesses 130.08 ns, copying packet takes 500/4
  130.08 16.26 µs interrupt 10 µs
• Total time
• 10us(int.)16.26(copying packet)1.504us(per-pac
  ket processing) 27.764 µs
• gt speed is 500bytes/27.764us144.1 Mbps
• overhead such as doing route-table computation
  not taken into account
• In conclusion, not efficient in resources, can be
  improved
• ???

29
Second generation switch

• Line cards are more intelligent, with a processor
  inside
• able to decide output port of some packets
  themselves, and put packets on bus and directly
  to output linecards, saving lots of time
• Some port mapping done by line cards themselves,
  instead of center CPU
• Need line cards keeping a local
  routing/translation cache, while center computer
  keep a larger routing table
• For routers, If a packet can not be routed by
  line card, its sent to center processor, which
  will forward it and update routing cache of this
  line card then. This may result in packet
  reordering.
• For ATM switches, because translation cache in
  line card is setup by path signaling process,
  port mapping can surely be fulfilled in line
  cards, no packets sent to center, no packet
  reordering. But during path setup, line card need
  control from center.

30
Second generation switch(contd)

• A shared bus or a ring to connect line cards and
  center processor
• a bottleneck for system, bandwidthsum of b of
  all ports
• Improvement
• In connection-oriented switches, using a shared
  port mapper card to replace all distributed port
  mappers in port cards.
• Packets go input line card-gtport mapper card
  -gtoutput line card
• Ipsilon IP switching
• goal to route IP packet over ATM, assume
  underlying ATM network
• by default, assemble cell packets back to IP
  packet and use an adjunct processor to route it.
• if detect a flow, ask upstream to send on a
  particular VCI, and install entry in port mapper
  gt implicit signaling, no need to assemble
  anymore.
• ???

31
Third generation switches

• Bottleneck in second generation switch is the bus
  (or ring)
• Third generation switch, shared bus is replaced
  by a switch fabric
• Switch fabric is composed of interconnected buses
  and switching elements.

32
Third generation (contd.)

• Port mappers are in ILCs
• Switch fabric features
• self-routing fabric
• output buffer is a point of contention
• unless we arbitrate(manage) access to fabric
• ???

33
Outline

• Circuit switching
• Packet switching
• Switch generations
• Switch fabrics
• Buffer placement
• Multicast switches

34
Switch fabrics

• Transfer data from input to output according to
  the output port given by port mapper, ignoring
  scheduling and buffering,
• Usually consist of links and switching elements
• 1st and 2nd gens may use software. 3rd gen
  hardware
• Many designs,
• Crossbar
• Broadcast
• Switching element
• Banyan
• Sorting and merging fabrics
• Batcher-Banyan
• Be aware that hard part in designing a switch is
  not the switch fabric, but deciding where to
  place buffers and how to schedule access to the
  buffers and bandwidth.

35
Crossbar

• Simplest switch fabric
• Quite same as that in telephone circuit switch
• Used here for packet routing cross-point is left
  open long enough to transfer a packet from an
  input to an output
• A switching scheduler is used to tell where,
  when, and how long
• (differs from schedulers of output port)
• For packets with fixed-size or a known arrival
  pattern(ATM), can compute schedule in advance
• Otherwise, need to compute a schedule
  on-the-fly(by reading the length message of
  packet).

36
Buffered crossbar

• What happens if packets at two inputs both want
  to go to same output? ---- output blocking. To
  avoid,
• Can have the crossbar N times faster than the
  inputs(expensive)
• Or, buffer crosspoints(practical)
• ???

37
Broadcast

• Packets are tagged with output port by port
  mappers
• Then are broadcast to all output ports
• Each output matches tags, so need to match N
  address items in parallel at each output(should
  be N times faster than input)
• Need not a switching schedule
• Useful only for small switches, or as a stage in
  a large switch
• ???

