One of the key enabling solutions to in-orbit extract information from Earth Observation images is given by deep learning techniques. However, the accuracy of these algorithms is strictly related to the availability of large datasets of satellite images for training purposes. Limitations on the available transmission bandwidth in the orbital context may prevent the possibility to downlink all acquired images to a node where centralized training happens. For this reason, we propose a communication strategy to support a completely distributed learning (DL) technique to train a deep learning model in-orbit, by leveraging the fact that satellites may form a network thanks to the potential availability of Inter-Satellite Links (ISLs) within and between orbital planes. Our proposal is different from a FL approach since we provide for each satellite to receive all the information needed to calculate an updated global model by itself, without leaning on a central parameter server. Numerical results show that DL outperforms FL in number of learning rounds completed in the unit time, allowing for reaching validation accuracy convergence in a shorter time.

One of the key enabling solutions to in-orbit extract information from Earth Observation images is given by deep learning techniques. However, the accuracy of these algorithms is strictly related to the availability of large datasets of satellite images for training purposes. Limitations on the available transmission bandwidth in the orbital context may prevent the possibility to downlink all acquired images to a node where centralized training happens. Instead, Federated Learning (FL) could be fruitfully leveraged in this scenario, since it provides for each satellite to train a local model only with its own dataset, and then to share its trained model with a central server, which receives models trained by the different satellites and aggregates them into a new global model being eventually shared with all the satellites, and this repeats until convergence is reached. However, because communication with a node acting as a central parameter server may be still limited by short visibility time, the described process may need a long time because of limited communication windows, negatively impacting the time needed to reach model convergence. For this reason, we propose a communication strategy to support a completely distributed learning technique to train a deep learning model in-orbit, by leveraging the fact that satellites may form a network thanks to the potential availability of Inter-Satellite Links (ISLs) within and between orbital planes. Our proposal is different from a FL approach since we provide for each satellite to receive all the information needed to calculate an updated global model by itself, without leaning on a central parameter server. Numerical results show that distributed learning outperforms FL in number of learning rounds completed in the unit time, allowing for reaching validation accuracy convergence in a shorter time, as it has been verified on a land coverage classification task based on the EuroSAT dataset.

The processing capacity available on satellites within Low Earth Orbit constellations allows for the implementation of data processing on-board of satellites. Furthermore, by endowing satellites with Inter-Satellite Links, it is possible to obtain a network within the constellation, enabling data routing and allowing processing to take place on any satellite of the constellation, even different from the task originating one. Edge computing solutions for Earth Observation (EO) applications have been proposed in satellite environments, where the data processing is accomplished by either the satellite running the observation task (Always First) or by a datacenter connected to a ground station (Always Ground). We propose a solution in which any satellite of the constellation may process the EO task and its choice is performed so as to minimize the sum of the memory, processing and transmission costs. We propose and evaluate the effectiveness of a heuristic that jointly performs the following operations: i) the choice of the element (satellite, datacenter) performing the task processing; ii) the choice of the routing path on which the task or its processing result is routed from the observation satellite to the datacenter. The proposed solution is compared to the Always First and Always Ground benchmark solutions and both the cost advantages and the potential delivery delay reduction are evaluated.

By endowing satellites in Earth Observation (EO) constellations with Orbital Edge Computing (OEC) capability, i.e, with the ability of processing acquired images directly on-board, it is surely possible to use bandwidth more effectively, since only the actual useful information extracted from the image is transferred on ground. However, OEC can be fruitfully leveraged also to minimize the energy consumption on ground stations due to image processing. In fact, in EO satellites energy is allocated in advance by endowing them with appropriate solar panels and batteries, and this amount of energy is always generated, even when it is not necessary. On the contrary, energy on ground station is closely related to its demand. For this reason, we propose and investigate a strategy to allocate processing and bandwidth resources in OEC-endowed EO satellite network to minimize energy consumption on ground stations due to on-ground processing. By taking into account the energy budget on Sentine1-2, having a 26% of available energy being unused, we show how with our approach it is possible to obtain a substantial reduction in energy consumption on ground stations even by using the 1% of the total energy budget when appropriate computational capacity is available on board, while this consumption can even drop to zero by increasing the energy dedicated to OEC operations to the 15% of the energy budget.

