Professional Documents
Culture Documents
Switching Systems
Switch structures
Jorma Kekalainen
208
Structure of a switch
Switches can categorized as
circuit switches and
Packet switches
209
Switch structures
Page 1
Lecture notes
Basic concepts
Accessibility
Blocking
Complexity
Scalability
Reliability
Throughput
210
Accessibility
A network has full accessibility ( connectivity)
when each inlet can be connected to each
outlet (in case there are no other I/O
connections in the network)
A network has a limited accessibility when the
above given property does not exist
Interconnection networks applied in todays
switch fabrics usually have full accessibility
211
Switch structures
Page 2
Lecture notes
Accessibility
Full accessibility
Limited accessibility
212
Blocking
Blocking is defined as failure to satisfy a
connection request and it depends strongly on
the combinatorial properties of the switching
networks
213
Switch structures
Page 3
Lecture notes
Blocking
Non-blocking - a path between an arbitrary idle inlet and
arbitrary idle outlet can always be established independent of
network state at set-up time
Blocking - a path between an arbitrary idle inlet and arbitrary
idle outlet cannot be established owing to internal congestion
due to the already established connections
Strict-sense non-blocking - a path can always be set up between
any idle inlet and any idle outlet without disturbing paths already
set up
Wide-sense non-blocking - a path can be set up between any idle
inlet and any idle outlet without disturbing existing connections,
provided that certain rules are followed.
These rules prevent network from entering a state for which new
connections cannot be made
215
Switch structures
Page 4
Lecture notes
Complexity
Complexity of an interconnection network is
expressed by cost index
Traditional definition of cost index gives the number
of cross points in a network
used to be a reasonable measure of space division switching
systems
Complexity
217
Switch structures
Page 5
Lecture notes
Scalability
Due to constant increase of transport links and data rates on
links, scalability of a switching system has become a key
parameter in choosing a switch fabric architecture
Scalability describes ability of a system to evolve with
increasing requirements
Issues that are usually matter of scalability
218
Example of scalability
A switching equipment has room for 20 line-cards and the
original design supports 10 Mbit/s interfaces (one per line card)
Throughput of switch fabric is scalable from 500 Mbit/s to 2
Gbit/s
When new line cards that each implement two 10 Mbit/s
interfaces are introduced, the interface logic may have to be
upgraded
When new line cards that implement a 100 Mbit/s interface (one
per line card) are introduced, the switch fabric has to be
upgraded (scaled up) to 2 Gbit/s speed and the interface logic
has to be upgraded to 100 Mbit/s speed
Buffering memories need to be replaced by faster (and possible
larger) ones
Larger number (>20) of line cards implies at least new physical
design
Increase of line rates beyond 100 Mbit/s means redesign of
switch fabric
219
Switch structures
Page 6
Lecture notes
Reliability
Reliability and fault tolerance are system measures that have an
impact on all functions of a switching system
Reliability defines probability that a system does not fail within
a given time interval provided that it functions correctly at the
start of the interval
Availability defines probability that a system will function at a
given time instant
Fault tolerance is the capability of a system to continue its
intended function in spite of having a fault(s)
Reliability measures:
MTTF (Mean Time To Failure)
MTTR (Mean Time To Repair)
MTBF (Mean Time Between Failures)
220
Throughput
Throughput gives forwarding/switching
speed/efficiency of a switch fabric
It is measured in bits/s, octets/s, cells/s,
packet/s, etc.
Quite often throughput is given in the range
(0 ... 1), i.e. the obtained forwarding speed is
normalized to the theoretical maximum
throughput
221
Switch structures
Page 7
Lecture notes
Switching mechanisms
A switched connection requires a mechanism that
attaches the right information streams to each other
Switching takes place in the switch fabric, the
structure of which depends on networks mode of
operation, available technology and required capacity
Communicating terminals may use different physical
links and different time-slots, so there is an obvious
need to switch both in time and in space domain
Time and space switching are basic functions of a
switch fabric
222
Space-Division Switch
In space-division switching, the paths in the
circuit are separated from one another
spatially.
This technology was originally designed for
use in analog networks but is used currently in
both analog and digital networks.
It has evolved through a long history of many
designs.
223
Switch structures
Page 8
Lecture notes
224
Crossbar Switch
Crossbar matrix introduces the basic structure of a
space switch
A crossbar switch connects n inputs to m outputs in a
grid, using electronic microswitches at each
crosspoint.
The major limitation of this design is the number of
crosspoints required.
