Offline Optimization of Wavelength Allocation and Laser to Deal with Energy-Performance Tradeoffs in Nanophotonic Interconnects

Transcription

1 Offline Optimization of Wavelength Allocation and Laser to Deal with Energy-Performance Tradeoffs in Nanophotonic Interconnects Cédric KILLIAN Universty of Rennes 1 Cairn Team IRISA, Inria OPTICS workshop, in conjunction with DATE 2018 March 23 rd 2018, Dresden Germany Jiating Luo, Cedric Killian, Sebastien Le Beux, Daniel Chillet, Olivier Sentieys, and Ian O Connor Offline optimization of wavelength allocation and laser power in nanophotonic interconnects. ACM J. Emerg.Technol. (January 2018). DOI: / Institut de Recherche en Informatique et Systèmes Aléatoires

2 Context Basic concepts of photonic Network-on-Chip l l n Source: [3] C. Sciancalepore, et al. [4] W. Bogaerts et al., 2013 P crossing l n l = ln P drop [5] Z. Wang et al [6] L. Vivien et al λ 0 λ 0 p p 2 Silicon waveguide: Handle multiple optical signal (wavelength) transmissions On-Chip Laser: Transform electric signal to optical signal Ring Micro-resonator: If on, move one optical signal from one waveguide to another If off, do nothing Wevelenght dependent PhotoDetector: Convert optical signal to electric signal On-chip laser with direct modulation for λ x Add-Drop filter for λ x Photodetector OFF ON OFF ON λ x λ x 4

3 Context Global view of a multicore based on ONoC λ 0 λ 0 p p 2 5

4 Context Global view of a multicore based on ONoC λ 0 λ λ 0 1 λ 0 01 λ 0 p 1 1 p 4 p 2 01 λ 0 λ0 λ p 3 6

5 Context Adapt the communication performances Wavelenght Data Modulation (WDM) Communication resources are shared Wavelength allocations impact application performances λ 0 λ 1 λ 2 λ 3 p 0 p 1 p 2 p 3 No wavelength available between P 0 and P 3 0 t 1 t 2 t 3 t 4 t 5 t 6 t 7 t 8 t On-chip laser with direct modulation for λ x Add-Drop filter for λ x Photodetector OFF ON OFF ON λ x λ x [7] J. Luo et al,

6 Context Dynamic Energy losses in ONoC Illustration of crosstalk a b p i p i 100 FSR 100 FSR Transmission on drop port (%) Transmission on drop port (%) 0 λ 0 +FSR λ signal (nm) 0 λ 0 +FSR λ signal (nm) Pin SNR < SNR Pin Received optical 0Power λ 0 +FSR λ signal (nm) Wavelength Allocations impact on the SNR signal Xtalk By targeting a given Bit Error Rate, a degradation of SNR results as an increase on Laser Power è increase Energy of communication. More Wavelengths in parallel for a communication increase performance, but increase energy. è How to provide a tradeoff between energy and performance? ç Received optical 0Power λ 0 +FSR λ signal (nm) 10

7 Context Accuracy for laser power tuning impacts Energy N P_lvl = Emitted optical N P_lvl =2 Power (uw) N P_lvl = i bias i 1 i 2 i 3 i 4 I laser (ma) t P out p 0 0 λ 0 λ 1 λ 2 λ signal (nm) t t More power levels è fine tuning of power Laser è Lower energy consumption But increase laser driver complexity. In fact, how many power level are really required? 13

8 Outline 1. Context, key elements of ONoCs, and problematics 2. Our vision on Energy Vs Performance tradeoffs 3. How to generate a set of wavelenght allocations 4. Case Study and trends 5. Conclusions 14

9 Our vision: mapping of communications depends on needs c 0à1 t 0 t 1 t 2 c 0à2 c 1à2 On-chip laser with direct modulation for λ x Low power High power OFF (β pj/bit) (α pj/bit) Add-Drop filter for λ x OFF λ x ON λ x Photodetector Energy d a c b Execution time a λ 0 λ 1 λ 2 λ 3 p 0 p 1 p 2 b λ 0 λ 1 λ 2 λ 3 p 0 p 1 p 2 c λ 0 λ 1 λ 2 λ 3 p 0 p 1 p 2 d λ 0 λ 1 λ 2 λ 3 λ 0 λ 1 λ 2 λ 3 p 0 p 1 p 2 [7] J. Luo et al,

10 Outline 1. Context, key elements of ONoCs, and problematics 2. Our vision on Energy Vs Performance tradeoffs 3. How to generate a set of wavelenght allocations 4. Case Study and trends 5. Conclusions 17

11 How to to generate a set of wavelenght allocations Application Task graph Architecture [8] Le Beux et Al System specification ONoC interface Mapping p 0 t 0 TSVs t 0 t 1 t 2 IP core p 3 Configuration Manager p 2 -Task execution time -Data size -BER requirements Optical layer: ONoC Electrical layer : cores -nb of wavelengths -nb of waveguides -nb of interfaces -distance between interfaces t 3 p 2 t 1 Technological parameters λ 0 Energy number of power level (N P_lvl ) maximum power data rate loss sensitivity Transmission Δ λ Solutions Cf g 4 Cf g 3 Cf g 2 Execution time population Individual (example) Offline multi-objective Optimization t 0 t x t 0 Simulation t 0 task t 1 Selection Mutations Crossovers Simulation Simulation t t 2 1 t 1 t 2 communication t 2 remove individual Constraints verification Estimation: -execution time -energy [7] J. Luo et al,

