1. Reconfigurable optical interconnections using multi-permutation-integrated fiber modules JSAP conference, 27 March 2003 Alvaro Cassinelli *, Makoto Naruse **,***, Masatoshi Ishikawa *, and Fumito Kubota **. Univ. of Tokyo *, Communications Research Laboratory **, JST PRESTO *** output input Introduction. In previous work we explored a way to alleviate wiring congestion in massively interconnected multi-chip architectures using cascaded optoelectronic arrays and fiber-based, plane-to-plane (2D) optical interconnection modules [1]. The target application for these modules was packet-switching using a buffered, highly scalable multistage interconnection network. We study here multi-permutation modules containing a set of independent addressable permutations. Addressing can be done by minute mechanical displacement of the modules (figure). Cascading these modules without intermediate optoelectronic arrays gives a transparent multistage architecture adequate for circuit switching in weak-interconnected multiprocessors.

2. Introduction Multistage architecture: parallel computers, switching networks Dense optical interconnect: interconnection folded in 2D… Optical Multistage Architecture Paradigm +

3. Fiber-Modules vs. Free-Space Fiber have better efficiency (than holograms) for long-range interconnections. no cross-talk in 3D, just like free-space optics, No space-invariance possible. Theoretically more volume efficient than free-space Precise and robust alignment possible… multiple interleaved permutations possible. Maybe “hard” to build? Boring, but not a fundamentally difficult (can be automated, can be done by “layers”). Alignment of both output and input needed… Power dissipation may be a fundamental limitation, but we are far from these limits… 2D folded perfect shuffle permutation module  (2) …wave-guide arrays for fixed, point-to-point and space variant interconnections are an interesting alternative to free-space optics

4. Interconnection module … Elementary Processor Array VCSEL array Photo- detector array 2D input data flow Fixed inter-stage interconnections… FIXED interconnections Optoelectronic processing/switching …useful for pipeline processing of data (eg. FFT) or packet switching.

5. … or reconfigurable inter-stage interconnections Reconfigurable Interconnection module 2D input data flow High bandwidth transparent circuit-switched networks for permutation routing in multi-processors Reconfigurable Interconnection module c2c2 c1c1 c3c3 c4c4 16 processors interconnected using the four-dimensional hypercube topology. The network provides four cube permutations c 1, c 2, c 3, c 4 2D output data flow … One or more reconfigurable modules

6. Cascaded Multi-permutation Module Paradigm Interleaved fiber-based permutation modules: A C : Rem: Dynamic alignment is tightly coupled with dynamic reconfiguration of the interconnect. Cf. Naruse’s presentation. A C : Rem: Dynamic alignment is tightly coupled with dynamic reconfiguration of the interconnect. Cf. Naruse’s presentation. A small mechanical or optical perturbation can produce a drastic change of the interconnection pattern from input to output! Cascaded multi-permutation modules: This architecture simplifies module design (bi-permutations), while maintaining whole network interconnection capacity. Cascaded optical permutation modules output input {c 2, id} A multistage version of most direct topologies (hypercube, cube-connected-cycles, deBruijn) can be implemented using specially designed interconnection modules.

7. Unfolded Folded [ exchange  (k) ]  (k) {b n, … b k+1, b k, b k-1, … b 2, b 1 } If k  n/2, (  (1) and  (2) ) exchange only rows:  (1)  (2)  (3)  (4) …If k>n/2, (  (3) and  (4) ) exchange only columns. The modules are just the same than previous ones, rotated. Only two modules are needed… Module design: layered modules [slide not shown in main presentation] Example: exchange (cube) permutation for N=2 4

8. For N=4=2 1 x2 1 (n=2) we have :  (1) = id  (2) =  row (1).  col (1).L = L For N=8=2 2 x2 1 (n=3) we have :  (1) = id  (2) =  row (2) (new)  (3) =  row (2).  col (1).L =  row (2). L For N=2=2 1 x2 0 (n=1) we have :  (1) =  row (1) = id For N=16=2 2 x2 2 (n=4) we have :  (1) = id  (2) =  row (2)  (3) =  row (2).  col (1).L =  row (2). L  (4) =  row (2).  col (2).L =  row (2).  R row (2). L.. But is not always so simple: shuffle permutation [slide not shown in main presentation]

