THE TANDEM-CROSSPOINT SWITCH
246
9.2.3 Multicasting Operation
Next, we explain the multicasting mechanism in the TDXP switch. Unicast is a subset of multicast. For multicasting, Steps 1 and 3 are modified a little.
The remaining procedures, Steps 2, 4, and 5, are the same as above. The modified procedures, Steps 1
X
and 3
X
, for multicast are as follows.
X
Ž .
Step 1 A cell at the head of the input buffer sends request signals REQ to all the destination tandem crosspoints according to the routing bits
written in the cell header. The routing bits for multicasting are, for example, written using a bit-map scheme. Then go to Step 2.
Step 3
X
The cell at the head of the input buffer that sent REQ before is sent to all the destination crosspoints on the first switch plane, if no
NACK is received from any of the crosspoints polled to the input buffer within a certain time. Then go to Step 4. If even one NACK is
received by the input buffer within a certain time, the cell is not sent to any of the destination crosspoints; after the next cell time, go to Step 1
X
.
9.3 PERFORMANCE OF TDXP SWITCH
The performance of the TDXP switch was evaluated by event-driven com- puter simulation. The simulation programs were written using the C
language. The performance parameters of interest were the maximum throughput, the delay time, and the buffer size required to guarantee a
specified cell loss ratio. The maximum throughput is defined as the ratio of the total number of cells transmitted to output ports to the total number
of offered input cells. In the estimation of the maximum throughput, all offered input loads are set at 1.0.
Assume that cell arrival at N input ports follows a Bernoulli process. When the input traffic load is , an incoming cell arrives with probability
in a cell time, and there is no arrival with probability 1 y . The incoming cells are distributed uniformly to all output ports. The input traffic is
assumed to be homogeneous, and it is distributed uniformly to all input ports. Bernoulli traffic is considered. In addition, a simple arbitration mecha-
Ž .
nism for cell output in the switch round-robin arbitration is established between the crosspoints belonging to the same output port.
First we present the performance of the TDXP switch for unicasting traffic. Table 9.1 shows how many tandem switch planes K are needed to
obtain the maximum throughput in the TDXP switch architecture. To evalu- ate the maximum throughput, it is assumed that the sizes of the input and
output buffers are infinite. We can see that the maximum throughput is almost saturated with K s 2.
We thus conclude that K s 2 is large enough to obtain the maximum throughput. Therefore, in the following performance evaluation of the TDXP
PERFORMANCE OF TDXP SWITCH
247 TABLE 9.1
Maximum Throughput Determined by the Number K
a
of Tandem Switch Planes
Maximum Throughput K
N s 8 N s 16
N s 32 N s 64
N s 128 1
0.618 0.599
0.598 0.588
0.588 2
0.922 0.948
0.970 0.983
0.990 3
0.939 0.962
0.979 0.989
0.994 4
0.942 0.965
0.981 0.990
0.995
a
N s switch size.
switch, K is set to 2. In addition, we also show the performance of the conventional internal speedup switch with input and output buffers for
w x reference. The internal speedup factor L is set to 2 for rough equivalency 9 .
The maximum throughput of the TDXP switch increases with switch size N, and is higher than that of the double-speedup switch. Figure 9.5 compares
the maximum throughput with those of the internal double-speedup switch Ž
. L s 2 and the input buffering switch. The maximum throughput of the
TDXP switch decreases with small N, say N F 8, and increases with larger N. The reason is as follows. The probability P
that two successive cells at
suc
Ž .
Fig. 9.5 Comparison of maximum throughput. 䊚1997 IEEE.
THE TANDEM-CROSSPOINT SWITCH
248
the same input buffer are destined to the same output port is 1rN. When N is small, P
is large. As a result, the later cell in the input buffer is likely to
suc
Ž .
be blocked NACK is received because the first cell is still being handled in the tandem crosspoint. On the other hand, if N is large, P
is small, i.e., the
suc
blocking probability is small. The maximum throughput approaches 1.0 with N ™ ⬁, because of P
™ 0. In the double-speedup switch, the maximum
suc
throughput decreases to a certain value with increasing N, as is true for the input buffering switch. Thus, the maximum throughput of the TDXP switch
is higher than that of the double-speedup switch when N is large. The cell transmission delay of the TDXP switch is compared with that of
the double-speedup switch with N s 32. The maximum delay, defined as the 99.9 value, and the average delay are shown in Figure 9.6. The maximum
and average delay of the TDXP switch are almost the same as those of the Ž
. double-speedup switch with small F 0.85 . However, when the offered
load is larger than 0.85, the maximum and average delay values of the double-speedup switch increase strongly, because the limitation of the maxi-
mum throughput is smaller than that of the TDXP switch. In the TDXP switch, although the internal speed is the same as that of the input and
output lines and a cell is buffered in a tandem crosspoint before it is
Fig. 9.6 Delay performance.
