## Abstract

We propose a most economical design of the *Optical Shared MemOry Supercomputer Interconnect System (OSMOSIS)* all-optical, wavelength-space crossbar switch fabric. It is shown, by analysis and simulation, that the total number of on-off gates required for the proposed *N* × *N* switch fabric can scale asymptotically as *N *ln* N* if the number of input/output ports *N* can be factored into a product of small primes. This is of the same order of magnitude as Shannon’s lower bound for switch complexity, according to which the minimum number of two-state switches required for the construction of a *N* × *N* permutation switch is log_{2} (*N*!).

© 2012 OSA

## 1. Introduction

High-performance computing (HPC) systems typically use thousands of microprocessors grouped into clusters, connected among themselves and with several TBytes of distributed memory using an interconnection network. Contemporary HPC system interconnects are mostly based on electronic switch fabrics [1]. Given that HPC systems’ performance is anticipated to approach 1 Exaflop by 2020 [1], it is necessary to employ some form of photonic switching, in order to achieve a low-cost, low-latency, high throughput interconnection between the microprocessors and the shared memory.

The Optical Shared MemOry Supercomputer Interconnect System (OSMOSIS) project [2–8] was a joint research effort between Corning, Inc. and IBM. The purpose of the project was to design and demonstrate a scalable, commercially available, hybrid HPC interconnect architecture, combining electronics for buffering and storing with a *N* × *N* wavelength-space crossbar switch fabric for optical cell switching. The latter was implemented using a broadcast-and-select architecture, with fixed-wavelength transmitters and discretely-tunable, direct-detection receivers. Receiver tunability was achieved using optical multiplexer-demultiplexer (MUX/DMUX) pairs and 2*N*^{3/2} semiconductor optical amplifiers (SOAs) as on-off gates.

Since the future viability of HPC systems’ optical interconnects critically depends on the drastic reduction of the cost per input/output port, a drastic reduction of SOA-based on-off gates would be clearly beneficial.

In this paper, we propose a most economical design of a fast *N* × *N* all-optical, wavelength-space crossbar switch fabric for optical interconnects. The total number of on-off gates required for the proposed *N* × *N* switch fabric can scale asymptotically as *N *ln* N* if the number of input/output ports *N* can be factored into a product of small primes. This is of the same order of magnitude as Shannon’s lower bound for switch complexity, according to which the minimum number of two-state switches required for the construction of a *N* × *N* permutation switch is log_{2} (*N*!). This result is reminiscent of the reduction in computational complexity achieved by highly efficient algorithms, e.g., Quicksort [9].

The reduction in the number of on-off gates is achieved by hierarchical multiplexing and by optimizing the design of tunable, direct-detection receivers. Hierarchical multiplexing first groups wavelengths into wavebands, then, groups wavebands into second-order wavebands, and so forth. Hierarchical demultiplexing is performed at the tunable, direct-detection receivers using several successive selection stages, with progressively finer selectivity, in tandem. More specifically, each tunable direct-detection receiver contains a fast, lossless, discretely-tunable optical bandpass filter (also referred to as *wavelength selector*). The latter can be constructed using several stages of optical multiplexer/demultiplexer (MUX/DMUX) pairs and SOAs as on-off gates [10]. We show that an optimal multi-stage wavelength selector configuration exists, which exploits the periodicity of arrayed waveguide grating (AWG) MUX/DMUX transfer functions, in order to reduce the number of SOAs to a minimum.

Hierarchical multiplexing/demultiplexing, in conjunction with AWG MUX/DMUXs, is key to the minimization of SOA on-off gates, in the same way that hierarchical routing is used to minimize the length of routing Tables [11].

A drawback of the aforementioned configuration is that the optical signal passes through a longer concatenation of AWG MUX/DMUXs and SOAs than in the original OSMOSIS optical interconnect architecture [2]. Consequently, the proposed interconnect is more vulnerable to transmission effects, esp. in SOAs, i.e., self-gain and cross-gain modulation, and polarization dependent gain. However, by proper selection of the system design parameters, the impact of the aforementioned transmission impairments on the proposed optical interconnect can be limited.

