The usual Machine Learning (ML) uses raw clients in a centralized mechanism and generalized training, which demands a lot of time and processing to send the network this sensitive data and to process only on Edge. Federated Learning (FL) improves the ML by using local training in each client device, selecting some specific clients to train, and sending only their local parameters for an edge. In this way, a client selection mechanism is essential to select clients to ensure data diversity and high-quality data, which enables global model to generalize well across different data distributions. In addition, while high-quality data contributes to the accuracy and reliability of the learned model. However, it is important to asses the device importance based on multiple criteria, aiding in identifying and prioritizing the most valuable and promising clients for participation in the training process. In this paper, we introduce a Robust multi-criteria Client Selection Mechanism called RICAm, which address the challenges of non-IID data FL environments. RiCAm uses client selection to search for the best-fit clients to train, calculated using the multicriteria method MACBETH. In our evaluation, RiCAm reached only 16 rounds, 80% of which is around 10% more than the standard random selection approach random selection approach.
Serverless computing offers an effective way to improve resource utilization and can significantly leverage edge-cloud environments. In this context, serverless request response times are one of the main focuses for optimization. However, the end-to-end routing path from a request to a function, which considerably contributes to the overall delay, is often overlooked. This oversight raises concerns about the actual performance of serverless in edge-cloud deployments, as serverless architectures are initially designed for cloud computing where node-to-node delay is negligible. To address this concern, our paper examines the current serverless networking design in a realistic edge-cloud setup. Our findings reveal significant performance degradation when serverless is applied in such environments due to key mismatches in the networking architecture. To mitigate these issues, we propose a novel latency-aware scheme and a load balancing mechanism for edge-cloud environments, called FaaSt. This proposal is implemented using open sources and standards, resulting in a significant reduction in response times.
Cloud-native applications, characterized by scalability, resilience, and flexibility, adopt microservices architectures to decompose applications into smaller, independently manageable services. Although microservices offer significant benefits, this architectural approach introduces challenges in service-to-service communication, commonly relying on advanced orchestration and communication frameworks such as Kubernetes and Istio, respectively. However, the added complexity imposes substantial overhead by introducing longer packet processing paths. This paper discusses performance bottlenecks arising from service meshes and proposes eZtunnel to address some of the identified challenges. Leveraging extended Berkeley Packet Filter (eBPF) to transparently offload networking traffic, the elongated network path is bypassed, optimizing resource utilization and enhancing application performance. Experiments show that eZtunnel can reduce median latency over 20% and jitter to almost 10%.
After identifying an artifact as malware, the additional challenge arises of correctly classifying it into a specific family. This raises the challenge of classifying malware amid growing complexity and the increasing volume of cyber threats. Therefore, we introduce FORTUNATE, a framework that employs variable-length instruction sequences to classify malware more efficiently and accurately, focusing on real and active malware. By introducing effective methods for extracting opcodes and their representation in the smallest possible vectors, FORTUNATE surpasses the limitations of previous approaches, enabling the correct identification of malware in their respective families, while avoiding redundancies in the data. The results achieved with FORTUNATE demonstrate advances in both precision and computational efficiency, consolidating it as a significant contribution to the field of cybersecurity.
Modern cloud-native applications, built with containerization and microservices architectures, present significant hurdles for proactive failure prediction and classification. Traditional anomaly detection and threshold-based algorithms often fall short: they either lack proactive capabilities or fail to provide granular enough output for effective remediation and root cause analysis. This paper proposes a novel, multi-stage framework that tackles these limitations. The framework leverages a combination of time-series forecasting models, such as NHITS, to predict resource consumption patterns. Subsequently, machine learning models like KNN and Random Forest are utilized to classify failures across various dimensions, including service names, HTTP methods, and statuses. This multistage approach empowers fine-grained identification of potential failures and their root causes. Our evaluation demonstrates the framework's effectiveness in a simulated environment. The NHITS model outperformed even cutting-edge architectures in predicting resource utilization. Additionally, the chosen machine learning models exhibited promising accuracy in classifying failure types.