PERFORMANCE OF TDXP SWITCH
249
Ž .
Fig. 9.7
Delay performance vs. switch size. 䊚1997 IEEE.
transmitted to the output ports, the delay performance is better than that of the double-speedup switch. This is simply because the TDXP switch has the
higher maximum throughput. The switch size dependence of the cell transmission delay is shown in
Figure 9.7. We can see that the maximum and average delay of the TDXP switch do not increase with N, while the delay of the double-speedup switch
does. This results from the lower HOL blocking probability of the TDXP switch with large N, as explained in Figure 9.5. Thus, the TDXP switch has
scalability in terms of N.
Figure 9.8 shows the required buffer sizes per port for an input buffer, an output buffer, and the sum of the input and output buffers, assuming a cell
loss ratio below 10
y 9
. As is true for the delay performance results in Figure 9.6, the total size of the TDXP switch is almost the same as that of the
double-speedup switch, but with offered loads higher than s 0.85, that of the double-speedup switch increases very strongly. Therefore, buffer re-
sources in the TDXP switch are used effectively.
Next, the results of the TDXP switch for multicasting traffic are pre- sented. The multicast cell ratio
is defined as the ratio of offered total
mc
input cells to offered multicast cells. The distribution of the number of copies is assumed to follow a binomial distribution.
THE TANDEM-CROSSPOINT SWITCH
250
Ž .
Fig. 9.8
Buffer size requirement. 䊚1997 IEEE.
Figures 9.9 and 9.10 show the blocking ratio B for multicasting traffic.
block
B is defined as the ratio of the offered input cells to the cells not
block
transmitted from the input buffer on account of HOL blocking. In this case, the offered input traffic load is set to 1.0. The multicast mechanism employed
is that presented in Section 9.2.3. The B of the TDXP switch is smaller
block
than that of the double-speedup switch with N s 32. The reason is as follows. In the double-speedup switch, if at least one of the destination
output ports is occupied by another cell belonging to a different input port, the multicast cell is blocked at the HOL in the input buffer. On the other
hand, in the TDXP switch, even when the destined output ports are so occupied, the cell is buffered in the tandem crosspoint, and an attempt is
made to send it through the next switch plane in the next time slot. Therefore, the following multicast cell in the input buffer can be transmitted
to all destination tandem crosspoints as long as none of them is buffering a cell. This benefit of the TDXP switch is large when N is large and the
Ž .
average number of copies ANC is smallᎏin other words, NrANC is large ᎏ
as shown in Figure 9.10. However, if NrANC is small, B of the TDXP
block
switch is higher than that of the double-speedup switch. For example, we can see such a case with N F 15 and ANC s 4.
PERFORMANCE OF TDXP SWITCH
251
Fig. 9.9
Blocking ratio for multicasting.
Fig. 9.10 Blocking ratio for multicasting vs. switch size and average number of
copies.
THE TANDEM-CROSSPOINT SWITCH
252
REFERENCES
1. G. Balboni, M. Collivignarelli, L. Licciardi, A. Paglialunga, G. Rigolio, and F. Zizza, ‘‘From transport backbone to service platform: facing the broadband
switch evolution,’’ Proc. ISS ’95, pp. 26, 1995. 2. T. Chaney, J. A. Fingerhut, M. Flucke, and J. S. Turner, ‘‘Design of a gigabit
ATM switch,’’ Proc. IEEE INFOCOM ’97, pp. 2᎐11, 1997. 3. M. Collivignarelli, A. Daniele, P. De Nicola, L. Licciardi, M. Turolla, and
A. Zappalorto, ‘‘A complete set of VLSI circuits for ATM switching,’’ Proc. IEEE GLOBECOM ’94, pp.134᎐138, 1994.
4. K. Genda, Y. Doi, K. Endo, T. Kawamura, and S. Sasaki, ‘‘A 160-Gbrs ATM switching system using an internal speed-up crossbar switch,’’ Proc. IEEE
GLOBECOM ’94, pp. 123᎐133, 1994. 5. M. J. Karol, M. G. Hluchyj, and S. P. Morgan, ‘‘Input versus output queueing on
a space-division packet switch,’’ IEEE Trans. Commun., vol. COM-35, pp. 1347᎐1356, Dec. 1987.