The rest of the paper is organized as follows: In Section 2, the operating principle of the proposed optimal crossbar switch fabric architecture for optimal interconnects is illustrated. In Section 3, we present an approximate analytical model for the calculation of the optimal number of stages of the multiplexing hierarchy and the optimal partition of tributaries per stage, for a given number of channels, in order to minimize the number of SOAs. In the same Section, the validity of the analytical model is checked by exhaustive search of all possible switch realizations and different optical crossbar switch sizes. It is shown that the number of SOAs per receiver card can very closely approach the analytically calculated minimum, when the number of channels can be factored into a product of primes. Section 4 presents a simplified control algorithm for setting the states of the on-off gates of the optimal wavelength selector. Finally, in Section 5, we present an improved analytical model for the calculation of the optimal number of stages of the multiplexing hierarchy and the optimal partition of tributaries per selection stage, by taking into account both the cost of SOAs, as well as of AWG MUX/DMUXs.

## 2. Operating principle of the optical interconnect

The block diagram of the proposed wavelength-space crossbar switch is shown in Fig. 1(a)
. Each transmitter uses a different carrier frequency (from a set of *N* equally spaced frequencies). All channels are broadcasted to all receivers using a *N* × *N* star coupler. Each receiver can select a different channel from the received WDM optical signal.

A most rudimentary implementation of the *N* × *N* star coupler and the tunable receivers in the proposed architecture is shown in detail in Fig. 1(b). Each receiver card contains a discretely tunable channel selector [10], which is used to discriminate one out of the *N* received carrier frequencies, followed by a direct-detection receiver with no inherent frequency selectivity. The channel selector is composed of a MUX/DMUX pair and *N* SOAs as on-off gates. By biasing one out of the *N*SOAs above the transparency current level, it is possible to select one channel, while attenuating all others. Since there are, in total, *N* SOAs per channel selector and one channel selector per receiver card, the total number of SOAs required in this particular implementation of the optical interconnect is *N*^{2}.

As an example, assume *N* = 16. The schematic of the 16-channel selector is shown in Fig. 2(a)
. Figure 2(b) illustrates the operation of the 16-channel selector in the frequency domain.

The 1 × 16 DMUX spatially separates 16 channels at carrier frequencies *f*_{1} through *f*_{16.}(corresponding to wavelengths *λ*_{16} through *λ*_{1}, respectively), so that only one of the input channels appears at each of its output arms. The idealized power transfer function of all arms is shown as a series of rectangles. The shaded (red) rectangle shows the power transfer function of the selected channel. Broken (blue) rectangles indicate the power transfer functions of all other channels. The channel selected is denoted with a solid (red) arrow, whereas all other channels are denoted with broken (blue) arrows.

Several variants [12–16] of the original single-stage channel selector proposed in [10] appeared in the literature, which take advantage of the periodic nature of the transfer function of arrayed waveguide grating (AWG) MUX/DMUXs, in order to achieve the same functionality as in Fig. 2, while using a smaller number of SOAs.

For instance, a two-stage channel selector [14–16] was proposed, in order to reduce the number of required SOAs. The configuration of the aforementioned selector is illustrated in Fig. 3(a)
, for the case of 16 channels. The operating principle can be better understood in the frequency domain (Fig. 3(b)): First, notice that the carrier frequencies of the 16 transmitted channels are slightly different, compared to the channel plan of Fig. 2(b), i.e., they are grouped in four wavebands of four channels each. Wavelengths within each waveband and successive wavebands have equal spacing. However, there is a guardband between adjacent wavebands. The MUX/DMUXs of the first selection stage of the channel selector do not have to be AWGs. The bandwidth of their transfer function is ${B}_{W}\cong 4\Delta f,$ where Δ*f* is the channel spacing within a waveband. The second selection stage uses an AWG MUX/DMUX pair withfree spectral range *FSR*_{λ} equal to the waveband spacing. The bandwidths of the individual arms are wide enough to pass a single wavelength channel without significant distortion. At the first stage of the channel selector, one of the four wavebands is selected by appropriate biasing of one out of the four SOAs (red rectangle). At the second stage of the channel selector, one of the four wavelengths of the selected waveband is chosen (red arrow) by appropriate biasing of one out of the four SOAs (red rectangle).

