Application-specific Network Research Articles

Multi-Objective Surrogate-Model-Based Neural Architecture and Physical Design Co-Optimization of Energy Efficient Neural Network Hardware Accelerators

In this paper, we propose a methodology for co-optimizing application specific neural network (NN) accelerators for accuracy and energy expenditure per inference. The architecture of the NN is co-optimized with the concrete ASIC implementation of the accelerator to provide reliable estimates of the energy efficiency. While not constrained to a specific application or NN accelerator architecture, the method is demonstrated on an application specific NN accelerator for the detection of atrial fibrillation in human electrocardiograms that is implemented in 22FDX/FDSOI technology. The NN accelerator is highly parameterizable, i.e., it can map NNs with different architectural properties to a synthesizeable register transfer level representation. The parameter space is further expanded by the parameters of the physical implementation (e.g. memories, clocking, voltage domains). Since the evaluation of accuracy and energy efficiency for a specific parameter combination is computationally expensive, different hyperparameter optimization methods are used and evaluated, including Bayesian Optimization, which tries to find the optimal neural network architecture and physical implementation parameters with a minimum number of training, simulation and evaluation steps.

Priority based dynamic spectrum management using Virtual Utility Functions in Cognitive Radio enabled Internet of Things

Fast growth of wireless devices in IoT (Internet of Things), industry 5.0 and 6G cellular networks, is dealing with the challenge of spectrum scarcity and uneven utilization of available resources. To achieve Ambient Intelligence (AmI) in IoT communication network, dynamic utilization of available network resources, is a challenge faced by resource constraint IoT devices. Cognitive radio technology enables these devices to intelligently sense the free spectrum holes and smartly manage available channels opportunistically. To jointly deal with the challenge of spectrum scarcity and application specific network service provision, we propose a novel multi-attribute based two-tier framework called Cognitive Radio Efficient Spectrum Utilization (CR-ESU) which integrates the concept of Network Function Virtualization (NFV) as Virtual Utility Functions (VUFs) with cognitive radio enabled IoT devices to achieve the dynamic and flexible network services as well as spectral efficiency. The proposed framework is one of the pioneer works that incorporate the two ground breaking technologies, NFV and CRN to create a smart and flexible CR-enabled IoT network. CR-ESU also ensures the best channel selection using Dynamic Channel Reservation (DCR) scheme to decrease the channel switching rate. Continuous Time Markov Chain (CTMC) model is used for network modeling and mathematical expression formation. Promising results show the potential of CR-ESU framework to form a reconfigurable and efficient network.

Towards software defined low maintenance structured peer-to-peer overlays

Structured Peer-to-Peer (P2P) overlay networks are scalable and efficient networks. This efficiency comes at a cost as there is a need for regular maintenance of an overlay structure to handle the continuous churn of peers. This results in huge overhead in terms of maintenance traffic generated by peers. This problem becomes more critical when logical overlay topology differs from actual physical network topology, a common phenomenon in P2P overlay networks. In this paper, we present a novel approach to designing a Distributed Hash Table (DHT) based P2P overlays with the help of Software Defined Networks (SDN) concepts. SDN is an upcoming technology that provides enough flexibility in deciding application-specific network policies and rules dynamically. In our approach, we have attempted to bring the overlay topology closer to actual physical topology and reduce the maintenance traffic as much as possible. We have also tested our work on a hybrid network consisting of both virtual and physical switches. The experimental results suggest and establish the applicability of our approach.

Quality of service provisioning in network function virtualization: a survey

Network function virtualization (NFV) is a promising paradigm for network service provisioning which decouples the network functions from the physical infrastructure, allowing to deploy the virtualized network functions (VNF) as modular software components over commodity servers. In this architecture, service function chaining (SFC) creates an application-specific network overlay which determines the placement and ordering of network functions in a specific service delivery chain. The diverse choices for VNF placement over physical hardware, scheduling of VNF instances to run over the servers and design of optimal chaining mechanisms for traffic steering in the network have direct consequences on the quality-of-service (QoS) perceived by the network users. Previous surveys have focused on categorizing and evaluating various aspects of SFC, NFV and SDN. Due to the tight coupling of QoS provisioning with NFV problems, our work aims primarily at complementing the existing surveys in the sense that we review and classify the state-of-art research efforts in NFV design and implementation from the perspective of QoS provisioning. Our survey provides a closer look at the design space of QoS-related NFV solutions and sheds light on the open research topics in relevance with QoS-based SFC. We also present a case study to evaluate the performance of some recent VNF placement algorithms.

An Introductory Review of Deep Learning for Prediction Models With Big Data.

