Академический Документы
Профессиональный Документы
Культура Документы
1
Outline
2
Introduction
Transponder/
Framer Network Traffic
Transceiver
Processor Manager
Switch
Fabric
Transponder/
Framer Network Traffic
Transceiver
Processor Manager
Switch
Fabric
5
Introduction
For IP Routers
Use packet switching. Time for each switching is much shorter
OC192 OC768 OC3072
Line rate 10Gbps 40Gbps 160Gbps
IP router
They are delivered to appropriate output ports by a switch
6
Introduction
7
Outline
Introduction to Packet Switching
Fundamental Switching Concepts
Switching and Routing
Unicast and Multicast
Throughput and Speedup
Non-blocking and Output Contention
Cell-mode Switching and Packet-mode Switching
Switch Fabric Classification
Buffering Strategy in Crossbar-based
Switching
8
Fundamental Switching Concept
9
Fundamental Switching Concept
Typical switching example
10
Fundamental Switching Concept
Routing
Represents a large-scale perspective, where
information is exchanged between two nodes
that can be separated by a large distance in the
network.
Needs the cooperation of other network nodes,
and is based on a routing protocol.
11
Fundamental Switching Concept
Typical routing example
12
Fundamental Switching Concept
13
Fundamental Switching Concept
14
Fundamental Switching Concept
16
Fundamental Switching Concept
10 0
1 0 0 0 0 0 0
0 1
1 0 0 0 0 0 0 0
Internal blocking in a Clos network (5 to 3)
0 0 0 0 1 0 0 0 0
0 0 0 11 0 0 0 0 0
0 011 0 0 0 0 0 0
0 0 0 0 0 1 0 0 0
0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 0 1
17
Fundamental Switching Concept
An N x N crossbar switch
consists of a 2-dimensional
array of N2 cross-points,
one corresponding to each
input-output pair.
18
Fundamental Switching Concept
Output Contention
is due to the nature of bursty IP traffic, but not
the design of switch architecture.
if more than one input port requests the same
output port, only one can be granted, so the
others have to wait.
the output contention resolution greatly affects
the overall utilization of the switching fabric.
19
Fundamental Switching Concept
Output port contention
20
Fundamental Switching Concept
Cell-mode Switching and Packet-mode Switching
Cell-mode Switching
in ATM networks, information is packed into fixed-size cells.
in an ATM switch, scheduling takes place per cell time, which is
called the time slot.
But in IP networks, variable-length packets are adopted
21
Fundamental Switching Concept
Packet-mode Switching
A packet is divided into fixed-size cells like in
cell-mode , but cells of the same packet are
scheduled contiguously, not independently.
Advantage
Both the scheduling of cells and the reassembly of
packets become much easier.
It benefits from lower average packet delay towards
cell-mode scheduling when the variance of packet
length distribution is small.
23
Outline
24
Switch Fabric Classification
Classification of switching architecture
The switch fabrics can be classified based on their switching
Packet Switches
Shared Shared
Single Path Multipath
Medium Memory
25
Switch Fabric Classification
Time-Division Switching
Advantage
Since every cell flows across the single
communication structure, it can be easily
extended to support multicast/broadcast
operations.
Disadvantage
Limited by the internal communication
bandwidth that should reach the aggregated
forwarding bandwidth of all the input ports.
26
Switch Fabric Classification
27
Switch Fabric Classification
Besides the bandwidth of shared medium,
the switch size is also limited by the
memory read/write access time, within
which N incoming and 1 outgoing cells in a
time slot need to be accessed.
Suppose
the time for a cell slot is Tcell
the memory access time slot is Tmem
the number of switch size N can be calculated
using the following equation:
Tcell
(N+1) ≤
Tmem
28
Switch Fabric Classification
The shared-memory structure
A memory unit is used to store The memory contains N logical
cells from all input ports. queues, one per output port,
and each queue stores the cells
destined for the corresponding
output.
29
Switch Fabric Classification
Shared-memory switch
Compare with shared-medium switches
Same
Cells from all the input ports are time-division
multiplexed into a cell stream
Difference
They are not forwarded immediately and flood to all
the outputs.
Instead, cells are buffered in a central memory, and
then scheduled to the de-multiplexer, where cells
destined for different output ports are separated.
30
Switch Fabric Classification
Shared-memory switch
Advantage
The memory is shared by all the output ports
and the utilization of memory is maximized.
Disadvantage
In each time slot the memory must
accommodate N concurrent write accesses and
N concurrent read accesses, which limits the
switch to a small size.
31
Switch Fabric Classification
Space-Division Switching
One or multiple data paths is available between
the input and the output ports.
Cells of the different input-output connections can
be forwarded concurrently, when no blocking is
present.
Total switching capacity is then the product of the
bandwidth of each path and the number of paths
that can transmit cells simultaneously.
The SDS switches are classified based on the
number of available paths between any input-
output pair.
Single-path switches
Multiple-path switches
32
Switch Fabric Classification –
Single-path switches
Only one path exists for any input-output
pair.