In conclusion, the two-stage, 16-channel selector is functionally equivalent to the single-stage, 16-channel selector shown in Fig. 2(a). However, the two-stage selector of Fig. 3(a) uses only half of the SOAs required in the single-stage 16-channel selector shown in Fig. 2(a). Figure 4 and Fig. 5 illustrate why clustering/partitioning of wavelengths into wavebands and the aforementioned structure of the 16-channel selector in Fig. 3(a) are beneficial.

In Fig. 4(a), the hierarchical multiplexing shown in Fig. 3(b) is represented by a tree diagram. The highest-ranking tributary is the root of the tree, whereas the smallest-ranking tributaries (i.e., the individual wavelengths) are the leaves of the tree. Then, wavelength selection performed by the discretely-tunable, direct-detection receivers of the optical interconnect is analogous to the traversal of a tree from root to leaves.

Figure 4(b) shows why optical demultiplexers with periodic transfer functions are key components, in order to exploit the hierarchical organization of wavelengths. Due to the periodicity of the optical demultiplexer transfer functions, the original tree representing the multiplexing hierarchy collapses into an equivalent folded tree, where the number of nodes and edges is significantly reduced. While retrieving a single wavelength, the receiver has to make as many binary decisions as the number of edges emanating from all the nodes of the folded tree. A minimum number of SOAs is required when the original tree (before folding) is maximally symmetric.

Figure 5 shows an alternative implementation of the multiplexing/demultiplexing hierarchy of Fig. 4. At the transmitter, hierarchical multiplexing uses four fibers, carrying the same set of four wavelengths, instead of one fiber carrying four wavebands of four distinct wavelengths each. This reduces the number of used wavelengths and enables wavelength reuse. The waveband MUX/DMUX pair at the receiver of Fig. 4 is omitted here. The waveband MUX at the receiver of Fig. 4 is substituted by a combiner.

For a given number of wavelengths, there are several alternative hierarchical multiplexing schemes with different number of levels and multiplexing granularities (Fig. 6 ). These alternative hierarchies can be represented by trees with a different number of levels, nodes, and edges, given the same number of leaves. From Fig. 6 we observe that the optimal multiplexing granularity is achieved when an equipartition or a quasi-equipartition of tributaries into the wavebands of different orders exists. Then, wavelength selection (i.e., tree traversal) can be extremely efficient.

It is worth pointing out several important details in the above example:

- • DMUXs spatially separate the constituents of the WDM signal they receive at their input. This implies that the role of a DMUX is to transform a WDM signal into a spatial division multiplexed (SDM) signal. Alternatively, for the last multiplexing stage at the transmitter, one could use SDM of several fibers. In this case, the DMUX corresponding to the first channel selector must be omitted.
- • When SDM of several fibers is used in place of the last level of WDM multiplexing at the transmitter, all fibers can carry the same set of wavelengths (wavelength reuse). This is desirable, since it reduces the number of different transmitters required for the implementation of the optical interconnect architecture. However, it is not necessary for the fibers to carry the same set of wavelengths. They can carry adjacent wavelength sets, as long as the spacing of subsequent wavelength sets is equal to the FSR of the MUX/DMUXs of the subsequent selection stage of the discretely tunable receiver.
- • The channel selection stages act as composite filters (if we neglect the nonlinear behavior of the SOAs). Therefore, their order can be interchanged without significant change in the performance of the multi-stage channel selector.
- • The DMUX used at the first selection stage of a multi-stage channel selector can always be aperiodic. When SDM of several fibers is used at the last multiplexing level of the transmitter instead of WDM, and if the same set of wavelengths is repeated in all fibers, the DMUX that follows the fiber selection stage can be aperiodic.

## 3. Formulation of the problem

A natural extension of the previous example is to further increase the stages of the channel selector until the number of required SOAs is reduced to a minimum. The question arises as to what the optimum number of stages is and the optimum number of tributaries (i.e., subsets of channels) selected by each consecutive stage of the channel selector is, in order to minimize the number of the required SOAs.

The aforementioned question can be formulated as a constrained minimization problem, where the cost function to minimize is the number of SOAs, the constraint is the number of channels, and the (discrete) variables to optimize are the number of selection stages and the number of tributaries per stage. The originality of the current study lies in the mathematical formulation of this optimization problem and its approximate analytical solution, using the method of Lagrange multipliers [17]. The analytical results are verified by exhaustive search of all possible realizations of 64 × 64, 72 × 72, 96 × 96 and 256 × 256 switch fabrics and by direct enumeration of the number of SOAs required in each realization.