9. c3c3 c4c4 c2c2 c1c1 c2c2 c1c1 c3c3 c4c4 Example: Multistage Spanned Hypercube …topology mapped on a plane (optical interconnects, VLSI integration) “spanned” hypercube using four bi-permutation modules four-dimensional hypercube connected multiprocessor… {c 2, id} {c 1, id} {c 3, id} {c 4, id} Spanned version of a 4-dimensional weak- interconnected hypercube (16 nodes, 1 bit wide data-bus). It uses four bi-permutation modules, each providing a cube permutation and the identity, which gives a total of 2 4 =16 global permutations for the whole network. Alternatively, using only two of these modules, one can implement an hypercube of dimension 2, with a four bit wide data-bus.

10. Time slotted permutation switching Time slot Permutation appearance period time Red link Blue link Green link Orange link Interconnect 1 Interconnect 2 Interconnect 3 Interconnect N Interconnect 1 Interconnect 2 Interconnect 3 Interconnect N Interconnect 1 Interconnect 2 Interconnect 3 Interconnect N

11. time Burst Interconnects Computation one-stage (ex. 1 ms) Burst interconnection within “short” time slot (Ex. 10Gbps, 100nsec  1kbit) Interconnect 1 Interconnect 2 Interconnection switching interval (Ex. 1ms) ＝ …Slow switching okay

12. Channels are single mode fibers: MFD = 9.5  m Grad diameter 125  m  1  m NA: 0.1  0.01 Module Prototype is not integrated as a single block Experiment Setup using two bi-permutation modules. Output (to CCD) Input (from VCSEL array) Exit first module Input second module {c 2, id} input output {c 2, id} {c 1, id} Displacement stage (piezo)

13. Input (exit VCSEL array) Output first two modules (CCD image) id.id C 1. C 2 id. C 2 C 1. id Preliminary results Inter-module Coupling Efficiency: 1.7dB (no additional optics, matching oil or antireflection coating).  Validation of simple cascaded architecture. …displacement operated manually using a piezo-stage {c 2, id} {c 1, id} Alignment tolerance:  5  m (half peak power). Displacement pitch for commutation: 125  m

14. Conclusion Design and characterization of multi-permutations modules Architectural considerations: Modularity / scalability / reusability of modules and systems Input/output module alignment Micro-lenses, fibers with round ends. Modules built from fiber bundles. Active alignment using electromechanical modules Applications: Transparent time division multiplexed permutation network with relatively slow switching time (ms range) Buffered architecture using bi-permutation modules for packet routing. Simulation results are encouraging and besides control simplicity, an additional advantage is that MEMS actuators could be used in AC mode (at their resonant frequencies). [ Ongoing research ] A C : Multi-function modules: the use of optical fiber modules fits well with the all optical approach; for instance, one can imagine a module with several different interconnection patterns, but also other “optical-functions” like optical delay lines: However, in all-optical networks the “switches” may be very fast (electro optical devices, not MEMS), because the delay time for avoiding the drop of ATM cells is ?? for a typical Gigabit network!!! A C : Multi-function modules: the use of optical fiber modules fits well with the all optical approach; for instance, one can imagine a module with several different interconnection patterns, but also other “optical-functions” like optical delay lines: However, in all-optical networks the “switches” may be very fast (electro optical devices, not MEMS), because the delay time for avoiding the drop of ATM cells is ?? for a typical Gigabit network!!! The switching fabric studied here provides a limited number of long-range, all- optical interconnections useful for high throughput massively interconnected multiprocessors requiring relatively slow switching time (ms range)

15. Electro-optical reconfiguration of the interconnection module. nanosecond range reconfiguration time Interconnection + optical function modules Mixed interconnections, and other optical functions (ex.: delay lines!) Further research directions References: [1] Cassinelli et al., JSAP 2002. [2] Naruse et al., JSAP spring meeting 2002. [3] Goulet et al., OJ2000, pp.247-248. All optical switching (modules with integrated permutations and directional couplers for instance) must be used if switching speed needs to be orders of magnitude higher. Proposed all-optical bi-permutation switch module