Deep learning models stand for a new learning paradigm in artificial intelligence (AI) and machine learning. Recent breakthrough results in image analysis and speech recognition have generated a massive interest in this field because also applications in many other domains providing big data seem possible. On a downside, the mathematical and computational methodology underlying deep learning models is very challenging, especially for interdisciplinary scientists. For this reason, we present in this paper an introductory review of deep learning approaches including Deep Feedforward Neural Networks (D-FFNN), Convolutional Neural Networks (CNNs), Deep Belief Networks (DBNs), Autoencoders (AEs), and Long Short-Term Memory (LSTM) networks. These models form the major core architectures of deep learning models currently used and should belong in any data scientist's toolbox. Importantly, those core architectural building blocks can be composed flexibly—in an almost Lego-like manner—to build new application-specific network architectures. Hence, a basic understanding of these network architectures is important to be prepared for future developments in AI.

An ANN-Based Single Calibration Impedance Measurement System for Skin Impedance Range

Impedance measurement has been widely used as an effective indicator for characterizing samples. Traditional high accuracy impedance analyzers are complicated, expensive, and non-portable. Many kinds of research require low cost, portable, and high precision impedance analyzer devices. AD5933 impedance analyzer integrated circuit has been popularly used for fulfilling these requirements. There are many successful applications of the AD5933 integrated circuit; however, the most significant drawback is nonlinear calibration requirements for high precision measurement in a specified range. The calibration and unknown impedances must be close enough to each other for better measurement accuracy. In the literature, calibration impedance arrays increasing the complexity and processing time are commonly used for high accuracy measurements. In this study, an artificial neural network-based signal post-processing algorithm is proposed to overcome the calibration requirements of the AD5933 integrated circuit, which requires different impedances for different ranges. In the literature, a neural network-based solution has not been applied to this phenomenon. An application specific artificial neural network topology is developed and trained for high precision impedance measurement using a fixed calibration impedance. The proposed measurement system is designed for operating in the range of nominal skin impedance. The average mean square error of measurements is obtained as 0.206%. Although a fixed calibration resistance is used, the proposed signal post-processing approach significantly improved the measurement accuracy of the AD5933 integrated circuit. The high accuracy measurement results prove the effectiveness of the proposed measurement system. The developed system offers portable, simple, and low cost high precision impedance analyzer.

Design and analysis of application specific network on chip for reliable custom topology

Network on Chip (NoC) is a communication sub-system formed among the Intellectual Property (IP) cores in System on Chip (SoC). Routers, Links and network interface are the main components of Application SpecificNetworks-On-Chip(ASNoC).As the number of cores increases, the communication demand between the cores also raised insystem on chip. Especially, routers are the main reason for the power consumption in network on chip. In this paper, an Ant Lion Optimization (ALO) based reliable custom topology is proposed in order to reduce the power consumption, area with high speed. Normally, custom topology uses less number of routers than the regular topology. The planned work includes the design of router and ALO based reliable custom topology. Xilinx ISE design suite 14.5 is used for the design and validation of proposed method. From the implementation result, proposed topology is operated with low power and less area and high speed. The performance of flip flops, LUTs, Slices and frequency are shown in result section. When compared with other existing techniques our proposed method gives better outcomes.

Ant Lion Optimized Bufferless Routing in the Design of Low Power Application Specific Network on Chip

Network on chip is widely restricted with power utilization and area occupation due to the usage of buffers. Hence, the design of bufferless architecture entirely eliminates such kind of limitations. However, conventional methodologies will not provide low-power design with enhanced features by means of operational frequency and area. This work introduces an optimization algorithm along with bufferless routing in chip design. Ant lion optimized (ALO) routing topology in bufferless router achieves very less power. Power dissipation of ALO-bufferless technique is evaluated with conventional topologies, such as spin, octagon and cliche. Xilinx ISE design suite 14.5 is used for the purpose of design and validation of the planned work, and it is compared with fault-tolerant deflection routing and hierarchical FTDR in terms of throughput and fault rate. ALO-based bufferless routing achieves operational frequency of 780.153 MHz with 0.413 mW power consumption; while ant lion optimized buffered routing achieves operational frequency of 426.995 MHz and 0.750 mW for speed and power, respectively.

Note: Development of sTGC strip front-end readout prototype for ATLAS new small wheel upgrade.