#### 3.1 Analytical model: Optimization of the number of tributaries

Consider that the transmitted channels are grouped into a multiplexing hierarchy of $K$ levels (Fig. 7(a)
). At each level *i* of the multiplexing hierarchy, *n _{i}* tributaries (i.e., wavelengths, wavebands of different orders, optical fibers) are grouped together (Fig. 7(a)).

Then, the total number of transmitters $N$ is factored as follows:

The number of tunable receiver cards is equal to the number of transmitters $N.$ At each receiver card, there are $K$ selection stages with*n*,

_{i}*i*= 1,…,

*K*, SOAs each (Fig. 7(b)).

The number of SOAs per receiver card Ω is

The total number of SOAs that must be used in the optical interconnect is thenWe want to minimize Eq. (2) subject to the constraint Eq. (1), given a fixed number of channel selection stages*K*.

There are several different approaches for solving integer optimization problems [18]. For instance, for multiplexing hierarchies up to two levels, the integer optimization problem under study can be solved graphically (Fig. 8 ). In this Section, we apply a relaxation technique, following a methodology common to statistical physics [19]. In Section 3, we will apply an enumerative technique.

Relaxation techniques drop the integrality conditions and solve the resultant continuous optimization problem. In our case, this means that although *n _{i}* are positive integers, for the minimization, we assume they take values in a continuous range. Despite this approximation,it will be shown by example that the results hold in the discrete case as well, at least for the order of magnitude of

*N*of practical interest.

Taking the differential of Eq. (2) yields:

Taking the differential of Eq. (1) yields:*λ*[17]

*K*equationswhere

*i*= 1,…,

*K*.

This important result indicates that the number of SOAs is minimized when the number of tributaries in all *K* channel selection stages is equal (*equipartition of tributaries*).

The value of the Lagrange multiplier *λ *can be calculated by replacing Eq. (9) in Eq. (1)

*K*channel selection stages is calculated by replacing Eq. (10) into Eq. (2)The minimum total number of SOAs required for

*K*channel selection stages is

#### 3.2 Analytical model: Optimization of the number of channel selection stages K

Equation (11) and Eq. (12) were derived assuming a fixed number of channel selection stages *K.* It would be desirable to optimize the number of channel selection stages *K* in order to achieve a global minimum of the number of SOAs.

Although *K *is a positive integer, in the following, it is assumed to take values in a continuous range. The optimum number of channel selection stages *K _{opt}* can be calculated by taking the derivative of Eq. (11) or Eq. (12) with respect to

*K.*It is straightforward to show that

*e*is the constant $2.718\dots $.

There is a geometric interpretation of the results of the analytical model. Assume we have a K-dimensional rectangular hyperbox with edge lengths *n*_{1},…*n _{K}*. Equation (1) (i.e., the product of the length of all edges) can be interpreted as the “volume” of a hyperbox, whereas Eq. (2) (i.e., the sum of the length of all edges) can be interpreted as the “perimeter” of the hyperbox. Then, the constrained optimization problem under study can be stated as follows: we want to minimize the “perimeter” of a rectangular hyperbox in K-dimensions whose “volume” is given. In two dimensions, the answer to this problem is well known: the rectangle of minimum perimeter, when the area is fixed, is the square. The generalization of the problem in K dimensions shows that the hyperbox of minimum perimeter, when the volume is fixed, is the hypercube.

What is rather surprising is that the formal optimization indicates that among all hypercubes of the same volume N but with different number of dimensions K, the one with the globally minimum perimeter is when there are *K* = ln *N* dimensions and the edge length is equal to e.

The reader is cautioned that Eqs. (13)–(15) are approximate because of the discrete nature of *K.* In addition, the global minimum might not be achievable, because (i) *K _{opt}* is not, in general, an integer; and (ii) an equipartition of the tributaries into

*K*channel selection stages might not exist, as illustrated in the numerical examples in Section 3.

_{opt}In spite of their inaccuracy, Eq. (14) and Eq. (15), should be considered to be lower bounds to the number of SOAs required per receiver card and per interconnect, respectively.