To connect n inputs to m outputs using a crossbar
switch requires n
m crosspoints.
For example, to connect 1000 inputs to 1000 outputs
requires a switch with 1,000.000 crosspoints.
Such a switch is inefficient because only small
minority of crosspoints are use at any given time.
225
Switch structures
Page 9
Lecture notes
226
227
Switch structures
Page 10
Lecture notes
228
Time-Division Switch
Time-division switching uses time-division
multiplexing (TDM) inside a switch.
The most popular technology is called the time-slot
interchange (TSI).
229
Switch structures
Page 11
Lecture notes
Time-Division Switch
The previous switch combines a TDM multiplexer, a
TDM demultiplexer, and a TSI consisting of random
access memory (RAM) with several memory locations.
The size of each location is the same as the size of a
single time slot.
The number of locations is the same as the number of
inputs (in most cases, the numbers of inputs and
outputs are equal).
The RAM fills up with incoming data from time slots
in the order received.
Slots are then sent out in an order based on the
decisions of a control unit.
230
Time-slot interchange
Each input line wants to send data to an output line according to the
following pattern:
1 3 2 4 3 1 4 2
231
Switch structures
Page 12
Lecture notes
232
Time-slot interchange
233
Switch structures
Page 13
Lecture notes
234
235
Switch structures
Page 14
Lecture notes
236
Time-Space analogy
A time switch can be logically converted into a space
switch by setting time-slot buffers into vertical
position => time-slots can be considered to
correspond to input/output links of a space switch
237
Switch structures
Page 15
Lecture notes
Space-Space analogy
A space switch carrying time multiplexed input and
output signals can be logically converted into a pure
space switch (without cyclic control) by distributing
each time-slot into its own space switch
238
An example conversion
239
Switch structures
Page 16
Lecture notes
240
Time switches
size of switch memory (SM)
and control memory (CM)
grows linearly as long as
memory speed is sufficient,
i.e. SM + CM + input
buffering + output buffering
= 2 2 number of timeslots
a simple and cost effective
structure when memory
speed is sufficient
speed of available memory
determines the maximum
switching capacity
241
Switch structures
Page 17
Lecture notes
Multistage Switch
The solution to the limitations of the crossbar switch is the
multistage switch, which combines crossbar switches in several
stages.
In a single crossbar switch, only one row or column (one path) is
active for any connection.
So we need N x N crosspoints.
If we can allow multiple paths inside the switch, we can
decrease the number of crosspoints.
Each crosspoint in the middle stage can be accessed by multiple
crosspoints in the surrounding stages.
242
Multistage Switch
243
Switch structures
Page 18
Lecture notes
2.
3.
N
(n k ) + k N N + N (k n ) = 2kN + k N
n
n n n
n
244
Note
245
Switch structures
Page 19
Lecture notes
Example
Design a three-stage, 200 200 switch (N = 200) with k = 4 and
n = 20.
Solution
In the first stage we have N/n=200/20 or 10 crossbars, each of
size (nk)=20 4.
In the second stage, we have k=4 crossbars, each of size (N/n
N/n )= 10 10.
In the third stage, we have N/n=10 crossbars, each of size (k n)=4
20.
The total number of crosspoints is 2kN + k(N/n)2, or 2000
crosspoints.
Note. This is 5 percent of the number of crosspoints in a singlestage switch (200 200 = 40,000).
246
Blocking (1)
The multistage switch in the previous Example has one drawback
blocking during periods of heavy traffic.
The whole idea of multistage switching is to share the
crosspoints in the middle-stage crossbars.
Sharing can cause a lack of availability when the resources are
limited and all users want a connection at the same time.
Blocking refers to times when one input cannot be connected to
an output because there is no path available between them all
the possible intermediate switches are occupied.
247
Switch structures
Page 20
Lecture notes
Blocking (2)
In a single-stage switch, blocking does not occur because every
combination of input and output has its own crosspoint; there is
always a path.
Cases in which two inputs are trying to contact the same output do
not count. That path is not blocked; the output is merely busy.
248
Blocking (3)
In large systems, such as those having 10,000 inputs
and outputs, the number of stages can be increased
to cut down on the number of crosspoints required.
As the number of stages increases, however,
possible blocking increases as well.
Note. Many people have experienced blocking on public
telephone systems caused by overload of the system.
249
Switch structures
Page 21
Lecture notes
Clos criterion
Clos investigated the condition of nonblocking in multistage
switches and came up with the following result:
In a nonblocking switch, the number of middle-stage switches must
be at least (2n-1).