38
Switch fabrics built with fabric element

• Can build complicated fabrics from a simple basic
  switch fabric element
• A switch fabric element
• Consisting of 2 inputs, 2 outputs, and an
  optional buffer
• Routing rule check a bit of the output port
  number. if 0, send packet to upper output, else
  to lower output
• If both packets go to the same output, buffer or
  drop

39
Features of fabrics built with switching elements

• NxN switch can be built with many bxb switching
  elements. Itl have stages with
  switching elements per stage
• E.g 4096x4096 switch can be built with 4 stages
  of 8x8 switching elements, with 512 elements in
  each stage. 8x8 switching elements can further
  be built with 2x2 elements, with 3 stages, per
  stage with 4 elements.
• Fabric is self routing, given an output port
  number.
• Recursively composed from smaller to larger
• Can be synchronous or asynchronous
• Async. need a buffer, fit for variable-size
  packet switching
• Regular and suitable for VLSI implementation
• ???

40
Banyan(1973)

• Simplest self-routing recursive fabric
• Input packets are tagged with output port in
  binary, i th element looks at the i th bit of
  the tag, forwarding rule 0---upper layer, 1---
  lower layer
• 2n outputs need n stages with 2n-1 elements in
  each stage
• ???
• What if two packets both want to go to the same
  output of a switching element?
• Blocking happens at output of the switching
  element

110-gt
41
Blocking

• Can avoid with a buffered banyan switch, every
  element being associated with a buffer
• but this is too expensive
• hard to achieve zero loss even with
  buffers(impossibly large enough)
• Instead, can check if path is available before
  sending packet
• three-phase scheme
• send requests
• inform winners
• send packets
• Or, use several banyan fabrics in parallel
• intentionally misroute and tag one of a colliding
  pair
• divert tagged packets to a second banyan, and so
  on to k stages
• expensive
• can reorder packets
• output buffers have to run k times faster than
  input
• Or, use sorting --- Batcher-Banyan
• ???

42
Batcher-Banyan

• A Batcher sorting network sorts packets according
  to their output port
• A trap network removes duplicate packets
• Where to remove?
• recirculate to beginning
• or run output of trap to multiple banyans
  (dilation)
• A shuffle-exchange network shuffles the packets

43
Batcher-Banyan(contd)

• Principle
• blocking in Banyan can be avoided by choosing
  order in which packets appear at input ports
• E.m two packets individually are to output port
  000 and 001
• if put at input of 000 and 001, packets will be
  blocked at first stage
• if put at input of 000 and 010, no blocking
• Procedure
• packets at inputs sorted by output number, by
  sorting network
• remove duplicates and remove gaps
• Shuffle with a perfect-shuffle network
• Enter into normal Banyan network
• It has be shown that Banyan network with above
  procedure is internally nonblocking.
• For example
• Ordered set of out port X, 011, 010, X, 011,
  X, X, X
• -(sort)-gt 010, 011, 011, X, X, X, X, X
• -(remove dups)-gt 010, 011, X, X, X, X, X, X
• -(shuffle)-gt 010, X, 011, X, X, X,
  X, X
• Bater-Banyan cheaper than per-element buffering
  in avoiding blocking
• ???

44
Sorting

• (5,7,2,3,6,2,4,5)-sort-gt(2,2,3,4,5,5,6,7)
• Can build sorters from merge networks
• Merging
• merge two sorted list into one sorted list
• E.g (5,7) (2,3)-merge-gt(2,3,5,7)
• (2,6) (4,5)-merge-gt(2,4,5,6)
• E.g a 4m4-gt8 merging network using comparator
• Sorting by merging
• E.g
• 1m1-gt2 merging,
• (5)m(7)-gt(5,7), (2)m(3)-gt(2,3), (6)m(2)-gt(2,6),
  (4)m(5)-gt(4,5)
• 2m2-gt4
• (5,7)m(2,3)-gt(2,3,5,7),
• (2,6)m(4,5)-gt(2,4,5,6)
• 4m4-gt8
• (2,3,5,7)m(2,4,5,6)-gt(2,2,3,4,5,5,6,7)
• Can sort any size list by merging two sorted
  lists recursively
• ???