#### 3.3 Analytical model: Figures of merit

For a given number of input/output ports (i.e., number channels *N*), numerous alternative discretely tunable receiver architectures may exist using different numbers of selection stages and tributaries per stage. The number of SOAs Ω required for each receiver architecture can be used as such a figure-of-merit.

To quantify the benefits of receiver architectures with a given number of input/output ports, one should adopt a normalized figure-of-merit. We propose the *optimality factor* defined as

_{min}is the theoretically minimum (albeit non-attainable) number of SOAs per receiver card given by Eq. (14), and Ω is the actual number of SOAs per receiver card given by Eq. (2). Obviously, 0 ≤

*F*≤ 1.

The number of SOAs required per receiver card, when using a single-stage channel selector, is equal to the number of channels *N.* The number of required SOAs per receiver card when using a multi-stage channel selector Ω is given by Eq. (2). An alternative figure-of-merit could be the *gain* in the number of SOAs defined as

#### 3.4 Model validation by direct enumeration

### Optimal architecture of a 64 × 64 optical crossbar switch fabric

In the following, without loss of generality, we assume that *N*, *n _{i}* are powers of two, i.e.,

*partition*: an additive partition of a positive integer

*N*′ is a set of

*K*strictly positive integers, whose sum is

*N*′. Obviously, additive partitions which correspond to a permutation of the

*n*′ yield the same number of SOAs. Therefore, the order of terms in Eq. (20) is disregarded (

_{i}*unrestricted partitions*). The 11 unrestricted additive partitions of 64 channels and the number of required SOAs given by Eq. (2) are calculated using

*Mathematica*and are listed in reverse lexicographic order in Table 1 below. It is observed that, for a given number of channel selection stages

*K*, equipartition of tributaries, when possible, yields the minimum number of SOAs, as predicted.

The number of possible equipartitions of tributaries can be calculated by finding the divisors *m* of *N*′ so that

*m*= {1,2,3,6}. Neglecting the trivial case

*m*= 6, which corresponds to a single selection stage

*K*= 1, the other three possible equipartitions of tributaries are

From Table 1, it is observed that, from the aforementioned equipartition cases, the minimum number of SOAs occurs for *m* = 1(*K* = 6) and *m* = 2(*K* = 3) and is equal to Ω = 12 SOAs per receiver card. The approximate formulae Eq. (13), Eq. (14) predict *K _{opt}* = 4.16, Ω

_{min}= 11.3, respectively, which cannot be achieved in practice. It is worth noting that there are non-equipartition cases for

*K*= 4,5 that also yield a minimum number of SOAs per receiver card (Ω = 12).

In the following graph (Fig. 9
), we compare the number of SOAs given by direct enumeration (Table 1) (points) and the theoretical formula Eq. (11) (solid curve). It is observed that Eq. (11) is accurate when the number of selection stages is *K* = 1,2,3,6 because then, equipartitions of tributaries occur.

In the above numerical example, it is observed that there are four (sub-optimal) minima for *K* = 3 – 6 that correspond to Ω = 12 SOAs per receiver card. Obviously, it is preferable to organize the tributaries into tetrads (*K* = 3), since this reduces the numbers of required MUX/DMUX pairs, as well. This implementation is shown in Fig. 10(a)
and 10(b).

In conclusion, when the number of channels and the number of tributaries per stage are restricted to powers of two, the optimal number of tributaries is:

- - Four, when log
_{2}*N*is even. - - Two, for one selection stage and four, for all other selection stages, when log
_{2}*N*is odd.

In both cases, the minimum total number of SOAs required for the optical crossbar switch fabric architecture is ${\Omega}_{tot}^{\mathrm{min}}=2N{\mathrm{log}}_{2}N,$ which is sub-optimal (i.e., slightly higher than the minimum given by Eq. (15)).

Further validation of the theoretical model is done for the following cases: (i) when the number of ports is a realistically large power of two (256 × 256); and (ii) when the number of ports is not a power of two and there are no possible equipartitions of tributaries (72 × 72 and 96 × 96 optical crossbar switch fabrics). In Fig. 11(a)
-11(c), we compare the number of SOAs given by direct enumeration (points) and Eq. (11) (solid curve). It is observed that Eq. (11) is accurate when the number of selection stages *K* is such that equipartitions of tributaries occur.

