Performance Analysis of Convolution Algorithms for Deep Learning on Edge Processors

Because of the solid performance of providing quality of service for various applications for decades, bandwidth reservation has been increasingly used in recent years for large amounts of data transfer to achieve guaranteed performance. However, effective scheduling strategy to achieve the trade-off between data transfer cost and data transfer performance still remains to be investigated. In this paper, we focus on the trade-off between cost and the most common performance parameter, i.e., completion time, of data transfers using bandwidth reservation in dedicated networks. We consider the scheduling of two types of bandwidth reservation requests regarding such trade-off: (1) to achieve the minimum data transfer cost given the data transfer deadline, and (2) to achieve the earliest data transfer completion time given the maximum data transfer cost. We propose two bandwidth reservation algorithms with rigorous optimality proofs to optimize the scheduling of these two types of bandwidth reservation requests. We then compare the proposed algorithms with two scheduling algorithms originating from one widely used scheduling algorithm in production networks, and the efficacy of the proposed optimal algorithms is verified through extensive simulations.

Cloud computing provides distributed computing resources such as servers, storage and applications to the end users through data centers. The data centers are geographically located at different locations. The client applications or requests being serviced by cloud service providers on “pay per use”. So, different pricing models are adapted to compute the cost and revenue of the data centers. The cost of VM is computed depending on its placement in the data centers through broker policies. So, the broker policies have a significant role in evaluating the cost of the VM which directly impacts on revenue of the service provider. In the computing competition, the VM cost should be minimized by which service demand will be maximized. Moreover, the response time and processing time of the data centers need to be minimized to posses commendatory Quality of Service. In this work, we propose a new service broker policy to minimize the total cost. The total cost considers the VM cost and the data transfer cost. The proposed mechanism also reduces the response time and data processing time of the data centers. The proposed policy is simulated in Cloud Analyst. We examine the performance of the proposed mechanism with ten different scenarios. Finally, we compare performance results with respect to VM cost, data transfer cost, total cost, processing time and response time of data centers with the existing policies and observe better than these.

The cost of data transfers, and in particular of I/O operations, is a growing problem in parallel computing. This performance bottleneck is especially severe for data-intensive applications such as multimedia information systems, databases, and Grand Challenge problems. A promising approach to alleviating this bottleneck is to schedule parallel I/O operations explicitly.

General Purpose Graphics Processing Units (GPGPU) have been used in high performance computing platforms to accelerate the performance of scientific applications such as simulations. With the increased computing resources required for large-scale network simulation, one GPU device may not have enough memory and computation capacities. It is therefore necessary to enhance the system scalability by introducing multiple GPU devices. It is also attractive to investigate the performance scalability of Multi-GPU simulations. This paper describes the simulation of information propagation on multiple GPU devices, including the optimized network simulation algorithms, the network partitioning and replication strategy, and the data synchronization scheme. The experimental results for scalable random networks show that the number of simulation steps, computation time, synchronization time, and data transfer time all affect the overall simulation performance. In order to compare with random networks, we also conduct simulations of scale-free networks. We can observe that the node replication ratio in scale-free networks is smaller than that in random networks and therefore the cost of data transfer and synchronization is significantly reduced. This indicates that the network structure is also an important factor that influences the simulation performance in a Multi-GPU system.

The study aims to ensure the security and authentication efficiency of the bill image, and the encryption and decryption methods and security protection of the electronic bill are studied in the experiment. First, aiming at the not high traditional electronic bill security performance, a method is proposed, namely, embedding a watermark into a binary image with edge information. Second, aiming at the weak compression robust character of electronic bill image, the method of chaotic encryption of digital watermark through wavelet coefficient matrix algorithm is proposed to be combined with the binary sequence. Finally, the deep learning algorithm combined with the convolution algorithm can detect the quality of electronic bill watermark images. The results show that the method of embedding watermark with edge information effectively has improved the confidentiality of electronic bills. The method of chaotic encryption of digital watermarks by wavelet coefficient matrix algorithm combined with binary sequence has improved the anti-compression ability of digital watermarks. The multi-watermark encryption method has enhanced the tamper-proof ability of electronic bills and has improved the security performance of bills, and the deep convolution algorithm has improved the security and efficiency of electronic bill processing.

