Discover how a Bayesian optimization system is utilized for reconfiguring wavelengths in a global network, addressing challenges such as slow reconfiguration and amplifier control. Learn about the complexities involved in optimizing optical backbones to meet the growing demands of online traffic and the significance of accurate gain parameters in amplifier control.
Please find below an Image/Link to download the presentation.
The content on the website is provided AS IS for your information and personal use only. It may not be sold, licensed, or shared on other websites without obtaining consent from the author. If you encounter any issues during the download, it is possible that the publisher has removed the file from their server.
You are allowed to download the files provided on this website for personal or commercial use, subject to the condition that they are used lawfully. All files are the property of their respective owners.
The content on the website is provided AS IS for your information and personal use only. It may not be sold, licensed, or shared on other websites without obtaining consent from the author.
E N D
Presentation Transcript
BOW: BOW: First Real First Real- -World Demonstration of a Bayesian World Demonstration of a Bayesian Optimization System for Wavelength Reconfiguration Optimization System for Wavelength Reconfiguration Zhizhen Zhizhen Zhong Zhong Manya Manya Ghobadi Ghobadi Maximilian Maximilian Balandat Balandat Sanjeevkumar Sanjeevkumar Katti Katti Jonathan Jonathan Leach Leach Mark Mark McKillop McKillop Ying Ying Zhang Zhang Abbas Abbas Kazerouni Kazerouni
Online traffic is growing Online traffic is growing Machine Learning, Internet of Things, Data analytics, COVID-19, Internet traffic during COVID-19 https://www.nytimes.com/interactive/2020/04/ 07/technology/coronavirus-internet-use.html 2
Reconfiguring optical backbones to meet traffic growth Reconfiguring optical backbones to meet traffic growth Global wide-area network (WAN) Reconfiguring wavelengths is challenging. Reconfiguring wavelengths is challenging. 3
Wavelength reconfiguration is a non Wavelength reconfiguration is a non- -trivial effort trivial effort Total time required: 671 seconds (11 minutes) Total time required: 671 seconds (11 minutes) ?3=122 s ?1=179 s ?2 =370 s 24 OSNR (dB) 22 Start 20 End 18 20:28 20:33 20:38 20:43 36 s Experiment setup: State-of-the-art devices at Facebook. Provisioning 7 wavelengths. 152 s 129 s 24 OSNR (dB) 22 Start End 20 Two problems: Reconfiguration is slow. proprietary vendor control software. 18 22:44 22:48 22:52 317 seconds (5 minutes) 317 seconds (5 minutes)
Amplifiers pose the main challenge Amplifiers pose the main challenge The gain parameter is an average gain of all wavelengths. The actual gain and noise depend on: Set of wavelengths and power Device manufacture Router Router Electrical Electrical Optical Optical Fiber Optical Device Optical Device Signal Signal optical power (dBm) optical power (dBm) optical power (dBm) Noise Noise Distance (miles) Wavelength (nm) Wavelength (nm) 5
State State- -of of- -the the- -art amplifier control art amplifier control Amplifier models are proprietary and empirical. Need lots of real-time sample data points to decide the new gain parameters. The process is slow and forces single vendor devices Empirical method (detailed amplifier behavior model) ?????_????= [?1,?2, ,??] ???? Optical Device Optical Device Fiber ?3 ?2 ?1 6
Our solution: Bayesian Our solution: Bayesian- -Optimized Wavelengths (BOW) Optimized Wavelengths (BOW) Bayesian Optimization (BO) uses a probabilistic model to learn wavelength-dependent gain and noise. Unlike prior approaches, BO does not require a priori knowledge about the device. BO is sample efficient and requires less time to converge. Bayesian Optimization ?????_????= [?1,?2, ,??] ???? Optical Device Optical Device Fiber ?3 ?2 ?1 7
Vanilla BO has its problem: possible unsafe parameters Vanilla BO has its problem: possible unsafe parameters BO s sample efficiency comes at the cost of possible unsafe parameters due to its inherent Gaussian stochastic process. BO may generate a set of gains, e.g., all of them are close to lower or upper bounds. This too high or too low amplifier gains may result in signal interruptions. ????? ????? ?? ????? ????? Bayesian Optimization ?????_????= [?1,?2, ,??] ???? Optical Device Optical Device Fiber ?3 ?2 ?1 8
Ensuring practicality with firewall Ensuring practicality with firewall- -based BO based BO We use GNPy, an optical network simulator, as the firewall to pre-evaluate the BO-generated parameters. Only BO-generated parameters that pass the firewall will be deployed. ?????_????= [?1,?2, ,??] ?3 ?2 ?1 Bayesian Optimization ???? ???? No Safe to deploy? Yes ?????_????= [?1,?2, ,??] ???? Optical Device Optical Device Fiber ?3 ?2 ?1 9
Real Real- -world demonstration world demonstration BOW Site #2 FCR/SSH control A production network in Southeast Asia. 3 ROADM sites, 14 amplifier sites, and over 900 km fiber. Simulator Firewall Bayes. Opt. sockets Site #1 Amplifier control ROADM control Site #3 Transponder control Spectrum monitor m n EDFA m n 1 n WSS 1 n WSS WSS WSS ROADM #1 ROADM #2 1 n WSS Transponders Transponders Transponders Transponders ROADM: Reconfigurable Optical Add Drop Multiplexer Transponders 10 Transponders ROADM #3
Experimental results: BOW Experimental results: BOW 141 seconds 88 s 53 s 24 OSNR (dB) 22 Start 20 End 18 19:19 19:23 19:27 Time (hh:mm) BOW is 4.7x faster than state-of-the-art. BOW is open-source, compatible with multiple vendors. 11
Final Remarks Final Remarks Practical wavelength reconfigurable is challenging! BOW (Bayesian-Optimized Wavelengths) achieves 4.7x faster results and is open-source. More details at http://bow.csail.mit.edu Contact: zhizhenz@mit.edu 12