251
Switch structures
Page 22
Lecture notes
Note
252
Example
Redesign the previous three-stage, 200 200 switch, using the
Clos criteria with a minimum number of crosspoints.
Solution
We let n = (200/2)1/2, or n = 10. We calculate k = (2n 1)= 19. In
the first stage, we have N/n=200/10, or 20, crossbars, each
with (nk)= 10 19 crosspoints.
In the second stage, we have k=19 crossbars, each with (N/n
N/n )= 10 10 crosspoints.
In the third stage, we have N/n=20 crossbars each with (k n)=
19 10 crosspoints.
The total number of crosspoints is 20(10 19) + 19(10 10) +
20(19 10) = 9500.
253
Switch structures
Page 23
Lecture notes
254
Comment 2
A multistage switch that uses the Clos criteria and a
minimum number of crosspoints still requires a huge
number of crosspoints.
For example, to have a 100,000 input/output switch,
we need something close to 200 million crosspoints
(instead of 10 billion with the single stage).
This means that if a telephone company needs to
provide a switch to connect 100,000 telephones in a
city, it needs 200 million crosspoints.
The number can be reduced if we accept blocking.
Today, telephone companies use time-division
switching or a combination of space- and time-division
switches.
255
Switch structures
Page 24
Lecture notes
256
257
Switch structures
Page 25
Lecture notes
258
259
Switch structures
Page 26
Lecture notes
260
261
Switch structures
Page 27
Lecture notes
262
263
Switch structures
Page 28
Lecture notes
264
Time-space-time switch
265
Switch structures
Page 29
Lecture notes
Time-Space-Space switch
Time-Space-Space switch can be applied to
increase switching capacity
266
Space-Time-Space switch
Space-Time-Space switch has a high blocking
probability (like ST-switch) - not a desired
feature in public networks
267
Switch structures
Page 30
Lecture notes
268
Switch structures
269
Page 31
Lecture notes
270
271
Switch structures
Page 32
Lecture notes
272
Note:
Switch structures
273
Page 33
Lecture notes
274
Number of cross-points
Fan-out
Logical depth
Blocking probability
Complexity of switch control
275
Switch structures
Page 34
Lecture notes
Cross-points
Number of cross-points gives the number of on-off
gates in space switching equivalent of a fabric
Minimization of cross-point count is essential when crosspoint technology is expensive (e.g. electro-mechanical and
optical cross-points)
Very Large Scale Integration (VLSI) technology implements
cross-point complexity in Integrated Circuits (ICs) => more
relevant to minimize number of ICs than number of crosspoints
Due to increasing switching speeds, large fabric
constructions and increased integration density of ICs,
power consumption has become a crucial design criteria
higher speed => more power
large fabrics => long buses, fan-out problem and more driving
power
increased integration degree of ICs => heating problem
276
Switch structures
Page 35
Lecture notes
278
Blocking probability
Blocking probability of a multi-stage switching network is
difficult to determine
Lees approximation gives a coarse measure of blocking
Assume uniformly distributed load
equal load in each input
load distributed uniformly among intermediate stages (and their
outputs) and among outputs of the switch
279
Switch structures
Page 36
Lecture notes
Control complexity
Given a graph G, a control algorithm is needed to find and set up
paths in G to fulfill connection requirements
Control complexity is defined by the hardware (computation and
memory) requirements and the run time of the algorithm
Amount of computation depends on interconnection structure
and degree of blocking tolerated
In general, computation complexity grows exponentially as a
function of the number of terminals
There are interconnection networks that have a regular
structure for which control complexity is substantially reduced
There are also structures that can be distributed over a large
number of control units
280
281
Switch structures
Page 37
Lecture notes
Example
Suppose that driving current of a switching gate (cross-point) is
100 mA and its maximum input current is 8 mA
How many output gates can be connected to a bus, driven by one
input gate, if the capacitive load of the bus is negligibly small ?
Fan-out= floor[100/8]= 12
Management complexity
Network management involves adaptation and
maintenance of a switching network after the
switching system has been put in place
Network management deals with
Switch structures
Page 38
Lecture notes
Example
284
Erlang-B, also known as the Erlang loss formula, is a formula for the blocking
probability derived from the Erlang distribution to describe the probability of
call loss on a group of circuits (in a circuit switched network, or equivalent).
It is, for example, used in planning telephone networks.
The formula is not limited to telephone networks, since it describes a probability
in a queuing system (albeit a special case with a number of servers but no buffer
spaces for incoming calls to wait for a free server).
