Based on the above terminology we state below a set of useful identities, for any switch Sk,j , in an N x N CNET: ( N - 2) > k > 0 (1) { H L }= {LH} {H}"{L}" = {L}"{H}" m, n < rnin (k, N - 2 - k) (2) = {H}m{L}m m < rnin (k, N - 2 - k ) ( 3 ) {HL}" = { H m m o d N } { p m o d N } {HL}" m < rnin (k, N {HL}" Lm mod N = { H L }m mod N {H,r}mmodN - 2 - k ) (4) for all m (5) = (HLJmmodN reduction rule at level 0 (6) - L mmodN inverse reduction rule at level 0 (7) ON THE CNET: A REARRANGEABLY NO N B LOCK1NG 0PTICAL INTERCON N ECTlON NETWORK Indexing terms: Switching and switching circuits, Optical switching, Optical communications, Networks A planar, circular, rearrangeably nonblocking optical switching architecture using N ( N - 1) Ti : LiNbO, directional coupler switches is suggested. Various figures of merit like the signal-to-crosstalk ratio (SXR), insertion loss and fault tolerance are presented. The outline for a distributed routing algorithm in O ( N )time is given. We propose an interconnection network the CNET, with a hierarchical planar ring structure and N ( N - 1) optical switches, which is rearrangeably nonblocking (Fig. 1). The CNET and its properties are described first and its possible optical implementation is outlined later on. In an N x N CNET, there are N - 1 levels numbered from 0 to N - 2. Each switch Si, (jth switch in the ith level) is connected to two other switches, as shown in Fig. 1. Let the connection to level k + 1 from level k be a H-connection, and the connection to level k - 1 from level k be an L-connection. Any path @ is a sequence of H- and L-connections, denoted as {H}"{L}", called an 'operator sequence', for that source and destination pair. The sequence { X X X .. . (n terms)} is referred to as { X } " . The number of X operator sequences in a path @ is denoted by n(X, @), or simply n(X). Weighted X indicates the weighted sum w where w = weight associated with the position of X counted from the right with the LSB as 0. 1 A 03 Fig. 1 Digraph representation of 6 x 6 CNET (Ii = ith input; O i= ith output) For an N x N CNET to be nonblocking, there must be at least ( N - 1) levels of N switches each. Consider a digraph of degree 4 representing an N x N CNET with i levels (e.g. i = N - 1 in Fig. 1) with a cut QQ' across all i levels. We notice that cyclic shifts are permutations that require the longest paths, and therefore, the largest number of switches. We choose one such bijection from the family of permutations' d;, k : (a2, al) + (a2, nt-j)(al)), j < N (cyclic shifts within segments of size 2'"-j)) where n f ) : x + x + (k mod 2") (cyclic shift of amplitude k). For 0 < j < N , d;, can be realised with fewer than ( N - 1) levels of N switches each. However, for j = 0, k = N - 1, across QQ, there must be at least ( N - 1) edges, as there are (IV - 1) distinct inputs to be connected to (N - 1) distinct outputs on either side of QQ'. Thus there must be at least ( N - 1) levels in an N x N CNET. If a path @ is to be found from Sk, to Sk+,. j + m , then n(H, @) - n(L, @) = n, and there are at least (m mod N ) { H L } sequences in @ if n < 0, and at least ((m- n) mod N ) { H L } pairs if n > 0. If the difference in level is n, then there must be at least n level transitions, i.e. I n(H) - n(L)I = I n I. If n > 0, then n(H) > n(L) and if n < 0 then n(H) < n(L). Thus, n(H, @) - n(L, @) = n. If n, L level transitions to S,,,,, j + m are necessary, then they must occur from Sk,j + m . To reach Sk,j + m from Sk,j , at least (m mod N ) { H L } pairs are required. If nHtraversals are necessary to reach S,,,, j + m then they must occur from Sk,j + m - n . To reach S,, j + m - n from Sk, we require at least (m - n) mod N { H L } pairs. @', ..., for all n > 0 and n E I, be a set Let @ = of paths between any input-output pair. Let Si.xbe a switch in i = 1, ..., n. We choose the path that includes min (Si,x) for all E @, where the min (Sx,J function is defined as min (Sx,y , Sm,") = S x , y ,if x < m and min (S,...,,, Sx, = Sx,y if y < z. If more than one has the same minimum element Si, x , i.e. E a1and Si,xzE and Si,x l = rnin (a1) = min (a2), then we take the sum of all x over Si, x over all Si,x E as the deciding criterion, i.e. we choose that path which has the maximum weighted L. The routing strings are arranged in order of decreasing weighted L. Using eqns. 1-7, it