45
Shuffle

• Use a shuffle algorithm
• E.g
• 010, 011, X, X, X, X, X, X -(shuffle)-gt 010,
  X, 011, X, X, X, X, X
• Need to read more reference to understand the
  algorithm
• http//www.nist.gov/dads/HTML/perfectShuffle.html
  
• ???

46
Effect of packet size on switching fabrics

• A major motivation for using small, fixed packet
  size in ATM is ease of building large parallel
  fabrics.
• But it seems small packet size may not help in
  building switches
• In general, smaller size gt more per-packet
  overhead, but less packetization delay
• At high speeds, overhead(such as interrupt time
  in 1st gen switch) dominates!
• Fixed size packets helps build synchronous switch
• But we could fragment at entry and reassemble at
  exit
• Or build an asynchronous fabric
• Thus, variable size doesnt hurt too much
• Conclusion maybe Internet routers with
  comparatively larger,variable-size packets can be
  almost as cost-effective as ATM switches
• ???

47
Outline

• Circuit switching
• Packet switching
• Switch generations
• Switch fabrics
• Buffer placement
• Multicast switches

48
Buffering

• All packet switches need buffers to match input
  rate to service rate
• or cause heavy packet loses
• Where should we place buffers?
• input
• in the fabric
• output
• shared

49
Input buffering (input queuing)

• An arbiter is used to schedule release of packets
  from input queues when an access to both switch
  fabric and output trunk is available
• Advantages
• No speedup in buffers or trunks (unlike output
  queued switch), all elements(queues, switch
  fabrics, etc) inside only need to run as fast as
  input lines
• the arbiter needs to run at sum of speed of input
  lines
• therefore, can possibly build fast speed switches

50
Input Queuing(contd)

• Problem head of line blocking(HoL)
• If a packet in the queue blocked, all followed
  packets are blocked even if the paths for them
  are available.
• with randomly distributed packets, utilization of
  queue resources is only at most 58.6, even worse
  with hot spots (some ports are more favored)
• Dealing with HoL
• Per-output queues at inputs, so N outputs -gt N
  queues in every input queue group. If one output
  access is blocked, others still can go.
• Arbiter must choose one of the input ports for
  each output port
• But how to select?
• Parallel Iterated Matching
• inputs tell arbiter which outputs they are
  interested in
• output selects one of the inputs
• some inputs may get more than one grant, others
  may get none
• if gt1 grant, input picks one at random, and tells
  output
• losing inputs and outputs try again
• Used in DEC Autonet 2 switch
• Arbitration schemes for input-queued switches are
  still on the research frontier
• ???

51
Output queueing

• Dont suffer from head-of-line blocking
• Can also allow control of output queues in fine
  grain
• But output buffers need to run much faster than
  trunk speed
• Need to store packets from all inputs, so sum of
  all input speed for buffer
• Make it more expensive than to build an input
  queuing switch.
• At the same cost, can build a faster input
  queuing switch
• Can reduce some of the cost by using the knockout
  principle
• Unlikely (though possible) that all N inputs will
  have packets for the same output
• So set a number SltN
• Need drop extra packets, fairly distributing
  losses among inputs
• ???

52
Shared memory

• Route only the header to output port
• Switch fabric only switches packet headers,
  easier to build
• Bottleneck is time taken to read and write
  multiported memory
• Need N times as fast as inputs (sum of input
  speed) to access memory
• Hard to build a large scale switch with many
  input
• For an 64 input ports and155Mbps each ATM switch,
  need to access 64 ATM cells at 53bytes/155Mbps2.7
  3us gt access speed1word/3.05ns, too high,
  common EDO memory 50ns/word
• But can form an element in a multistage switch

53
Datapath switch(IDTI) clever shared memory design

• An 8x8 integrated shared-memory ATM switch module
  in one chip
• Includes a controller, 4k cell memory
• Reduces read/write cost by doing wide reads and
  writes(using parallel shift registers)
• 1.2 Gbps switch for 50 parts cost
• ???