Has simpler routing control.
Classified into
crossbar-based switches.
fully interconnected switches.
banyan-based switches.
33
Switch Fabric Classification –
Single-path switches
Crossbar Switch Basically, an N x N crossbar switch
consists of a 2-dimensional array
34
Switch Fabric Classification –
Single-path switches
The self-routing property
The bar state of a cross-point can be triggered
individually by each incoming cell when its
destination matches with the output address.
No global information about other cells and their
destinations is required.
The control complexity is significantly reduced in the
switching fabric as the control function is distributed
among all cross-points.
The crossbar switch allows the transferring
of up to N cells with different input ports
and different outputs in the same time slot.
35
Switch Fabric Classification –
Single-path switches
36
Switch Fabric Classification –
Single-path switches
37
Switch Fabric Classification –
Single-path switches
38
Switch Fabric Classification –
Single-path switches
Fully interconnected switches
The complete connectivity between inputs
and outputs is usually accomplished by
means of N separate broadcast buses from
every input port to all output ports
N separate buffers are required in such a
switch, one at each output port.
39
Switch Fabric Classification –
Single-path switches
Operates in a similar manner to the shared
medium switch. A cell from any input port is
broadcast to every output port.
The difference is that multiple cells from
several input ports can be simultaneously
broadcasted to every output ports.
Therefore, separate cell filters and dedicated
buffers, one for each output port, are
required to filter out the mis-delivered cells
and to temporarily store the properly
destined cells.
40
Switch Fabric Classification –
Single-path switches
Advantage
Its simple and nonblocking structure, similar to
the crossbar-based switch.
Disadvantage
Space overhead requirement of the total
separate N 2 broadcast buses.
41
Switch Fabric Classification –
Single-path switches
Three different topologies of banyan-
based switches
42
Switch Fabric Classification –
Single-path switches
Banyan-based Switches
single-path space-division switches.
self-routing switch constructed from 2 x 2
switching elements with a single path
between any input-output pair.
there are three isomorphic topologies:
delta, omega, and banyan networks,
belonging to banyan-based family.
All of them offer an equivalent performance.
43
Switch Fabric Classification –
Single-path switches
Banyan switching structure
Example: 0
000
A cell in input port 1 destined 001
for output port 2, ('010') is
1
rounded to the top link (0) of
the 1st switch element, to the 0
010
down link (1) of the 2nd 011
switch element, and to the
top link (0) of the third switch
element. 100
101
110
111
44
Switch Fabric Classification –
Single-path switches
Banyan-based Switches
An NxN banyan network consists of log2N stages
of 2x2 switching elements.
Each stage has multiple 2 x 2 switching elements
whose inputs and outputs are labeled 0 and 1.
The inter-stage connection patterns are such that
only one path exists between any input of the first
stage and any output of the last stage.
45
Switch Fabric Classification –
Single-path switches
The banyan network has self-routing
capability.
When a cell enters in one of the input ports, it
uses the output address as the routing
information when traversing all the stages.
Advantage
It has fewer switch elements than the crossbar-
based and the fully interconnected switch,
O(NlogN) vs. O(N 2).
Self-routing property is also an attractive
feature in that no control mechanism is needed
for routing cells.
Disadvantage
It is an internally blocking switch. 46
Switch Fabric Classification –
Multiple-path switches
47
Switch Fabric Classification –
Multiple-path switches
More than one path exists for any input-
output pair.
Has better connection flexibility and fault
tolerance.
classified as
Augmented banyan switches
Clos switches
Multi-plane switches
Recirculation switches.
48
Switch Fabric Classification –
Multiple-path switches
49
Switch Fabric Classification –
Multiple-path switches
000 000
001 001
010 010
011 011
100 100
101 101
110 110
111 111
50
Switch Fabric Classification –
Multiple-path switches
010 010
011 011
100 100
101 101
110 110
111 111
51
Switch Fabric Classification –
Multiple-path switches
Advantage
By adding augmented stages, the cell loss rate is reduced.
Disadvantage
Its complicated routing scheme. Cells are examined at every
hardware complexity.
52
Switch Fabric Classification –
Multiple-path switches
53
Switch Fabric Classification –
Multiple-path switches
Three-Stage Clos Switches
...
...
...
...
...
...
54
Switch Fabric Classification –
Multiple-path switches
Example of internal blocking in a three-
stage Clos switch
57
Switch Fabric Classification –
Multiple-path switches
Three-Stage Clos Switches
Worst case situation for blocking occurs if all of
the remaining n - 1 input lines and n - 1 output
lines are busy and are connected to different
middle-stage switch modules.
Thus a total of (n - 1) + (n - 1) = 2n – 2
middle-stage switch modules are unavailable
for creating a path from a to b.
However, if one more middle-stage switch
module exists, an appropriate link must be
available for the connection.
Thus, a three-stage Clos switch will be
nonblocking if m ≥ (2n − 2) + 1 = 2n − 1 .