The next step in Earth Observation (EO) constellations will be leveraging Inter-Satellite Links (ISLs) to form a network where information generated by the EO application can be transmitted, in such a way that, by endowing spacecrafts with processing capacity, observation data may be processed directly in orbit by any satellite of the constellation. However, since bandwidth and on-board processing capacity are valuable resources, strategies to appropriately routing the information and deciding on which node it has to be processed shall be defined. In this work, we formalize and solve an optimal bandwidth and computing resource allocation problem in Low Earth Orbit (LEO) satellite constellation for EO applications. In order to deal with the complexity of the proposed optimization problem, we also present two heuristics requiring different computational effort. In the proposed problem formalization, processing can happen on any node of the network (i.e., either on the data source satellite, on any other satellite of the constellation or on ground station). After having validated the proposed heuristics by comparing their results to the optimization problem ones, we apply them to a real orbital scenario, showing their ability to reduce both total cost and data delivery delay to ground with respect to state-of-the-art solutions.

The increasing number of mobile devices and the enhanced user experience they require have a strong impact on mobile network development, since they result in an increased channel capacity demand to be obtained with a limited site densification. An interesting approach to face this challenge can be found in the combination of two technology enabling solutions: Cloud or Centralized RAN (C-RAN) and Advanced Antenna Systems (AAS). In this paper, we discuss the advantages given by these solutions, with a special focus on how Artificial Intelligence (AI)-based algorithms can improve their combination in terms of functional split and antenna mapping, stated as an optimization problem. In particular, AI can be beneficial in three main areas, such as the actual solution of the optimization problem, the tuning of parameters used in classical heuristic algorithms aiming at solving the optimization problem, and, finally, the traffic and resource allocation prediction at the base of proactive reconfiguration frameworks.

An innovative resource allocation framework for virtualized network environments and based on the application of Artificial Intelligence techniques is proposed and investigated. It integrates the needed processing capacity prediction procedure and the allocation one which determines the capacity under-allocation or over-allocation needed to minimize the resource allocation and Quality of Service degradation costs. The proposed solution is based on: i) a monitoring procedure in which the processing capacities required by virtual instances are periodically monitored; ii) an integrated allocation/prediction procedure in which the processing capacities to be allocated to the virtual instances are evaluated in time intervals successive to the monitoring periods. This second procedure uses a Convolutional/Long Short Term Memory neural network whose loss function is defined so as to minimize an overall cost dependent on both the cloud resource allocation and Quality of Service degradation costs. We evaluate the proposed solution in backbone and metropolitan traffic and network scenario. We show how in the traffic scenario of an Italian Mobile Operator in Milan zone, the proposed solution far outperforms the not integrated classical solution in which the capacity prediction and allocation procedures are separately performed. Furthermore its performance are very near to the one of an oracle that optimizes the capacity over/under dimensioning parameter.

The high reconfiguration time of cloud resources in Network Function Virtualization architecture has led to make ineffective the reactive cloud resource allocation procedures whose application lead to over-allocate resources or to degrade Quality of Service in decreasing/increasing traffic scenario. Recently many Artificial Intelligence (AI)-based allocation procedures have been proposed to pre-allocate cloud resource according to required processing capacity predictions. In this paper we illustrate how an ETSI NFV architecture can support these predictions procedures. Furthermore they aim to exactly predict the processing capacity to be allocated and they are all based on the minimization of a symmetric loss function of the neural network. For this reason we propose a resource allocation procedure with a asymmetric loss function whose parameters are dependent on an overall cost expressed in terms of allocation and QoS degradation costs. We prove that the proposed solution allows for a cost reduction in the order of 30% in a typical NFV traffic and network scenario.

Service Function Chaining (SFC) is an enabling technology to provide end-to-end service differentiation according to specific user requirements. Although emerging technologies such as Software-Defined Networking (SDN) and Network Function Virtualization (NFV) are perfect enablers for SFC, hardware limitation of Ternary-Content Addressable Memories (TCAMs) can be an obstacle when handling a large variability of SFC requests, derived from the increasing number of users, and the heterogeneity of applications and Quality of Service (QoS) requirements. This article introduces and investigates the problem of TCAM size limitation on the classification procedure of SFC requests in SDN-based SFC environments. To overcome this limitation, the classification of incoming SFC requests is proposed to be offloaded to transient nodes when the occupation of the ingress node flow table is close to its maximum. An Integer Linear Programming (ILP) formulation is provided to formalize the Chain Request Classification Offloading (CRCO) problem, that consists in maximizing the number of SFC requests that can be served. Furthermore, a heuristic algorithm is presented to solve the CRCO problem in feasible time. The performance evaluation carried out over two real topologies, shows that the proposed offloading strategy can greatly increase the number of accepted requests without significantly affecting the network QoS.