6. T. Kawamura, H. Ichino, M. Suzuki, K. Genda, and Y. Doi, ‘‘Over 20 Gbitrs throughput ATM crosspoint switch large scale integrated circuit using Si bipolar
technology,’’ Electron. Lett., vol. 30, pp. 854᎐855, 1994. 7. L. Licciardi and F. Serio, ‘‘Technical solutions in a flexible and modular architec-
Ž .
ture for ATM switching,’’ Proc. Int. Symp. Commun. ISCOM ’95 , pp. 359᎐366, 1995.
8. P. Newman, ‘‘A fast packet switch for the integrated services backbone network,’’ IEEE J. Select. Areas Commun., vol. 6, pp.1468᎐1479. 1988.
9. Y. Oie, M. Murata, K. Kubota, and H. Miyahara, ‘‘Effect of speedup in nonblock- ing packet switch,’’ Proc. IEEE ICC ’89, p. 410, 1989.
10. Y. Oie, T. Suda, M. Murata, D. Kolson, and H. Miyahara, ‘‘Survey of switching techniques in high-speed networks and their performance,’’ Proc. IEEE INFO-
COM ’90, pp. 1242᎐1251, 1990. 11. E. Oki, N. Yamanaka, and S. Yasukawa,‘‘Tandem-crosspoint ATM switch archi-
tecture and its cost-effective expansion,’’ IEEE BSS ’97, pp. 45᎐51, 1997. 12. E. Oki and N. Yamanaka, ‘‘A high-speed tandem-crosspoint ATM switch archi-
tecture with input and output buffers,’’ IEICE Trans. Commun., vol. E81-B, no. 2, pp. 215᎐223, 1998.
13. J. S. Turner, ‘‘Design of a broadcast packet switching networks,’’ IEEE Trans. Commun., vol. 36, pp. 734᎐743, 1988.
14. N. Yamanaka, K. Endo, K. Genda, H. Fukuda, T. Kishimoto, and S. Sasaki, ‘‘320 Gbrs high-speed ATM switching system hardware technologies based on copper-
polyimide MCM,’’ IEEE Trans. CPMT, 1996, vol.18, pp. 83᎐91, 1995.
H. Jonathan Chao, Cheuk H. Lam, Eiji Oki Copyright 䊚 2001 John Wiley Sons, Inc.
Ž .
Ž .
ISBNs: 0-471-00454-5 Hardback ; 0-471-22440-5 Electronic
CHAPTER 10
CLOS-NETWORK SWITCHES
In Chapter 6 and Chapter 7, we have studied how to recursively construct a large switch from smaller switch modules based on the channel grouping
principle. In such a construction, every input is broadcast to each first-stage module, and the input size of first-stage modules is still the same as that of
the entire switch.
In this chapter we consider a different approach to build modular switches. Ž
. The architecture is based on the Clos network see Fig. 10.1 . Switch modules
are arranged in three stages, and every module is interconnected with every module in the adjacent stage via a unique link. In this book, the three stages
are referred as input stage, middle stage, and output stage, respectively. The modules in those stages are accordingly called input modules, central mod-
ules, and output modules. Each module is assumed to be nonblocking and could be, for example, one of the crossbar switches described previously.
Inputs are partitioned into groups, and only one group of inputs are con- nected to each input module, thereby reducing the size of each module.
One may wonder why we are not just considering a two-stage interconnec- tion network in which every pair of modules of adjacent stages are intercon-
nected with a dedicated link. In that case, no two cells can be simultaneously transmitted between any pair of modules, because there is just one path
between them. In a Clos network, however, two cells from an input module can take distinct paths via different central modules to get to the same
Ž .
module at the output stage see Fig. 10.2 . The central modules in the middle stage can be looked on as routing resources shared by all input and output
modules. One can expect that this will give a better tradeoff between the switch performance and the complexity.
253
CLOS-NETWORK SWITCHES
254
Fig. 10.1
A growable switch configuration.
Fig. 10.2 Routing in a Clos network.
ROUTING PROPERTIES AND SCHEDULING METHODS