The formula applies under the condition that an unsuccessful call, because the
line is busy, is not queued or retried, but instead really lost.
It is assumed that call attempts arrive following a Poisson process
285
Switch structures
Page 39
Lecture notes
Erlang B
Switch structures
Page 40
Lecture notes
AN
P( K = N ) PN = N N ! n = E B ( N , A)
A
n = 0 n!
Erlangs blocking formula applies to systems fulfilling conditions
Example
E(5, 3) implies that we have a system of 5 inlets and
offered load is 3
289
Switch structures
Page 41
Lecture notes
Example
An exchange for 1000 subscribers is to be installed and it is
required that the blocking probability should be below 10 %. If
E1 links are used to carry the subscriber traffic to telephone
network, how many E1 links are needed?
- average call lasts 6 min
- a subscriber places one call during a 2-hour busy period (on the
average)
Amount of offered traffic is (2x1000x6 min /2x60 min) = 100
Erl.
Erlangs B formula gives N=97 for 10 % blocking and load of 100
Erl
=> required number of E1 links is ceil(97/30) = 4
(two incoming and two outgoing E1 links)
290
Calculation
291
Switch structures
Page 42
Lecture notes
Traffic capability
292
293
Switch structures
Page 43
Lecture notes
A=0. 1
A=0.2
A=0.5
A=1
A=2
A=3
A=4
A=5
A=10
A=15
A=20
A=30
A=40
A=50
9,0909091
0,4524887
0,0150807
0,000377
7,54E-06
1,257E-07
1,795E-09
2,244E-11
2,493E-13
2,493E-15
2,267E-17
1,889E-19
1,453E-21
1,038E-23
6,919E-26
4,325E-28
2,544E-30
1,413E-32
7,438E-35
3,719E-37
1,771E-39
8,05E-42
3,5E-44
1,458E-46
5,833E-49
16,666667
1,6393443
0,1091703
0,0054582
0,0002183
7,278E-06
2,079E-07
5,198E-09
1,155E-10
2,31E-12
4,201E-14
7,001E-16
1,077E-17
1,539E-19
2,052E-21
2,564E-23
3,017E-25
3,352E-27
3,529E-29
3,529E-31
3,361E-33
3,055E-35
2,657E-37
2,214E-39
1,771E-41
33,333333
7,6923077
1,2658228
0,1579779
0,0157953
0,0013163
9,402E-05
5,876E-06
3,265E-07
1,632E-08
7,419E-10
3,091E-11
1,189E-12
4,246E-14
1,415E-15
4,423E-17
1,301E-18
3,614E-20
9,51E-22
2,378E-23
5,661E-25
1,287E-26
2,797E-28
5,827E-30
1,165E-31
50
20
6,25
1,5384615
0,3067485
0,0510986
0,0072993
0,0009124
0,0001014
1,014E-05
9,216E-07
7,68E-08
5,908E-09
4,22E-10
2,813E-11
1,758E-12
1,034E-13
5,746E-15
3,024E-16
1,512E-17
7,2E-19
3,273E-20
1,423E-21
5,929E-23
2,372E-24
66,666667
40
21,052632
9,5238095
3,6697248
1,2084592
0,344086
0,0859476
0,0190958
0,003819
0,0006944
0,0001157
1,78E-05
2,543E-06
3,391E-07
4,239E-08
4,987E-09
5,541E-10
5,833E-11
5,833E-12
5,555E-13
5,05E-14
4,391E-15
3,66E-16
2,928E-17
75
52,941176
34,615385
20,610687
11,005435
5,2157115
2,1864315
0,8132439
0,2703484
0,0810388
0,0220966
0,0055238
0,0012747
0,0002732
5,463E-05
1,024E-05
1,808E-06
3,013E-07
4,757E-08
7,135E-09
1,019E-09
1,39E-10
1,813E-11
2,266E-12
2,72E-13
80
61,538462
45,070423
31,067961
19,906687
11,716247
6,2748943
3,0420058
1,3339673
0,5307549
0,19263
0,0641688
0,0197403
0,0056398
0,0015039
0,000376
8,847E-05
1,966E-05
4,139E-06
8,277E-07
1,577E-07
2,867E-08
4,985E-09
8,309E-10
1,329E-10
83,333333
67,567568
52,966102
39,834289
28,486782
19,184726
12,051864
7,0047852
3,7457786
1,838457
0,8287368
0,3441188
0,1321784
0,0471843
0,0157256
0,004914
0,0014453
0,0004015
0,0001056
2,641E-05
6,289E-06
1,429E-06
3,107E-07
6,473E-08
1,295E-08
90,909091
81,967213
73,206442
64,666322
56,395218
48,45149
40,904078
33,831843
27,320794
21,458234
16,323233
11,973919
8,4338863
5,6819143
3,6496945
2,2301872
1,2948875
0,7142438
0,3745099