This note presents the development of front-end electronics prototype for A Toroidal LHC ApparatuS (ATLAS) New Small Wheel (NSW) detector small-strip Thin Gap Chamber (sTGC) to be used for level-1 triggering and precision muon tracking at the Large Hadron Collider. The sTGC front-end electronics must deal with a large number of strip channels with very limited space constraint, large dynamic range of charge readout with low noise, simultaneous fast trigger, and precision readout data movement. The prototype uses custom radiation-tolerant application specific integrated circuits, optimized layout, power, and ESD protection networks to realize the above stringent requirements. The electronic test indicates that the front-end channel has a dynamic range of 900 pC and inherent noise of 0.19 fC. The sTGC cosmic ray test verified the precision charge measurement of the prototype, reconstructed the muon track, and measured the performance of sTGC. The NSW trigger chain test verified the trigger processing and 4.8 Gbps transmission of the prototype.

Toward In-Network Deep Machine Learning for Identifying Mobile Applications and Enabling Application Specific Network Slicing

Toward In-Network Deep Machine Learning for Identifying Mobile Applications and Enabling Application Specific Network Slicing

Software based Method to Specify the Extreme Learning Machine Network Architecture Targeting Hardware Platforms

Extreme learning machine (ELM) is a biologically inspired feed-forward machine learning algorithm that offers a significant training speed. Typically, ELM is used in classification applications, where achieving highly accurate results depend on raising the number of ELM hidden layer neurons, which are randomly weighted independently of the training data and the environment. To this end, determining the rational number of hidden layer neurons in the extreme learning machine (ELM) is an approach that can be adapted to maintain the balance between the classification accuracy and the overall physical network resources. This paper proposes a software based method that uses gradient descent algorithm to determine the rational number of hidden neurons to realize an application specific ELM network in hardware. The proposed method was validated with MNIST standard database of handwritten digits and human faces database (LFW). Classification accuracy of 93.4% has been achieved using MNIST and 90.86% for LFW database. General Terms Neural Networks, Classification Applications

A Systematic Review of Shared Sensor Networks

While Wireless Sensor Networks (WSNs) have been traditionally tasked with single applications, in recent years we have witnessed the emergence of Shared Sensor Networks (SSNs) as integrated cyber-physical system infrastructures for a multitude of applications. Instead of assuming an application-specific network design, SSNs allow the underlying infrastructure to be shared among multiple applications that can potentially belong to different users. On one hand, a potential benefit of such a design approach is to increase the utilization of sensing and communication resources, whenever the underlying network infrastructure covers the same geographic area and the sensor nodes monitor the same physical variables of common interest for different applications. On the other hand, compared with the existing application-specific design, the SSNs approach poses several research challenges with regard to different aspects of WSNs. In this article, we present a systematic literature survey on SSNs. The main goal of the article is to provide the reader with the opportunity to understand what has been done and what remains as open issues in this field, as well as which are the pivotal factors of this evolutionary design and how this kind of design can be exploited by a wide range of WSN applications.

Communication Centric Floorplanning of NoC Based System on Chip

In the past, shared bus based architecture was used as a communication architecture in SoC. They consume more area, power and do not meet the proper bandwidth requirement. Network on chip (NoC) is evolving as a viable communication architecture because they offer better scalability, modularity, design predictability, lower power consumption and shorter latency compared to bus based systems. Application Specific Network on chip (ASNoC) topologies are found to be superior than regular NoC topologies for designing SoC with known communication demands. In this paper, a communication centric floorplanning algorithm is proposed in which communication architecture is synthesized along with the floor plan. This is achieved by grouping IP cores based on their communication requirements in pre floorplanning stage and placing network components judiciously in the post floorplanning stage. Simulated annealing is used as a search engine to obtain the optimal location of IP cores and network components in the floorplan.

An agent-based framework for production software defined networks

Software defined networking SDN is lauded to be the paradigm of choice for the next generation networks. While academia explores use cases in various contexts, industry has focused on data centre networks' limited but intense needs. There is a significant range of complex and application-specific network services that can potentially benefit from SDN, but introduction and adoption of such solutions remains slow in production networks. One impeding factor is the lack of a simple yet expressive enough framework applicable to all SDN services across production network domains. Without a uniform framework, SDN developers create disjoint solutions, resulting in untenable management and maintenance overhead. In this paper, we propose an agent-based framework, which enabled the development of three distinct SDN services addressing the needs of network operators, application providers, and application users. The architecture facilitates application-oriented SDN design with an abstraction composed of software agents on top of the underlying network.