12 Outline 1. Context, key elements of ONoCs, and problematics 2. Our vision on Energy Vs Performance tradeoffs 3. How to generate a set of wavelenght allocations 4. Case Study and trends 5. Conclusions 19

13 Case Study and trends Simple generic application Application Task graph Mapping t 0 t 1 ONoC interface t 0 5kcc C 0à3 6kb t 3 5kcc C 2à3 6kb C 3à5 8kb Op0cal layer: ONoC t 4 t 5 t 3 t 2 TSVs IP core C 1à2 C 2à4 C 4à5 t 1 t 2 t 4 t 4kb 5 8kb 4kb Electrical layer : cores 5kcc 5kcc 5kcc 5kcc Parameter Value Waveguide propagation loss db/cm Photodetector sensitivity -20 dbm Laser efficiency 15% λ FSR 3dB MR bandwidth 0.4 nm 8 nm 0.26 nm [7] J. Luo et al,

14 Case Study and trends Simple generic application Impact on number of wavelengths Increase number of wavelengths à may improve perf but increase energy consumption Targeted BER:

15 Case Study and trends Simple generic application Impact on number of power levels of laser Increase power levels à decrease energy consumption Targeted BER: and 8 wavelenghts, NP Lvl =3 è level power [1,2.5,4] mw NP Lvl =5 è level power [1,1.75,2.5,3.25,4] mw NP Lvl =7 è level power [1,1.5,2,2.5,3,3.5,4] mw 22

16 Case Study and trends Simple generic application Energy and performance trade-off capabilities Energy High perf d c b Execution time Low power 44% energy reduction 71 % execution time reduction èoffers possible tuning of 44% energy reduction or 71% execution time reductionç 26

17 Bandwidth distribution Power distribution Bandwidth distribution Power distribution Case Study and trends Power and wavelengths distribution Np lvl = 8 and Nb λ = 8 Energy High perf d c b Execution time Low power Low power 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% Bandwidth allocation 1 λ TG1 TG2 TG3 TG4 TG5 TG6 TG7 TG8 Graph ID 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% Power distribution 2 mw 4 mw 6 mw 8 mw 10 mw TG1 TG2 TG3 TG4 TG5 TG6 TG7 TG8 Graph ID High performance 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 1 λ 2 λ 3λ 4λ 5λ 6 λ 7λ 8 λ TG1 TG2 TG3 TG4 TG5 TG6 TG7 TG8 Graph ID 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 2 mw 4 mw 6 mw 8 mw 10 mw TG1 TG2 TG3 TG4 TG5 TG6 TG7 TG8 Graph ID 27

18 Outline 1. Context, key elements of ONoCs, and problematics 2. Our vision on Energy Vs Performance tradeoffs 3. How to generate a set of wavelenght allocations 4. Case Study and trends 5. Conclusions 28

19 Conclusions Introduction of key concepts: Performance on ONoC with WDM Energy problematics with Crosstalks and Laser power levels Be careful on allocation when performing WDM! Directly impact performance and energy We proposed a framework to provide a set of wavelength allocations, each providing an energy-performance tradeoff Multi-parameters as input (technological, applications, optimization algorithm, ) Offers possible tuning of 44% energy reduction or 71% execution time reduction 29

20 Thanks to Collaborators: Daniel CHILLET, Olivier SENTIEYS, Jiating LUO, Van-dung PHAM Sebastien LE BEUX, Ian O CONNOR Financers: 30

21 References [1] Semiconductor Industry Association, ITRS, Online, URL: [2] : A. Karkar et al. A Survey of Emerging Interconnects for On-Chip Efficient Multicast and Broadcast in Many-Cores. In:IEEE Circuits and Systems Magazine. [3] C. Sciancalepore et al., "Thermal, Modal, and Polarization Features of Double Photonic Crystal Vertical-Cavity Surface-Emitting Lasers," in IEEE Photonics Journal, vol. 4, no. 2, pp , April [4] W. Bogaerts et al. Technologies and Building Blocks for On-Chip Optical Interconnects,books [5] Zhao Wang, Dixon Paez, Ahmed I. Abd El-Rahman, Peng Wang, Liam Dow, John C. Cartledge, and Andrew P. Knights, "Resonance control of a silicon micro-ring resonator modulator under highspeed operation using the intrinsic defect-mediated photocurrent," Opt. Express 25, (2017) [6] Laurent Vivien, Johann Osmond, Jean-Marc Fédéli, Delphine Marris-Morini, Paul Crozat, Jean- François Damlencourt, Eric Cassan, Y. Lecunff, and Suzanne Laval, "42 GHz p.i.n Germanium photodetector integrated in a silicon-on-insulator waveguide," Opt. Express 17, (2009) [7] Jiating Luo, Cedric Killian, Sebastien Le Beux, Daniel Chillet, Olivier Sentieys, and Ian O Connor Online optimization of wavelength allocation and laser power in nanophotonic interconnects. ACM J. Emerg.Technol. Comput. Syst.1, 1, Article 1 (January 2018),19pages. DOI: / [8] S. Le Beux et al. Chameleon: Channel efficient Optical Network-on-Chip. In:2014 Design, Automation Test in Europe Conference Exhibition (DATE). Mar [9] C. Killian et al. Energy and Performance Trade-off in Nanophotonic Interconnects using Coding Techniques. In: DAC [6] J. Luo et al,

22 Thanks! Questions? Institut de Recherche en Informatique et Systèmes Aléatoires