Quality prediction of seabream Sparus aurata by deep learning algorithms and explainable artificial intelligence.

Advancements in digital technology have driven the rise of biometric security systems, notably in the field of finger vein detection. In most of the research on finger vein classification in the literature, achieving high accuracy is the main aim, while aspects such as generalization capacity and test distribution are mostly overlooked. In this study, two different datasets (MMCBNU_6000 and FV-USM) were tested with different test distributions, using a K-Fold structure for unbiased sampling in classification. In experiment part, two distinct image enhancement methods, namely Contrast Limited Adaptive Histogram Equalization (CLAHE) and Sobel filtering, were utilized on the datasets, and Convolutional Neural Networks (CNN) were used for feature extraction. Furthermore, machine learning algorithms were applied for classification, forming a Hybrid Convolutional Machine Learning algorithm. In this method, the model, which is fed through two different channels compared to conventional learning algorithms, combines classical machine learning classifiers with the CNN model. In the scope of this study, three tasks were outlined. The first two focused on implementing various machine learning algorithms for each dataset, while the third involved merging datasets and employing the Stacking Ensemble Classifier (SEC). For evaluating the models, accuracy and F1-score metrics were used. The results indicate that the highest accuracy rate was achieved in the third experiment, with a score of 98.94%. Additionally, it is also observed that increasing the amount of test data (the difference between 20% Test and 50% Test) has a minimal effect in reducing the model's accuracy metric compared to previous studies.

In this work, we present an innovative approach for damage detection of infrastructures on-edge devices, exploiting a brain-inspired algorithm. The proposed solution exploits recurrent spiking neural networks (LSNNs), which are emerging for their theoretical energy efficiency and compactness, to recognise damage conditions by processing data from low-cost accelerometers (MEMS) directly on the sensor node. We focus on designing an efficient coding of MEMS data to optimise SNN execution on a low-power microcontroller. We characterised and profiled LSNN performance and energy consumption on a hardware prototype sensor node equipped with an STM32 embedded microcontroller and a digital MEMS accelerometer. We used a hardware-in-the-loop environment with virtual sensors generating data on an SPI interface connected to the physical microcontroller to evaluate the system with a data stream from a real viaduct. We exploited this environment also to study the impact of different on-sensor encoding techniques, mimicking a bio-inspired sensor able to generate events instead of accelerations. Obtained results show that the proposed optimised embedded LSNN (eLSNN), when using a spike-based input encoding technique, achieves 54% lower execution time with respect to a naive LSNN algorithm implementation present in the state-of-the-art. The optimised eLSNN requires around 47 kCycles, which is comparable with the data transfer cost from the SPI interface. However, the spike-based encoding technique requires considerably larger input vectors to get the same classification accuracy, resulting in a longer pre-processing and sensor access time. Overall the event-based encoding techniques leads to a longer execution time (1.49×) but similar energy consumption. Moving this coding on the sensor can remove this limitation leading to an overall more energy-efficient monitoring system.

Instructors who wish to disseminate digital content to learners in an economical manner must consider the costs of (1) content creation (software and hardware), (2) content dissemination (data transfer costs), and (3) content access (hardware and software needed to view content). In this paper, we present a novel technique – photocasting – to greatly reduce these three costs in a one-to-one learning environment. Our technique has minimal requirements: a low-end smartphone for instructors and learners, and a modest additional effort by instructors after recording their digital notes. The delivery cost is primarily determined by the size of the audio, and is approximately 100 times smaller than the size of a traditional video of comparable quality. We have implemented our technique as a prototype for the Android platform, and we report initial results of a small-scale user study with three participants: a school teacher, a private tutor, and a high-school student.