Enterprises are adopting multi-cloud strategies, es-tablishing overlays atop two or more cloud providers (CP) backbones to connect resources and services, even across conti-nents. Simultaneously, there is a significant increase in submarine cable deployments by the CPs to enhance the reliability and performance of their backbone networks. However, enterprises face challenges in understanding how these deployments impact their adoption of multi-cloud strategies. These challenges include the evolving nature of submarine cable deployments, the high cost (and resulting paucity) of data collection efforts on multi-cloud network paths, and a general unawareness of how these deploy-ments impact multi-cloud network path latencies in practice. To address this problem, this work presents a third-party measurement study to corroborate the latency trends/changes characteristics of multi-cloud network paths with submarine cable deployments. To this end, we develop a three-step approach: (1) analyze latency characteristics of multi-cloud paths by comparing two sets of measurements obtained across three major CPs, (2) examine the possible root causes of latency trends by leveraging publicly available data sources on submarine cable deployments, and (3) validate our findings with submarine cable operators and providers. Our study reveals several insights into the impact of submarine cable deployments on multi-cloud network paths' latency characteristics, helping enterprises make informed decisions regarding their cloud overlays. To promote reproducibility and extension of our work, we will release the code and datasets to the community.
Network-level load balancing algorithms have been proposed to help provide application flows with adequate quality-of-service (QoS) by assigning traffic to different network paths to avoid congestion as well as improve overall network utilization. As datacenters scale up and become more geographically distributed, wide-area network (WAN) datacenter traffic, which typically consists of data-heavy tasks, has become increasingly prevelant. Some of the noteworthy challenges raised by the coexistence and interaction between WAN- and intra-datacenter (DC) traffic are the differences in their QoS requirements, link utilizations and round-trip times. To the best of our knowledge, these challenges have not yet been addressed by existing datacenter load balancers. To highlight this gap, in this paper, we conduct a comparative performance study of state-of-the-art datacenter load balancers considering datacenter topologies and workloads that reflect how datacenters have evolved to keep up with their driving applications. Through extensive simulations, we study state-of-the-art load balancers under different network topologies and workloads, including DC-, WAN-, as well as DC-WAN mixed workloads. Our study shows that current load balancers are not able to adequately distribute load under WAN and DCWAN mixed traffic. Additionally, we also observe that load balancer performance is highly sensitive to its parametric choices, especially when subject to WAN and DC-WAN mixed workloads. Our study also highlights that the interplay between transport-level congestion control and load balancing mechanisms plays a significant role and must be considered in order to meet desired performance goals. We also share our experience engineering our experimental methodology by (1) providing recommendations on how to set load balancer parameters and (2) introducing a datacenter workload modeling framework to represent both synthetic and more realistic datacenter workloads, including mixed DC and WAN traffic.
Resources in a cloud environment are limited. Containers in a Kubernetes Cluster are always in contention for resources. Kubernetes orchestrates the scheduling of pods on different nodes. Accommodating as many containers as possible on a node is a goal state that must be reached to ensure optimal resource utilization. This indirectly helps with operational and hardware costs. This paper evaluates the two most popular high-level container runtime systems, Docker and ContainerD. We study the impact of these two container runtimes on different computing resources such as CPU, I/O, Memory, and Network latency. The evaluation is performed on Google Kubernetes Engine (GKE) and Amazon Elastic Kubernetes Service (EKS). We study the performance of both cloud providers separately. The aforementioned has provided us with a more rigorous and accurate analysis since the underlying hardware would be the same. In the second part of the paper, we have bootstrapped our own custom Kubernetes Cluster from scratch using Terraform and Kubeadm on Amazon Web Services and performed an analysis similar to those mentioned above compared to the Managed AWS Kubernetes offering. Through the two experiments, we want to determine which container runtime system performs better in both the Managed Kubernetes Services. Secondly, whether deploying your own K8s cluster from scratch provides any significant performance gains or losses.
The translation of complex policies into practical network configuration actions is often conducted by human operators, resulting in errors that affect network performance. In Open RAN networks, this challenge becomes even more relevant, considering the objective of intelligent configuration automation. This paper proposes an intent-based management system, which receives commands in natural language through a chatbot and converts them into network policies. Proposal evaluation is conducted on a software-defined network emulation on Mininet, controlled by ONOS. The system prototype integrates Rasa chatbot with ONOS controller, allowing the configuration of network flows to be derived from natural language instructions. The results indicate that the system fulfills intents in constant time, up to 100 ms, even with an increasing load of flows in the network switches.
Data centers create the backbone of the modern Internet. However, internal network traffic characteristics are closed know-how of data centers. We have collected an internal network traffic analysis based on the data from one of major world data centers. We have analysed 4 internal network traffic characteristics (clustering of IP addresses, application traffic patterns, frequency of changes in cluster topologies and histograms of communicating IP address pairs). The data set has been published at GitHub.
This article focuses on the development of an intelligent data center at the Edge, using nodes that include servers. The primary advantage of these Edge data centers is their ability to quickly respond to user demands while utilizing only Edge resources. Additionally, they reduce energy consumption by eliminating the need for intermediate equipment like routers, switches, and repeaters on optical fiber. The article also aims to provide Artificial Intelligence (AI) functions, specifically focusing on handling Small Language Models (SLMs) to meet the growing demand for AI capabilities. The final section of this article presents a case study involving a prototype network of drones capable of autonomously responding to a variety of requests.