58
Switch Fabric Classification –
Multiple-path switches
59
Switch Fabric Classification –
Multiple-path switches
Substituting into Nx
60
Switch Fabric Classification –
Multiple-path switches
Three-Stage Clos Switches
Advantage
Reduces the hardware complexity from O( N 2) 3in
the case of the crossbar-based switch to O( N 2 )
while the switch could be designed to be
nonblocking.
Provides more reliability since there is more
than one possible path through the switch to
connect any input port to any output port.
Disadvantage
Some fast and intelligent mechanism is needed
to rearrange the connections in every time slot
so that internal blocking can be avoided.
61
Outline
Introduction to Packet Switching
Fundamental Switching Concepts
Switch Fabric Classification
Buffering Strategy in Crossbar-based Switching
Output Queueing (OQ)
Crosspoint Queueing
62
Buffering Strategy in Crossbar-
based Switching
Since the burst arrival of traffic are likely to contend for
the output ports and/or internal links, buffers are
required to temporarily store cells that lose the
contention.
Crosspoint Queueing
63
Buffering Strategy in Crossbar-
based Switching
The output queueing structure
64
Buffering Strategy in Crossbar-
based Switching
Characteristics of output queueing structure
All arriving packets can be transferred to output port immediately
Since only one packet can be transmitted via the output link at a
time, the remaining packets are queued at the output port.
Scheduling is needed at the output of output queues.
Advantage
OQ switch is working conserving: an output line is always busy
when this is a packet destined to it.
Highest throughput, lowest average latency
Disadvantage
Suffers from memory speed constraint and thus switch size
limitation.
Memory must run at the speed of N+1 times of the line rate (N times
writes, and once reads)
As the line speed or the switch port number increases, this scheme
becomes practically impossible to implement.
65
Buffering Strategy in Crossbar-
based Switching
Scheduling algorithms in OQ Switch
The purpose of the scheduling algorithms is to guarantee
the fairness among flows, and control the
throughput/latency of flows
Some typical OQ scheduling algorithm (already learned in
last week)
FIFO
Weighted Round Robin
Max-Min Fairness
GPS (Generalized Processor Sharing)
PGPS (Weighted Fair Queuing)
DRR (Deficit Round Robin)
Self-Clocked Fair Queuing (SCFQ)
Worst-Case Fair Weighted Fair Queuing (WF2Q)
…
66
Buffering Strategy in Crossbar-
based Switching
Example (memory bandwidth)
Assume we have an 8x8 OQ switch.
We use DDR SDRAM to store the packets.
The frequency of system clock is 200MHz, so the frequency of
data rate is 400MHz.
The data bus width of SDRAM is 64bit.
The peak throughput of the SDRAM is 25.6Gbps
In OQ switch, the memory must run at 9 times as fast as
line rate, so the maximum line rate supported in this OQ
switch is 2.8 Gbps
If the port number increases, the maximum line rate
supported will decrease.
67
Buffering Strategy in Crossbar-
based Switching
The input queueing structure
Head of Line
Blocking!
Scheduler 68
Buffering Strategy in Crossbar-
based Switching
IQ switch with
HoL blocking
OQ switch
69
2 − 2 ≈ 58%
Buffering Strategy in Crossbar-
based Switching
Advantage
Much more scalable than OQ: memory only need to run as twice
fast as the line rate (once read, once write) regardless of the
switch size
Eliminate HOL blocking.
Disadvantage
Since the logical queue number is N times the input queueing
structure, scheduling algorithm is much more sophisticated.
71
Buffering Strategy in Crossbar-
based Switching
Example (memory bandwidth)
Assume we have an 8x8 IQ switch.
We use DDR SDRAM to store the packets.
The frequency of system clock is 200MHz, so the frequency of
data rate is 400MHz.
The data bus width of SDRAM is 64bit.
The peak throughput of the SDRAM is 25.6Gbps
In IQ switch, the memory must run as twice fast as line
rate, so the maximum line rate supported in this IQ switch is
12.8 Gbps (2.8 Gbps for OQ switch)
With the increasing of port number, the maximum line rate
supported will not change
72
Buffering Strategy in Crossbar-
based Switching
VOQs Departing
First
SecondSwitching
Scheduling
Switching
Scheduling
Arriving
Scheduler
73
Buffering Strategy in Crossbar-
based Switching
What do we need in a switch with speedup between
1 and N
Input Queues
Output Queues
Schedulers at both crossbar and output side
Advantage
A tradeoff between complexity and performance
High performance can be easily achieved by simple
scheduling algorithm
Disadvantage
Memory must run at the speed of S+1 times of the line rate
(S is the speedup)
74
Buffering Strategy in Crossbar-
based Switching
76
Buffering Strategy in Crossbar-
based Switching
Advantage
Achieve the same performance as the OQ
switches because of no HOL blocking.
Disadvantage
Since there are N 2 discrete buffers, the sharing
effect is very poor.
With limited chip space, only a few cells can be
implemented.
As a result, this switch usually combines VOQs at
each input.
77