With the continuous increase of cloud storage adopters, data deduplication has become a necessity for cloud providers. By storing a unique copy of duplicate data, cloud providers greatly reduce their storage and data transfer costs. Unfortunately, deduplication introduces a number of new security challenges. We propose PerfectDedup, a novel scheme for secure data deduplication, which takes into account the popularity of the data segments and leverages the properties of Perfect Hashing in order to assure block-level deduplication and data confidentiality at the same time. We show that the client-side overhead is minimal and the main computational load is outsourced to the cloud storage provider.

Geographical knowledge cloud service is a typical online service that provides big spatial data analysis with the function of knowledge discovery or decision-making.The composition of geographical knowledge cloud service imposes stricter requirements for better overall QoS and execution efficiency of the service chain.In this paper, we present a data volume aware ant colony optimization approach called DVA-MOACO algorithm for geographical knowledge cloud service composition.Our algorithm utilizes a multi-index service quality evaluation model, and improves the transition probability while considering the data transfer cost and other QoS constraints simultaneously when ant finding path.Our algorithm could reach the Pareto near optimal solution rapidly with better QoS performance and lower data transfer cost from numerous candidate solutions.

Massively Multiplayer Online Cloud Game (MMOCG) is a combination of cloud gaming and Massively Multiplayer Online Game (MMOG). Running MMOGs on the cloud can reduce the cost for the players, attract potential players, and meet the nowadays preference of mobile devices. However, the MMOCGs introduce new challenges that need to be addressed. MMOGs usually consume more computational resources because of their sophisticated graphics. Hosting MMOCGs can be more expensive for game operators than hosting other genres of games on the cloud. In order to earn a decent profit, the cost of hosting MMOCGs must be reduced to an acceptable cost. Moreover, MMOCG introduce extra cloud servers to MMOG, which increased the burden of network transmission. So, the MMOCGs must have a good quality of service (QoS) and quality of experience (QoE) to retain the players. In this thesis, we study the server allocation problem to improve MMOCGs in many aspects, including money cost, QoS, and QoE. We proposed three different server allocation strategies to address the challenges of MMOCGs in different situations. We first consider the MMOCG system with several cloud gaming platforms that lie in different locations in the world, each can provide a type of server that can host MMOG on them. The cloud gaming system only obtains the information necessary for gaming. With only the log-in information, a server allocation problem is defined to reduce the money cost of hosting MMOCGs and provide an acceptable network latency for the players. The problem is formulated into an optimization problem that aims to minimize the weighted combination of server rental cost, data transfer cost, and network latency. A genetic algorithm is developed to solve the minimization problem. Extensive experiments are conducted to evaluate the proposed method and compare it with the state-of-the-art as well. The experimental results show that the method gives a lower money cost and a shorter network latency most of the time. We then consider the MMOCG system that can collect the past logging in and out data of the players. We proposed a server allocation problem that aims to reduce the cost of cloud server rental and improve the QoE of the system. A genetic algorithm with gaming time prediction is proposed to solve the problem. The gaming time of the players logging into the system is predicted by the model trained by a neural network. We conduct experiments to evaluate our method and the proposed gaming time prediction model. The results show that our prediction model can achieve almost highly accurate prediction and the server allocation strategy we proposed outperforms the prior art in both server rental and QoE. At last, we consider the influence of QoE on the movement of players during gaming time and propose a server allocation problem aiming to improve the QoE of players during gaming time. We propose a genetic algorithm with trace prediction to address this problem. A trace prediction that predicts the future positions of players for a few time steps when they log into the MMOCG system. Experiments are conducted to evaluate our method by measuring the network latency, mean opinion score, and the time consuming computation. The results show that our method outperforms the genetic algorithm without trace prediction in over 50% of the cases.