The optimality factor is used to compare the results of direct enumeration for 64 × 64, 72 × 72, 96 × 96 and 256 × 256 optical crossbar switch fabrics (Fig. 11(d)). It is observed that in all aforementioned cases, it is possible to approach within 95% of the optimum, by appropriately choosing the number of selection stages. This is due to the fact that, for all the above cases, it is possible to partition the number of channels *N* into a product of small primes.

### Optimality comparison: General case

In the general case, the number of channels *N* can be factored into a product of *M* primes *p*_{1},…,*p _{M}* as follows

*m*denotes the multiplicity of the prime factor

_{i}*p*.

_{i}The minimum number of SOAs is achieved by using *K* selection stages, where

Then, the minimum number of SOAs is

Substituting Eq. (24) into Eq. (16) and Eq. (17), it is possible to calculate the optimality factor and the corresponding gain in the number of SOAs as a function of the number of channels for the optimal architectures, without performing an exhaustive search of all possible realizations of the switch fabric (Fig. 12(a) and Fig. 12(b), respectively). The results vary greatly, depending on if the number of channels*N*can be factored into a product of small primes or not. As we can see from Fig. 12(a), the worst performance is obtained for large primes. In practice, it is rather unlikely that the number of ports of an optical crossbar switch will ever be a large prime. Nevertheless, one could achieve a reduction of the number of SOAs given by Eq. (24), even in this unlikely case, by using the same multiplexing hierarchy as for the smallest following integer with a high optimality factor and simply leaving some wavelength slots unpopulated. Figure 12(a) indicates that all powers of three approach within 1% of the optimum, whereas powers of two approach within 5% of the optimum. This somewhat surprising result in favor of the powers of three is counteracted by the fact that the powers of two allow one to organize the tributaries into tetrads, and this reduces the numbers of required MUX/DMUX pairs, as well. Therefore, minimization of a more fair cost function, that would factor the relative cost of MUX/DMUX pairs as well, would favor the powers of two, as shown in Section 5. It is worth noting that, with the current cost function, values of $N$ which can be factored into a product of powers of two and three, exhibit a better optimality factor than the powers of two. Finally, Fig. 12(b) indicates that the use of an optimized multichannel selector generally leads to larger SOA savings as the number of channels increases, provided that the latter can be factored into a product of small primes.

## 4. Simplified control algorithm

In the case of *K* channel selection stages and equipartition of tributaries into *λ *tributaries/selection stage, it is possible to use the following simplified control algorithm for setting the states of the switches:

- i. Sequentially number the transmitters and receivers in the decimal numeral system starting from zero, using the tags
*l*,*m*, respectively (i.e.,*l*,*m*$\in \left[0,\dots ,N-1\right].$ - ii. Sequentially number the SOAs of each selection stage in the decimal numeral system starting from zero, i.e., in the range [0,…,
*λ*– 1]. - iii. Express the transmitter tags
*l*in base-*λ*form, i.e., (*λ*…_{K}*λ*_{1}), where*λ*is the most significant digit and_{K}*λ*_{1}is the least significant digit. - iv. Represent the currents of the SOA array of each selection stage by a unit column vector with zeros everywhere except for a ONE in the (
*j*+ 1) – th positionA ONE at the (

*j*+ 1) – th row of the*i*– th unit column vector*υ*means that the_{i}*j*– th SOA on–off gate of the*i*– th channel selection stage is ON, whereas all other SOA on–off gates are OFF. - v. For an interconnection between the transmitter-receiver pair (
*l*,*m)*, set the on-off SOA gates of the*K*channel selection stages of the*m*– th receiver using the following rule:

According to this rule, depicted in Fig. 13
, there is a ONE at the (*j* + 1) – th row of the unit column vector *υ _{i}* if the number

*λ*

_{1}corresponds to the (

*j*+ 1) – th position of the discrete interval ${\lambda}_{i}\in \left[0,\dots ,\lambda -1\right].$

It is worth noting that there is a similarity between this algorithm and one previously proposed for the control of the Mach-Zehnder interferometer chain [20].