Network function virtualization foresees the virtualization of service functions and their execution on virtual machines. Any service is represented by a service function chain (SFC) that is a set of VNFs to be executed according to a given order. The running of VNFs needs the instantiation of VNF Instances (VNFIs) that in general are software modules executed on virtual machines. The virtualization challenges include: 1) where to instantiate VNFIs; ii) how many resources to allocate to each VNFI; iii) how to route SFC requests to the appropriate VNFIs in the right sequence; and iv) when and how to migrate VNFIs in response to changes to SFC request intensity and location. We develop an approach that uses three algorithms that are used back-to-back resulting in VNFI placement, SFC routing, and VNFI migration in response to changing workload. The objective is to first minimize the rejection of SFC bandwidth and second to consolidate VNFIs in as few servers as possible so as to reduce the energy consumed. The proposed consolidation algorithm is based on a migration policy of VNFIs that considers the revenue loss due to QoS degradation that a user suffers due to information loss occurring during the migrations. The objective is to minimize the total cost given by the energy consumption and the revenue loss due to QoS degradation. We evaluate our suite of algorithms on a test network and show performance gains that can be achieved over using other alternative naive algorithms.

Network virtualization techniques allow for the coexistence of many virtual networks hosted in the same substrate network. Virtual router migration allows for resource consolidation with the consequence to reduce the power consumption in low traffic periods. Unfortunately, virtual router migration has the effect of degrading the quality of service during the downtime in which the router is not able to carry on its forwarding function. For this reason, despite the advantages in power consumption saving, the migration technique should be applied with parsimony to avoid an excessive QoS degradation. The objective of this paper is to study virtual network embedding (VNE) problems aware of both operation and reconfiguration costs that are characterized by the energy consumption and the revenue loss due to QoS degradation. Both embedding optimization problem formulation and feasible solutions to the problem in cycle-stationary traffic scenario and when the possible embeddings are determined a priori will be given. Two migration policies will be introduced and compared. The first one, referred to as “global,” is based on the knowledge of the entire daily traffic profile and it is determined by solving a Markov decision process. The second one, referred to as “local,” is only based on the knowledge of the current traffic. The results achieved show how the application of the global policy allows better performance than the local one and its application can lead to a cost reduction in the order of 35% with respect to traditional migration policies.

This paper deals with an energy saving routing solution, called Energy Saving IP Routing (ESIR), to be applied in an IP network. ESIR operation is integrated with Open Shorthest Path First (OSPF) protocol and allows the selection of the links to be switched off so that the negative effects of the IP topology reconfiguration procedures are avoided. The basic mechanisms which ESIR is based on are the concepts of SPT exportation and move. These mechanisms allow to share a Shortest Path Tree (SPT) between neighbor routers, so that the overall set of active network links can be reduced. Properties of moves are defined and the energy saving problem in an IP network is formulated as the problem of finding the Maximum Set of Compatible Moves (MSCM). The MSCM problem is investigated in two steps: firstly, a relaxed version of the problem, named basic MSCM problem, is considered in which QoS requirements are neglected; in the second step, the solution of the full problem, named QoS-aware MSCM problem, is faced. We prove that the basic MSCM problem can be formulated as the well-known Maximum Clique Problem in a graph; instead the QoS-aware MSCM introduces a condition equivalent to the Knapsack problem. ILP formulations to solve both the problems are given and heuristics to solve them in practical cases are proposed. The performance evaluation shows that in a real ISP network scenario ESIR is able to switch off up to 30% of network links by exploiting over-provisioning adopted by operators in the network resource planning phase and typical daily traffic trend.

The author evaluates the performance of bufferless optical packet switches equipped with shared tuneable optical wavelength converters (TOWC) and supporting the quality of service (QoS). The TOWCs are shared per output fibre and, hence each output fibre has a dedicated pool of TOWCs. The TOWCs are not completely shared within the node. The QoS, expressed in terms of the packet loss probability, is differentiated by giving the packets different priorities in accessing both the output wavelengths and the TOWCs. A scheduling algorithm is proposed and its complexity is evaluated. The QoS technique's effectiveness is studied by means of an analytical model based on a probabilistic approach and validated by simulation results. The results show that the proposed technique to support the QoS facilitates a good differentiation of the QoS offered to packets.

An optical packet switch that shares both limited range and full range wavelength converters for contention resolution is proposed with the aim to guarantee an high conversion cost saving. To optimally dimension the number and the conversion range of the wavelength converters, an analytical model, validated by simulation, is introduced to evaluate the packet loss probability of the switch. Numerical results show that the proposed switch architecture allows for a conversion cost saving in the order of 90% with respect to a classical architecture in which only shared full range wavelength converters are used.

