The present invention generally relates to interconnection networks, and in particular to a scalable switching fabric architecture allowing for the building of large switching networks from a common module.
Multi-stage Interconnection Networks (MIN) enable building large switches from smaller switches in a structured manner. A three stage (stages 120, 130 and 140) network is shown in
With reference to
The Clos network 100 of
The input stage 120 and output stages 140 of a Clos network consist of m switches. The middle stage 130 of network 100 consists of k switches. The prior art suggest systems with integrated input and output stages and separate middle stage. The problem with prior art is that the structure is not modular, that is the v(k, n, m) network of
Therefore, there is a need for constructing a Clos network v(k, n, m) from m identical modules, furthermore there is a need for constructing a modular and scalable Clos networks v(k, n, m), wherein network of different sizes (different values of m) can be constructed from m identical modules and allows building larger networks from a module by adding such modules as needed. The advantage of a modular structure is that it allows integration into a single module or monolithic Integrated Circuit (IC). The advantages of modular and scalable structure is that it allows building networks of different sizes from the same module or IC.
Briefly, an embodiment of the present invention includes an expandable network comprising of modules having switches wherein the modules are identical. Furthermore, a method for building a network with varying sizes (different values of m) from a common module is disclosed.
These and other features and advantages of the present invention will become well understood upon examining the figures and reading the following detailed description of the invention.
Referring now to
The input stage 220, the middle stage 230, and output stages 240 of Clos network 200 of the present invention all consist of m switches. The switches 221, 231, and 241 are grouped together and are included in the module 201. In a similar manner, the other modules 202, . . . , 209 include the grouping of switches (222, 232, 242), . . . , (229, 239, 249), respectively. The network 200 is built from identical modules 201, 202, . . . , 209 where each module includes groups of switches (221, 231, 241), (222, 231, 242), . . . , (229, 239, 249). Each module includes three switches an input switch of size (n×k), a middle switch of size (k′×k′) and an output switch of size (k×n)
A discussion is now presented regarding the selection of the parameter k′. With reference to
With each (k′×k′) switch, a network of q (m×m) switches, can be built wherein q is the quotient of dividing k′ by m(q=Q(k′/m) where Q (x/y) denotes the quotient of x divided by y). The m*q inputs and outputs of (k′×k′) switch are used and the (k′−m*q) remaining inputs and outputs are unused. The minimum number of equivalent (m×m) switches that is required in the second stage is k. Since in the second stage of network 200, there are m(k′×k′) switches, a network of m*q (m×m) switches can be built, therefore parameter k′ must be selected such that:
m*Q(k′/m)≧k Eq. 1
The connectivity of the middle stage 230 of the network 200 is now described. A simple way of describing the connectivity of the middle stage 230 is in terms of an equivalent virtual (m×m) switch. That is, the middle stage 230 is first transformed, into an equivalent virtual k(m×m) switches and the connectivity of the equivalent virtual k(m×m) switches is the same as that described previously. There are m(k′×k′) switches 231, 232, . . . , 239 and as described above, with each (k′×k′) switch, a network of q(m×m) switches is built. Starting with the first switch 231, the q*m input and output lines are assigned to the q equivalent (m×m) switches and any remaining input and output lines of the switch are unused. The virtual (m×m) switches are labeled 1, . . . , q, the assigned input lines are labeled I2ij. The index i identifies one of the virtual switches and the index j identifies one of the m input lines of the virtual switch, the assigned output lines are labeled O2ij. The index i identifies one of the virtual switches and the index j identifies one of the m outputs. The same process is repeated for the second switch 232, the virtual (m×m) switches are labeled (q+1), . . . , 2q, the assigned inputs and outputs are labeled I2ij and O2ij respectively similar to the first switch. This process is repeated until the last switch 239. With k′ satisfying Eq. 1, then at least k equivalent (m×m) virtual switches is constructed. The inputs lines of said virtual switches are labeled I2ij (1≦i≦k, 1≦j≦m) where the index i identifies one of the k virtual (m×m) switches and the index j identifies one of the m input lines of the switch. The output lines of the virtual switches are labeled O2ij (1≦i≦k, 1≦j≦m). The index i identifies one of the k (m×m) virtual switches and the index j identifies one of the m outputs. The interconnection between stages is as follows:
We have disclosed a v(k, n, m) switching network using m identical modules and a method of building a v(k, n, m) switching network from m identical modules.
Next, methods for building a scalable v(k, n, m) switching network from a common module will be disclosed. The method includes constructing a switching network 200 from common module 201 and selecting parameter k, k′and specifying a subset of integers such that networks with values of m belonging to (mε) can be constructed with the same common module.