The implementation of the aforementioned control algorithm, in the optimal 64 × 64 optical crossbar switch fabric, is shown in Fig. 14(a)
and 14(b). In order to establish a connection between the transmitter-receiver pair (37, 16), first the transmitter index is written in base-4 (i.e., 37 = (211)_{4}). Then, each one of the three digits of the base-4 representation is used to set the state of the SOAs of a corresponding selection stage (shaded (red) rectangles in Fig. 14(b) denote SOAs at the ON state). Figure 15
shows that, eventually, the desired channel is selected.

## 5. Generalization for other cost functions

The theoretical model of Section 3 can be generalized for a cost function that takes into account the price of a MUX or DMUX *A* and an SOA *B*

Using the method of Section 3, it is straightforward to show that Eq. (9) still holds, i.e., the new cost function is minimized when the number of tributaries in all *K* channel selection stages is equal (*equipartition of tributaries*).

The minimum cost is calculated by replacing Eq. (10) into Eq. (25)

The optimum number of channel selection stages*K*can be calculated by taking the derivative of Eq. (26), with respect to

_{opt}*K*It is straightforward to show thatwhere

*x*is the root of the transcendental equationand

*r*= (2

*A*/

*B*) is the relative cost between a MUX/DMUX pair and a SOA.

The global minimum number of SOAs per receiver card and total number of SOAs required per crossbar switch fabric are, respectively:

where*y*denotes the auxiliary variableFigure 16 shows the normalized channel selector cost (per unit SOA cost) for

*N*= 64 and

*r*= 0 – 2 for the most efficient (in terms of SOAs) realizations listed in Table 1. For

*r*= 0, it is assumed that the MUX/DMUX pairs have negligible cost. Then, channel selectors with

*K*= 3 – 6 stages cost the same (i.e., the equivalent price of 12 SOAs, in agreement with Fig. 9). For

*r*> 0, this degeneracy is lifted. It is observed that for all values of

*r*> 0, the most cost-efficient, discretely tunable receiver should have three selection stages.

Figure 17
shows similar results for *N* = 256 and *r* = 0 – 2 for the most efficient (in terms of SOAs) realizations listed in Table 1. It is observed that for all values of *r* > 0, the best discretely tunable receiver should have four selection stages.

## 6. Conclusions

The purpose of the present theoretical study was to optimize the architecture of an all-optical, wavelength-space crossbar switch fabric for HPC optical interconnects. The latter was implemented using a broadcast-and-select architecture, with fixed-wavelength transmitters and discretely-tunable, direct-detection receivers. Discretely-tunable, direct-detection receivers were implemented using multiple stages of AWG MUX/DMUXs, with progressively decreasing bandwidths and free spectral ranges, and semiconductor optical amplifiers (SOAs) as on-off gates. A significant reduction of the number of semiconductor optical amplifiers was achieved by exploiting the periodicity of the transfer functions of the AWG MUX/DMUXs. The total number of on-off gates required for the proposed switch fabric can scale asymptotically as *N *ln* N,* if the number of input/output ports can be factored into a product of small primes. This is of the same order of magnitude as Shannon’s lower bound for switch complexity.

The feasibility of the aforementioned optimal optical interconnects depends on the transmission impairments induced by the optical components. Nonlinear effects, esp. self-gain and cross-gain modulation, and polarization dependent gain in semiconductor optical amplifiers are the major limiting factors of the maximum number of channel selection stages. A separate article addressed these issues [21].

## Acknowledgments

The authors are indebted to Prof. S. Daskalakis, University of Patras, Greece, for her remarks on integer optimization, which helped to improve the clarity of the mathematical model, and to Mrs. L. M. Szemiot, for corrections in the syntax of the manuscript. The OSMOSIS project was sponsored by the U.S. Department of Energy as part of the Accelerated Strategic Computing Initiative (ASCI).

## References and links

**1. **The Top 500 Supercomputer Sites, http://www.top500.org.

**2. **R. Hemenway, R. R. Grzybowski, C. Minkenberg, and R. Luijten, “Optical-packet-switched interconnect for supercomputer applications,” J. Opt. Netw. **3**(12), 900–913 (2004). [CrossRef]

**3. **R. Luijten, W. E. Denzel, R. R. Grzybowski, and R. Hemenway, “Optical interconnection networks: The OSMOSIS project,” in *Proceedings of IEEE Lasers and Electro-Optics Society 17th Annual Meeting,* Rio Grande, Puerto Rico, (2004), 563–564.