The Internet of Things (IoT) is a disruptive technology that underpins Industry 5.0 by integrating various service technologies to enable intelligent connectivity among smart objects. These technologies enhance the convergence of Information Technology (IT), Operational Technology (OT), Core Technology (CT), and Data Technology (DT) networks, improving automation and decision-making capabilities. While cloud computing has become a mainstream technology across multiple domains, it struggles to efficiently manage the massive volume of OT data generated by IoT devices due to high latency, data transfer costs, limited resilience, and insufficient context awareness. Fog computing has emerged as a viable solution, extending cloud capabilities to the edge through a distributed peer-to-peer (P2P) network, enabling decentralized data processing and management. However, IoT networks still face critical challenges, including connectivity, heterogeneity, scalability, interoperability, security, and real-time decision-making constraints. Security is a key challenge in IoT implementations, including secure data communication, IoT edge and fog device identity, end-to-end authentication, and secure storage. This paper presents an efficient blockchain-based framework that creates a secure end-to-end communication cooperative flow IoT network. The framework utilizes a hybrid blockchain network that collaborates to offer a collaborative flow of end-to-end secure communication from end devices to cloud storage. The fog servers will maintain a private blockchain as a next-generation public key infrastructure to identify and authenticate the IoT’s edge devices. The consortium blockchain will be maintained in the cloud and integrated with the permission blockchain system. This system ensures secure cloud storage, authorization, efficient key exchange, and remote protection (encryption) of all sensitive information. To improve the synchronization and block generation, reduce overhead, and ensure scalable IoT network operation, we proposed the threshold signature-based Proof of Stake and Validation (PoSV) consensus. Additionally, lightweight authentication protects resource-constrained IoT nodes using an aggregate signature, ensuring security and performance in real-time scenarios. The proposed system is implemented, and its performance is evaluated using key metrics such as cryptographic processing overhead, consensus efficiency, block acceptance time, and transaction delay. The findings show that threshold signature-based Proof of Stake and Validation (PoSV) consensus, reduces the computational burden of individual signature verification, which results in an optimized transaction latency of 80–150 ms, compared to the previous 100–200 ms without Non-PoSV. Additionally, aggregating multiple signatures from different authentication events reduces signing time by 1.98 ms compared to the individual signature time of 2.72 ms and the overhead of verifying multiple individual transactions is 2.87 ms is significantly reduced to1.46 ms along with authentication delay ranges between 95–180 ms. Hence, the proposed framework improves over existing approaches regarding linear computing complexity, increased cryptographic methods, and a more efficient consensus process.

Array redistribution is usually needed for more efficiently executing a data-parallel program on distributed memory multicomputers. To minimize the redistribution data transfer cost, processor mapping techniques were proposed to reduce the amount of redistributed data elements. Theses techniques demand that the beginning data elements on a processor not be redistributed in the redistribution. On the other hand, for satisfying practical computation needs, a programmer may require other data elements to be un-redistributed (localized) in the redistribution. In this paper, we propose a flexible processor mapping technique for the Block-Cyclic redistribution to allow the programmer to localize the required data elements in the redistribution. We also present an efficient redistribution method for the redistribution employing our proposed technique. The data transfer cost reduction and system performance improvement for the redistributions with data localization are analyzed and presented in our experimental results.

Cloud computing offerings are maturing steadily, and their use has found acceptance in both industry and research. Cloud servers are used more and more instead of, or in addition to, local compute and storage infrastructure. Deciding which workloads to outsource to what cloud provider in such a setting, however, is far from trivial. This decision should maximize the utilization of the internal infrastructure and minimize the cost of running the outsourced tasks in the cloud, while taking into account the applications' quality of service constraints. Such decisions are generally hard to take by hand, because there are many cost factors, pricing models and cloud provider offerings to consider. In this work, we tackle this problem by proposing a set of heuristics to cost-efficiently schedule deadline-constrained computational applications on both public cloud providers and private infrastructure. Our heuristics take into account both computational and data transfer costs as well as estimated data transfer times. We evaluate to which extent the different cost factors and workload characteristics influence the cost savings realized by the heuristics and analyze the sensitivity of our results to the accuracy of task runtime estimates.