A Comparative Study and Analysis of Data Aggregation Techniques in WSN

Background: Wireless Sensor Network (WSN) often consists of large number sensor nodes which co-operatively send sensed data to base station. These networks are used in various applications scenarios such as habitat monitoring, disaster management, security and military, etc. As the sensor nodes are generally operated by battery driven, an efficient utilization of power is essential. Therefore, to enhance the life time of sensor network, power efficient methods has to be adapted to gather and aggregate data. It's mainly because majority of the sensor node energy is consumed during transmission than computation. Data aggregation is one of the widely recognized solutions to gather and reduce the redundant number of packets to be sent to base station by aggregating them. Literature survey shows that several application specific sensor network data gathering protocols exists in enhancing the network lifetime. Methods: This paper provides a comprehensive survey of different data aggregation algorithms in wireless sensor networks. The different algorithm are compared and contrasted based on the performance measures such as network type, mobility, Data (D) or Node (N) centric protocol, computation characteristics, data delivery model, communication pattern, query based support, energy usage and application type. To evaluate the performance of some of the algorithms surveyed in this work, simulation experiments on Matlab were conducted. Results: It was observed from the results that that the network topology has a high impact on the routing protocols to perform data aggregation. There exist few other protocols using contemporary methodologies such as Tree, Grid and Hybrid approaches which are also discussed in this paper. Application: The survey will help to identify the data aggregation algorithm suitable for a specific application.

OPAIC: An optimization technique to improve energy consumption and performance in application specific network on chips

Network on Chip (NoC) is an appropriate and scalable solution for today’s System on Chips (SoCs) with the high communication demands. Application specific NoCs is preferable since they can be customized to optimize all requirements of the specific applications. This paper presents an OPtimization technique for Application specifIC NoCs (OPAIC), which aims not only to decrease the energy consumption but also to improve the area of NoCs. OPAIC is composed of three stages to find the optimum NoC; in the first stage, it uses a linearized form of a Quadratic Assignment Problem (QAP) to map tasks on cores to minimize the energy. In the second stage, a Mixed Integer Linear Problem (MILP) is proposed to find the optimum number of the routers for the layout earned in previous stage. Finally, a Greedy Algorithm is applied to optimize the number of virtual channel for every link based on its traffic needs.

Data De-Duplication Aware Clustering for Optimized Storage in Wireless Multimedia Sensor Networks

The progressive growth in the scope and scale of wireless sensor networks have explored new dimensions of potential research in this domain. Traditionally they are viewed as resource constrained and application specific localized adhoc networks. The capabilities of the sensor node have increased many fold with the advent of new sensors, miniaturization of electronic devices and storage facilities. Now a days the sensor nodes are equipped with on chip CMOS cameras, Flash memory and considerably large storage space and power resources. Many changes in communication protocols have led to the possible integration of heterogeneous sensor networks into the existing infrastructure based networks. With all these capabilities sensor nodes are capable to transmit multimedia data such as video, audio, image etc. seamlessly over the large scale networks leading to large scale Wireless Multimedia Sensor Networks (WMSNS). For optimizing the data transmission and storage of the generated huge multimedia data we have proposed De-Duplication aware clustering in WMSNS.

Internet factories: Creating application-specific networks on-demand

We introduce the concept of Internet factories. Internet factories structure the task of creating and managing application-specific overlay networks using infrastructure-as-a-service clouds. We describe the Internet factory architecture and report on a proof of concept with three examples that progressively illustrate its working. In one of these examples, we demonstrate the creation of a 163-node IPv6 network over 18 cloud locations around the world. Internet factories include the use of libraries that capture years of experience and knowledge in network and systems engineering. Consequently, Internet factories solve the problem of creating and managing on-demand application-specific overlay networks without exposing all their intricacies to the application developer.

Scalable architecture for ubiquitous healthcare using sensor cloud platform

In this paper we propose a hybrid computational paradigm associated with cloud computing and wireless sensor network. Attempts have been made to devise a protocol that establishes communication between sensors and cloud resources. The objective is to increase the availability of real time sensor information across the users. Further it will be possible to share the information of various application specific sensor network deployed at various locations. We suggest a virtualisation for sensor vs. cloud and cloud vs. users. This will facilitate the user to visualise sensor network and cloud.

Performance improvement of application-specific network on chip using machine learning algorithms

This paper presents a novel and efficient mapping algorithm based on machine learning methods. It produces the best mappings with different metrics which were totally evaluated by support vector machine SVM, decision tree classifier DTC, and multiple clustering. It helps to find the optimal application-specific network-on-chip NoC based on user's demands on how to customise and prioritise the impact of three key metrics on special mapping. The parameters are robustness index, contention factor and communication cost. In fact, as mapping generator produces a mapping, these parameters will be calculated and compared with some rules. The rules are extracted by SVM and DTC. They highly affect the fitness function of the genetic algorithm GA. So the algorithm can be controlled. Simulation results show that the proposed algorithm achieves more than fifteen times performance improvement and more supervision in finding the best mappings from numerous generated solutions in comparison with the ordinary algorithm.