Photonic Interconnect based Switching and Logic Synthesis for High Performance Computing Applications

Abstract

The widespread web-based applications are attracting extensive acceptance and deployment due to easy accessibility. The voluntary usage of social networking platforms by millions of users has forced various organizations to install extensively huge data centers (DCs). The large-scale industrial fields like health, engineering, commercial, defense, etc. employing High-Performance Computation (HPC) networks depend on these data centers. The “big data” flowing every second requires minimum possible error and faster processing to ensure high quality of service. The major issue of this multimedia transmission arises when the user-traffic is at its peak and DCs face a load-balancing problem and higher latency while switching this traffic. This questions the present DC configuration and its efficiency. HPC systems handling complex tasks along with artificial intelligence, on the other hand, demand highly scalable structures with low latency. The current hybrid electro-optic technology entails large power consumption during electro-optic (E-O) and optic-electro (O-E) conversions. In addition, if the switching and processing of big data are continued in the electronics domain then its full potential cannot be exploited due to the increasing complexity of cables and it will eventually limit the scalability of DCs. The all-optic platforms can overrun the limitations of these challenges by offering many other positive advantages like a large bit rate, massive bandwidth, low reception-error, and reliable cost. The attractive features of optical switching technology in DCs offer promising solutions to various networking problems. Although the extensive deployment of various optical switching technologies has proposed countless innovative schemes to overcome the challenges but the designing demands have always been an obstruction in DC switch performance. The high-radix interconnected structure utilizing MEMS-based technologies faces a large response time whereas liquid crystal-based configurations offer low power consumption but scalability remined an issue. Thermo-optic switches require large operating power and hence have a lower preference for DC applications. It is observed that SOA based technologies have the ability to overcome their internal non-linear issues and offer solutions to the above limitations. These structures don’t need any large coolants and offer low-power operation. In addition, the non-linear properties of SOA when exploited in Mach-Zehnder Interferometric (MZI) configuration can achieve many self-switching capabilities with the iii inevitable speed that adds to the superiority of the structure. The employment of MZI configurations suffers various limitations like poor crosstalk and low extinction ratio but with appropriate usage of its properties, its applicability in several fields can picture future switching structures with guaranteed scalability upgradations with time. This thesis facilitates a promising photonic switching technique that can balance the incoming traffic with definite speed and offers good service quality for high performance computing applications. Initially, the best suitable combination for photonic switching has been explored which is superior to previously reported techniques. The cross-phase modulation property (XPM) in Traveling Wave SOA i.e. TWSOA has been tested with Symmetric MZI i.e. SMZI to obtain the crossbar operation at the elementary level and power distribution in terms of gain has been presented. With the use of this elementary block, the TWSOA-SMZI photonic switch is proposed by increasing the input and output port-size to 4×4. The performance of this 4×4 crossbar switch is measured for various parameters like BER, latency, OSNR, ER, output power, eye-opening, and eye-height for ensuring scalability. In the next step, the high-radix configuration has been investigated by increasing the port-size further to 8×8. The scalability of the structure is monitored with the same parameters as 2×2 and 4×4. The performance of a three-stage 8×8 switch is examined by exploring the impact of varying injection currents of SOA. The other results are presented by decreasing the port spacing in THz to ensure the further possibility of scalability in the proposed configuration. The investigations explain the relation of carrier density with TWSOA gain w.r.t. injection current. The large values of injection current can vary the non-linearities occurring in the transmission route which further opens doors to explore the in-path non-linearities. The optical non-linear phenomenon like noise and crosstalk can worsen the performance of switching arrangements. They are able to modify the incoming input bits traveling through the channel by affecting the refractive index of the medium. The crosstalk is considered at its peak when a crossbar switch is operated in the cross-state. The switch parametric analysis of the proposed structure is examined parallelly with non-linearities of SOA that could affect the final performance at the destination. This thesis presents analytical modelling of non-linear crosstalk and proposes a solution to minimize its effects on the output signal. The relation of the occurrence of logic ’1’ at different places in the input sequence is also observed. The high logic appearance at alternate and consecutive places affects the output due to the MZI phase shifts occurring above the iv threshold. It further explains how the biasing current of TWSOA can control these phase shifts and hence the ER impairments. The other results explain the changing output effects with the length of input sequences. The structure is explored further on TWSOA’s birefringent property that certifies its polarization dependent output and further proposes a solution to remove the polarization-sensitivity by adding another type of amplifier in the middle stages. It further explains the benefits of adding EDFA-SMZI loops in the second stage and further upgrades the proposed structure to S-E-S hybrid compound amplification-based 8×8 photonic switching. This structure studies the fundamental components of polarization ellipse in terms of Stokes parameters and the Poincaré sphere. It compares the polarization sensitivity of basic SOA-MZI loop and EDFA-based hybrid structure for achieving ultrafast response. Switching the signal to an output port and blindly sending it on a particular I/O path that is not available at the moment will only increase or add to the switching time and hence negatively contributes to the latency. Hence, the proposed S-E-S switching configuration is further explored on an innovative level of smart switching for DC applications. This level includes the incorporation of neural networks parallelly with an optical switch. Firstly, the basic introduction to the working of Artificial Neural Networks (ANN) and their various types is presented. This investigation is further explored to discover the best suitable ANN for predicting the switching path. This brings in the study of the Recurrent Neural Network (RNN) based Long-Short-Term Memory (LSTM) neural network and its activation functions. Secondly, the implementation of the LSTM network for predicting the best suited port-destination path in terms of shortest distance and minimum error is done. This process achieves a prediction accuracy of 97% and the classification of minimum and maximum error paths is done with a classification accuracy of 99.5%. The LSTM network performance is further explored on the basis of the confusion matrix and the final improved results of the switching configuration are acquired. This investigation leaves with the exploration of the availability of port that is decided by the LSTM network based on error-rate. Hence, further upgradation is done by feeding both the current port power and error-rate at the training of the LSTM network so that it predicts and classifies the data on the basis of availability and minimum error-based shortest path. It first samples the data v for path availability by determining the average port power and calculates the minimum distance port-path combinations and then classifies those paths with BER. This solves the problem of finding an available port for switching traffic data. The training of the LSTM network is measured with batch loss and RMSE and testing parameters give a classification accuracy of 99.7%. The improved switching parametric results are obtained with a maximum throughput of 96 Tbps. Hence, it proves a better solution to switching challenges for HPC systems in DCs. The three-stage hierarchies of DC networks also face the issue of scaling the required information and forwarding it upper level based on priority. The investigations have also been done in Control Plane Interface (CPI) routing for achieving a smart solution for scheduling the incoming traffic. The logic gating-based smart routing technique is presented that utilizes the information from the header of the incoming node requests. The priority bit in the header tells whether the immediate transfer of the message is required or it could be put in wait-status for another important signal to pass through. The proposed AND-OR logic gating-based routing allows the appropriate scheduling and controlling of incoming traffic and it further proves the excellency in terms of insertion loss and contrast ratio. The ultrafast scalable compound hybrid amplification-based photonic smart switching structure utilizing TWSOA-EDFA-TWSOA in SMZI arrangement has been implemented in parallel incorporation of the deep learning LSTM model that provides a solution to various challenges of existing switching techniques in Data Centers. The logic-routing scheme for the Control plane Interface is also established for removing the centralized control problems.