54
Buffered fabric

• Have discussed buffered crossbar fabric, can also
  be used in all switch fabrics such as Banyan
• Buffers in each switch element
• From first stage to last stage
• advantage
• Speed up is only as much as fan-in
• Hardware backpressure can be used to reduces
  buffer requirements
• disadvantage
• costly (unless using single-chip switches), many
  distributed memory
• scheduling is hard, so hard for QoS
• As a result, purely buffered-fabric switch is
  impractical, more are actually composed of
  buffered single-chip switch elements such as
  Datapath chips

55
Hybrid solutions

• Buffers at more than one point
• Such that
• a few input buffers to deal with output blocking,
  lower output buffer speed,
• More output buffers to be well scheduled,
  providing QoS
• Using shared bus with each interface card
  providing both input and output buffering
• Becomes hard to analyze and manage
• But common in practice
• Buffer arrangement is a tough point in switch
  design
• Transmission cost vs. buffer cost
• Buffer is more economical, using as large as
  possible memory to reduce transmission link speed
• ???

56
Outline

• Circuit switching
• Packet switching
• Switch generations
• Switch fabrics
• Buffer placement
• Multicast switches

57
Multicasting

• Multicast send packets to several chosen
  destinations
• Unicast send packets to one destination
• Broadcast send packets to all available
  destinations inside a network such as a LAN
• Transmission point-to-point, point-to-multipoint,
  multipoint-to-multipoint(conference)
• better to do this in hardware in switch fabrics,
  how?
• Multicast packet arrives, port mapper retrieves
  the list of outputs(instead of one output port
  number)
• Incoming packet copied to these output ports
• Two sub-problems
• generating and distributing copies, how?
• VCI translation for the copies

58
Generating and distributing copies

• Either implicit or explicit
• Implicit
• suitable for bus-based, ring-based, crossbar, or
  broadcast switches
• multiple outputs enabled according to the port
  list given by port mapper after placing packet on
  shared bus
• Only put packet on bus one time
• used in Paris and Datapath switches
• Explicit
• Suitable for such as Banyan switch
• need to copy a packet at switch elements
• use a copy network before entering Banyan network
  to do this
• One input many output
• Propagate packets according to the port list
  given by port mapper
• Many packets with different port tags at output
  of copy network
• Output of copy network then as input of Banyan
• Both schemes increase blocking probability

59
Header translation

• Recall VCI is swapped in ATM switch
• Normally, in-VCI to out-VCI translation can be
  done either at input or output
• Just by looking up translation table and finding
  out out-VCI and replacing it with in-VCI
• Can be done at time of port mapping(input side)
• With multicasting, every copied packet need
  individual VCI swapping
• so translation easier at output port side
• Need two separate tables port mapping table and
  a normal translation table
• Port mapping table is at input side, for input
  packet to use in-VCI as index to find out a set
  of output ports(multicast).
• translation table at output side, for every
  copied packet to use in-VCI and output port to
  find out-VCI and swap it
• Need to do two lookups per packet
• ???

60
Summary

• 1st generation switches a computer with many
  interface cards
• Can achieve speed up to 300Mbps(1997)
• 2nd generation switches line cards are more
  intelligent and with a high speed bus connected
  together directly
• Can achieve speed up to 8 or more times of 1st
  gen. switches
• 3rd generation switches using many parallel high
  speed buses
• Can achieve speed up to Gbps

61
Assignments

• Exercises 8.2
• Exercises 8.9
• Exercises 8.11
• Exercises 8.5