0,186905
0,0889232
0,0404033
0,0175636
0,0073176
0,002927
93,75
87,548638
81,403763
75,324733
69,322726
63,411084
57,605723
51,925557
46,39294
41,034054
35,879157
30,962563
26,322168
21,998293
18,03164
14,460212
11,315292
8,616888
6,3695015
4,5593216
3,153945
2,1051476
1,3543285
0,8393506
0,5010868
95,238095
90,497738
85,781686
81,093135
76,435799
71,813995
67,232758
62,697966
58,216481
53,796317
49,446818
45,178854
41,005014
36,939785
32,999693
29,203347
25,571358
22,126034
18,890791
15,889196
13,143603
10,673395
8,492963
6,6096717
5,0221778
96,774194
93,555094
90,343305
87,139501
83,944435
80,758951
77,583993
74,420626
71,270044
68,133597
65,012811
61,909409
58,825348
55,762844
52,724418
49,712932
46,731641
43,784243
40,874938
38,008488
35,190279
32,42638
29,723599
27,089523
24,532532
97,560976
95,124851
92,691829
90,262127
87,835986
85,413666
82,995456
80,581668
78,172646
75,768771
73,370459
70,978169
68,592409
66,213741
63,842785
61,480232
59,126846
56,783479
54,451079
52,1307
49,823522
47,530861
45,254186
42,995141
40,755565
98,039216
96,079939
94,122254
92,166252
90,212032
88,259699
86,309367
84,361158
82,415205
80,47165
78,530647
76,592365
74,656983
72,7247
70,795729
68,870302
66,948673
65,031118
63,117939
61,209465
59,306058
57,408112
55,516063
53,630385
51,751602
294
295
Switch structures
Page 44
Lecture notes
Erlang B
296
297
Switch structures
Page 45
Lecture notes
298
Switch structures
Page 46
Lecture notes
A=g(N,EB)
B.H.T.
Blocking
Lines
1.000
0.010
2.000
0.010
3.000
0.010
4.000
0.010
10
5.000
0.010
11
10.000
0.010
18
20.000
0.010
30
30.000
0.010
42
40.000
0.010
53
50.000
0.010
64
100.000
0.010
117
150.000
0.010
170
301
Switch structures
Page 47
Lecture notes
Channels Servers
Users Calls
Calls arrive according to a Poisson process with rate =
Call inter-arrival time is an exponentially distributed random variable,
with mean 1/.
Note.
Note.
Efficiency
Efficiency = A/N for a given blocking probability,
PN=EB(N,A)
Typical planning values for cellular and conventional
fixed networks are PN=EB(N,A) = 1% or 2%
Example: for PN=EB(N,A) = 1%
303
Switch structures
Page 48
Lecture notes
Example Erlang B
A cellular provider owns 100 cell sites each having 60 channels.
Assuming each user makes three calls per hour and the average
holding time per call is 1 minutes. Determine the total number of
subscribers that the service provider can support with a blocking
probability less than 2%.
Solution:
N=60/cell
EB(60,A)= 2%
A=49 Erl/cell
Asubs=/=(3/60)1=0.05 Erl/subs
Number of subs = total traffic/traffic per subs=49/0.05=980/cell
Total number of subs = 100 980 = 98000 subs
304
N
A + N !1
N k = 0 k!
The average delay is given by
D = P(delay > 0 )
1
(N A)
Switch structures
Page 49
Lecture notes
Example Erlang C
A company has a system with five private telephone lines
connecting two of its sites. The number of employees is 60
where on average each employee makes a three-minute
telephone call every hour.
a) What is the average delay for an employee to get access to
the telephone?
b) What is the probability of an employee waiting more than one
minute for the access to the telephone?
Solution:
a) N=5, A=60(1/60)*3=3 Erl, =1/3
306
Example
a) The average delay
D = P(delay > 0 )
1
( N A)
D0.35 min=21 s
b) The probability of the delay exceeding 1 min
is given by
P(delay > 1 min ) = P(delay > 0) exp[ ( N A) 1 min ]
P(delay > 1 min ) 0.12
Note. Blocking is
Switch structures
307
Page 50