Given modern clouds' dynamic and scalable environment, cloud resource management is a complex and multifaceted task involving several challenges. In this context, self-parameterization of metaheuristics is an advanced technique that can optimize cloud resource management. Motivated by the need to tune the parameters of metaheuristics to improve their efficiency, we propose a self-parameterization method based on knowledge representation and inference that can be applied to optimize resource allocation, scalability, and load balancing in cloud environments. This method, which we call DT-Tuning, uses a Decision Tree to dynamically adjust the value ranges of the parameters of a metaheuristic. A broad set of experiments validated the functioning of DT-Tuning in the self-parameterization of the hybrid metaheuristic GRASP with Path-Relinking (GRASP+PR) in the data clustering problem. We compared DT-Tuning with manually parameterized GRASP+PR and the I/F-Race and BRKGA auto parameterization algorithms about solution quality and execution time. The results highlight the potential of DT-Tuning in improving solutions, reducing execution time, and improving the quality of clusters compared to manually parameterized metaheuristics.
Port scanning is an important technique for collecting sensitive information, highlighting the need for enhanced security systems. It is emphasized that port scanning, considered an anomaly, should be identified and suppressed early, especially given the significant number of reported incidents. In response to this challenge, this paper presents RAVEN, an intelligent and automated system capable of analyzing network flows to detect and classify scans quickly. Contributions include implementing and evaluating RAVEN, demonstrating performance improvement with the expansion of features, and making datasets available to the academic community. Additionally, the RAVEN system is capable of detecting and classifying attacks on web services, including those using the Tor network. Web attacks are quickly detected using Bloom filters, yielding high accuracy.
Digital twin network are emerging as key drivers for future automated and high-performance networks. Digital twin networks create virtual representations of physical networks, enabling real-time monitoring, simulation, and optimization. The accuracy and timeliness of data are crucial for building a precise digital twin network. However, constructing an accurate digital twin network faces significant challenges due to the difficulties in data collection. In network measurement and data collection, data uncertainty is often unavoidable due to observer effects, where the act of measurement itself imposes an impact on the system. This phenomenon can introduce biases or perturbations that compromise the accuracy of digital twin network models, leading to less precise representations of the network's actual state and behavior. This paper systematically reviews existing measurement schemes and proposes a new classification. We evaluate these measurement methods within a simple network environment, analyzing their impact on system performance and delving into the underlying causes of performance degradation. These insights contribute to the development of a more accurate and efficient digital twin network.
Vehicular network connectivity within Intelligent Transport Systems (ITS) is essential for enabling seamless data and resource sharing, including transmitting critical safety messages, traffic management information, entertainment, and comfort services. This connectivity enhances the user experience by supporting complex interactions between vehicles and infrastructure in dynamic network environments. Edge connectivity management has recently gained attention for maintaining connection stability while managing complex models effectively. In this context, connectivity refers to vehicles' ability to maintain a stable and robust link with network resources and other vehicles for optimal data exchange. In this paper, we propose an edge connectivity management approach, the Edge Connectivity Estimation Model ECEM, aimed at ensuring connection stability and strength. We design and implement the SARSA Reinforcement Learning (RL) algorithm to assess and estimate the overall connection reliability, determining the optimal vehicular edge - a selective combination of vehicles within groups - to ensure superior connection strength for data and resource sharing, even in high-mobility scenarios. This estimation process helps identify the most suitable edge to meet the data-sharing requirements for each vehicle. Our approach considers multiple parameters, including mobility, application parameters, and network density. Extensive realistic simulations have demonstrated that our proposed approach outperforms existing methods by reducing packet loss and delay while increasing throughput.
This paper presents UBOTA, a novel protocol designed for Over-the-Air software updates in vehicular systems. Despite replacing TCP by UDP, UBOTA guarantees reliable transmission and optimizes software delivery speeds through bursts of encrypted UDP packets, reducing the reliance on establishing connections in unstable mobile networks. Addition-ally, the protocol reduces the total data transmitted, leading to cost savings due to a more efficient use of network resources. Our implementation extends our prior work on securing OTA updates, adding a focus on transmission speed and data efficiency. Experimental results indicate a 50% improvement in speed under ideal lossless network conditions as well as a 500% increase in very lossy networks. Additionally, UBOTA protocol is also 45% more efficient in terms of data volume. Performance analysis underscores the critical role of the retransmission timer, revealing its impact on packet duplication and redundant requests.