**4. **R. Luijten, C. Minkenberg, R. Hemenway, M. Sauer, and R. Grzybowski, “Viable opto-electronic HPC interconnect fabrics,” in *Proceedings of ACM/IEEE SC2005 Conference on High Performance Networking and Computing*, Seattle, WA, USA, (2005).

**5. **R. Luijten, C. Minkenberg, B. R. Hemenway, and R. R. Grzybowski, “Implementation challenges in the OSMOSIS optical HPC switch,” in *Proceedings of IEEE Lasers and Electro-Optics Society 19th Annual Meeting*, Montreal, CA, (2006), 623–624.

**6. **M. Sauer, R. Hemenway, R. Grzybowski, D. Peters, J. Dickens, and R. Karfelt, “A scaleable optical interconnect for low-latency cell switching in high-performance computing systems,” Proc. SPIE **6124**, 61240N, 61240N-12 (2006). [CrossRef]

**7. **I. Roudas, B. R. Hemenway, and R. R. Grzybowski, “Optimization of a supercomputer optical interconnect architecture,” in *Proceedings of IEEE Lasers and Electro-Optics Society 20th Annual Meeting*, Orlando, FL, (2007), 741–742.

**8. **R. P. Luijten and R. Grzybowski, “The OSMOSIS Optical Packet Switch for Supercomputers,” in *Proceedings of IEEE/OSA Optical Fiber Communication Conference*, San Diego, CA, (2009), pp. 1–3.

**9. **W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, *Numerical Recipes in C: the Art of Scientific Computing*, 2nd ed. (Cambridge University Press, 1992).

**10. **M. Zirngibl, C. H. Joyner, and B. Glance, “Digitally tunable channel dropping filter/equalizer based on waveguide grating router and optical amplifier integration,” IEEE Photon. Technol. Lett. **6**(4), 513–515 (1994). [CrossRef]

**11. **L. Kleinrock and F. Kamoun, “Hierarchical routing for large networks–performance evaluation and optimization,” Comput. Netw. **1**, 82–92 (1977).

**12. **O. Ishida, H. Takahashi, and Y. Inoue, “Digitally tunable optical filters using arrayed-waveguide grating (AWG) multiplexers and optical switches,” J. Lightwave Technol. **15**(2), 321–327 (1997). [CrossRef]

**13. **A. Misawa, K. Sasayama, and Y. Yamada, “WDM knockout switch with multi-output-port wavelength channel selectors,” J. Lightwave Technol. **16**(12), 2212–2219 (1998). [CrossRef]

**14. **N. Kikuchi, Y. Shibata, H. Okamoto, Y. Kawaguchi, S. Oku, H. Ishii, Y. Yoshikuni, and Y. Tohmori, “Monolithically integrated 64-channel WDM channel selector with novel configuration,” Electron. Lett. **38**(7), 331–332 (2002). [CrossRef]

**15. **N. Kikuchi, Y. Shibata, H. Okamoto, Y. Kawaguchi, S. Oku, H. Ishii, Y. Yoshikuni, and Y. Tohmori, “Monolithically integrated 64-channel WDM wavelength selective receiver,” Electron. Lett. **39**(3), 312–314 (2003). [CrossRef]

**16. **N. Kikuchi, Y. Shibata, H. Okamoto, Y. Kawaguchi, S. Oku, Y. Kondo, and Y. Tohmori, “Monolithically integrated 100-channel WDM channel selector employing low-crosstalk AWG,” IEEE Photon. Technol. Lett. **16**(11), 2481–2483 (2004). [CrossRef]

**17. **G. Arfken, *Mathematical Methods for Physicists*, 3d ed. (Academic Press, 1985).

**18. **G. L. Nemhauser and L. A. Wolsey, *Integer and Combinatorial Optimization* (Wiley, 1999).

**19. **M. A. Saad, *Thermodynamics: Principles and Practice* (Prentice Hall, 1997).

**20. **P. E. Green, *Fiber-Optic Networks* (Prentice Hall, 1993).

**21. **F. Karinou, I. Roudas, B. R. Hemenway, and R. R. Grzybowski, “Physical layer performance of HPC optical interconnect architectures,” *J. Lightwave Technol.*29(21), 3167–3177 (2011).