is possible to arrive at the optimal weighted L routing string if the source and destination are specified. The value of the 'most significant bit' (either 'H'or 'L')in this string is taken as the routing bit and the switch is set accordingly. As regards fault tolerance, we assume a fault model as follows: any network component (switch or link) in levels 0 to N - 4 can fail, faults occur independently, link failures are equivalent to partial or total switch failures and retention of full access. The CNET is N ( N - 3) fault tolerant, as it can tolerate N faults in each of the N - 3 levels. To provide a basis for the comparison of switching architectures with various degrees of fault tolerance and numbers of switches, we introduce a figure of merit, the normalised fault tolerance (NFT), which is defined as the ratio of the maximum number of faults that can be tolerated to the total number of switches. On this basis, the NFT of the CNET is ( N - 3)/(N - 1). The upper bound on the number of passes is N . Optical switching architectures have been reviewed, in detail, by Spanke.' When implemented using integrated optic directional couplers3 as the basic switching elements, the worst case insertion loss for an N x N CNET is given by I, = (2N - 2) . L + 2 . W + (2N - 3) . G (in dB), where (al, ai, an), ai ai ai) (c L = directional coupler insertion loss, W = waveguide to fibre coupling loss and G = inter-switch waveguide loss. The differential loss is given by 1; = (2N - 3)L + (2N - 5)G, in dB (for N 2 2). The worst case insertion loss occurs for path AB for a 6 x 6 CNET, while minimum loss occurs for the path A’B’ (Fig. 1). In a nonfaulty CNET, we find that the worst case SXR will arise when the signal path passes through (2N - 2) switches (path AB as shown in Fig. 1). Thus, SXR = 10 log,, ((1 - 10x/10)(2N-2)/(1 - (1 - 10”’10)(ZN-2))}, where x is the crosstalk per switch. Typical differential loss and signal to crosstalk ratio values for the CNET as a function of the number of inputs are shown in Figs. 2 and 3, respectively. It may be mentioned here that if the CNET is faulty such that all switches in levels 0 to N - 4 are unusable, the signal path would run along the two outer-most levels. Under such fault conditions only one path would be set up at a time. The remaining N - 1 paths would be set up in successive passes. L:07 / the width of each level is Wd, interswitch waveguide length is a fraction ‘a’ of the device length ‘D’, the number of devices that can be put in level 0 is given by N = n/{2 sin-’ (D(l + a)/ 2(R - (N - I)&)}, where R is the radius of the outermost level. For N = 32, u = 0.2 and D = 6mm, R is equal to 73.34 mm and thus the above configuration can be realised easily. 4Or 31 c\ ”:: ‘ ‘06 / N , n u m b e r of Inputs 1891131 Fig. 3 Differential loss against number of inputs However, due to the narrow annular geometry, much of substrate real estate at the centre of the wafer goes unutilised. This area could be effectively utilised for putting local electronic switch controllers if a suitable optoelectronic substrate like GaAs/AlGaAs or InP/GaInAsP is chosen. 0 12 24 36 N, number of inputs 40 60 189112/ Fig. 2 Signal to crosstalk ratio against number ofinputs G = 0,4dB/cm W = 2.0dB With a system budget of say 30dB and a 5dB system margin, we can have a worst case insertion loss of 25dB. Hence, from Fig. 2, taking L = 0.3dB, the maximum subsystem size would be a 32 x 32 CNET. For an SXR of 12dB, the permissible subsystem size can be 32 x 32, assuming x = -30dB. This is significantly larger than other reported optical switching architectures realised using directional couplers. Assuming that the substrate wafer diameter is 76 mm, 24th November 1988 S. KAR A. SELVARAJAN Department of Electrical Communication Engineering Indian Institute of Science blangalore-560 012, India References 1 LENFANT, I.: ‘Parallel permutations of data: a Benes network control algorithm for frequently used permutations’, IEEE Trans. Comput., 1983, pp. 637-647 2 SPANKE, R. A , : ‘Architectures for guided-wave optical spaceswitching systems’, IEEE Commun. Mag., 1987,25, pp. 42-48 R. c.: ‘Ti : LiNbO, integrated optic 3 KOROTKY, s. K.,and ALFERNFS, technology’, in HUTCIIESON, L. D. (Ed.): ‘Integrated optical circuits and componentsdesign and applications’ (Marcel Dekker, New York, 1987), pp. 169-227