Let the set of divisors of k be denoted by ={m|m divides k}. If k′ is selected to be equal to k(k′=k), it is obvious that m*Q(k′/m)=k for all mεk), and Eq. 1 is satisfied. Please note that the set at least includes 1, and k. Therefore one method is as follows:
k′=k and =={m|m divides k} (A1)
Let m1, m2, . . . denote divisors of k other than 1, and k. If k is prime then the set {mi} is empty. With k′ and according to A1 then Eq. 1 is satisfied for all (mεk), and network of size {n, (m1)*n, (m2)*n, . . . , (k)*n}can be constructed from identical modules 201 as described previously.
Another method is as follows:
k′=k=as, and ={m|m=ar(1≦r≦s)} (A2)
With k′ and according to A2 then Eq. 1 is satisfied for all (mε), and network of size {n, (a1)*n, (a2)*n, . . . , (as)*n} can be constructed from identical modules 201 as described previously. In practice, integer a is typically two (2) and k=2s, and m=2r (1≦r≦s). For example, select k′=k=32 =25. Then networks of size n, (21)*n, (22)*n, (23)*n, (24)*n and (25)*n can be constructed from identical modules 201 as described in the embodiment of present invention.
The above method is an exponentially scalable solution. The exponentially scalable method is perfectly acceptable method in certain applications. We now present a linearly scalable method.
Another method to satisfy Eq. 1 is as follows:
k′=k+(M−1), ={m|1≦m≦M}, (A3)
With k′ and according to A2 then Eq. 1 is satisfied for all (mε). To demonstrate this let k=q*m+r (0≦r≦m−1), and let k′=k+(M−1),
q*m≦k<(q+1)*m
If k is divisible by m then r=0, and m*Q(k′/m)=k and Eq. 1 is satisfied.
If k is not divisible by m then 1≦r≦(m−1), and (k+(m−1))≧ (q+1)*m,
Since m≦M therefore k′=k+(M−1)≧(q+1)*m, and Q(k′/m)≧(q+1), therefore
m*Q(k′/m)≧(q+1)*m≧k and Eq. 1 is satisfied.
With k′ according to A3 then you can build a network of size n, 2n, 3n, 4n . . . , M*n, that is you can build a linearly scalable Clos network up to M*n from identical modules 201 as described in the embodiment of present invention.
Several methods for constructing a scalable modular Clos network has been disclosed. Although the method has been described in terms of specific values for {k, k′, }, different values of {k, k′, } that satisfy Eq. 1 are considered to fall within the true spirit and scope of the invention.
In the method described, some inputs and outputs, specifically (k′−m*q) where q=Q(k′/m), of the 2nd stage of module 201 will be unused since there is not enough inputs and outputs to make an m×m switch. In another embodiment, the parameter k′ is selected to be equal to k. After assigning the m*q inputs and outputs, the remaining (k−m*q) inputs of the switch and the first ((m+1)*q−k) inputs and outputs of the next switch are assigned to next virtual (m×m) switch, the remaining inputs and output of the next switch are grouped into groups of m inputs and outputs, forming complete virtual (m×m) switches. With remaining inputs and outputs, the same process is repeated. Since there is no unused inputs of outputs, k′=k, however in this system, some of the virtual (m×m) switches are split between two modules. In order to implement the split virtual (m×m) switches in the middle switch, additional inputs and outputs between adjacent modules is included. Specifically, there is m bi-directional (or 2m unidirectional) signals between adjacent modules to allow operation of split switches.
Although the present invention has been described in terms of specific embodiments it is anticipated that alterations and modifications thereof will no doubt become apparent to those skilled in the art. It is therefore intended that the following claims be interpreted as covering all such alterations and modifications as fall within the true spirit and scope of the invention.
Number | Name | Date | Kind |
---|---|---|---|
4991171 | Teraslinna et al. | Feb 1991 | A |
4993018 | Hajikano et al. | Feb 1991 | A |
5038343 | Lebizay et al. | Aug 1991 | A |
5084867 | Tachibana et al. | Jan 1992 | A |
5103220 | Brunle | Apr 1992 | A |
5237565 | Henrion et al. | Aug 1993 | A |
5263121 | Melsa et al. | Nov 1993 | A |
5276425 | Swanson et al. | Jan 1994 | A |
5323386 | Wiher et al. | Jun 1994 | A |
5325090 | Goeldner | Jun 1994 | A |
5337308 | Fan | Aug 1994 | A |
5631902 | Yoshifuji | May 1997 | A |
5982770 | Sekine | Nov 1999 | A |
6157643 | Ma | Dec 2000 | A |
6418142 | Wolf | Jul 2002 | B1 |