In this paper we evaluate the performance of Bufferless Optical Packet Switches equipped with shared Tunable Optical Wavelength Converters (TOWC) and supporting Quality of Service. The TOWCs sharing is partial, that is each Output Fiber shares one pool of TOWCs shared by all of the packets directed to that Output Fiber. The QoS, expressed in terms of Packet Loss Probability, is differentiated by giving the packets different priorities in accessing both the output wavelengths and the TOWCs. A scheduling algorithm is proposed and its complexity is evaluated. The QoS technique effectiveness is studied by means of a sophisticated analytical model validated by simulation results. The results show that the proposed technique to support QoS allows a good differentiation of the QoS offered to the packets.

In this paper we analyze intra-domain routing protocols improvements to support new features required by realtime services. In particular we introduce OSPF fast convergence and highlight the advantage of using a dynamic algorithm instead of the Dijkstra one to compute the shortest paths. Then we propose a new multi-path dynamic algorithm which uses multipath information to make a fast determination about the new shortest paths when a link failure occurs, reducing this way the network re-convergence time. To evaluate the proposed algorithm performance we have implemented it in the OSPF code of the Quagga open-source routing software. We compare our own algorithm with three different dynamic algorithms, like the one implemented in Cisco routers and the two others, well known in literature, proposed by Narvaez and Ramalingam-Reps. We show how, by exploiting multi-path information, our algorithm performs, in many case studies, better than the above algorithms, especially in a link failure scenario.

An architecture is proposed for a wavelength division multiplexed (WDM) optical packet switch equipped with both limited range wavelength converters (LRWCs) and shared full range wavelength converters (FRWCs). The FRWCs are used to overcome the performance degradation in terms of packet loss probability due to the use of LRWCs only. A probabilistic model is proposed to dimension the number of shared FRWCs so that the same packet loss probability of a switch equipped with only shared FRWCs is guaranteed. After introducing a cost model of the converters depending on the range conversion, we show that the architecture may allow conversion cost saving in the order of 85%.

It is cost-effective to install multiple fibers in each link of an all-optical network, because the cost of fibers is relatively low compared with the installation cost. In this paper we propose an architecture of an bufferless multi- fiber optical packet switch (OPS) using the wavelength dimension for contention resolution. In order to save the number of tunable optical wavelength converters (TOWC), the proposed architecture shares the TOWCs among the arriving packets. An analytical model, validated by means of simulations, is proposed to evaluate the Packet Loss Probability of the switch. A sensitivity analysis of the TOWCs saving of a multi-fiber switch with respect to a single-fiber switch is carried out as a function of the main switch and the traffic parameters (number of input lines, number of wavelengths per fiber, offered traffic,......). In some cases the saving can reach 50%.

This paper analyses the performance of an Optical Packet Switch equipped with multicast wavelength converters shared among the arriving packets. The architecture performance is evaluated by means of an analytical model and it is compared to the one of an Optical Packet Switch equipped with unicast wavelength converters. The obtained results show that in a multicast traffic scenario the proposed architecture allows the Wavelength Converters to be saved. In some cases the gain can be in the order of 95%.

OSPF (open shortest path first) is a widely used intra-domain routing protocol in IP networks. Processing delays in OSPF implementations impact the time needed for both intra-domain and inter-domain routing to reconverge after a topology change. In this paper we introduce a performance index, referred to as the switching time, allowing the router reconverge to be characterized when network topology changes occur. We propose a test methodology in order to measure the introduced performance index in router realized with the personal computer hardware and equipped with Quagga, the most used open routing software. An optimization of the Quagga routing software has been performed and better performance has been obtained in terms of switching time

This paper evaluates the performance of an optical packet switching architecture that uses the wavelength conversion technique to solve the packet contention problem. The architecture, referred to as shared per node (SPN), is equipped with limited-range wavelength converters shared per node. We evaluate for this architecture the performance of the optimum scheduling algorithm by resolving an integer linear programming problem. Unfortunately the optimum scheduling algorithm has a high time complexity and it cannot be adopted for optical packet switches requiring packet scheduling time in the order of 10/100 ns. For this reason we propose some scheduling algorithms allowing the best optimum scheduling performance to be reached in low computation time. In particular an algorithm, based on heuristic rules, is proposed able to reach good performance with a complexity O(M),M being the number of wavelengths used

This paper compares selected optical packet switching architectures that use a wavelength conversion technique to solve the packet contention problem. The architectures are equipped with shared and limited-range wavelength converters. This paper focuses on two architectures: the shared per node (SPN) and the shared per output fiber (SPOF) architectures. The analytical results show how the use of a reduced number of limited-range wavelength converters with small range allows these switches to obtain the same performance as an architecture equipped with all of the wavelength converters and using a full wavelength conversion.

The objective of this study is to propose a new optical packet switching architecture in which the wavelength converters, needed to solve output packet contentions, are shared per input line; according to this sharing strategy the packets arriving at a given input fiber shares a converter pool that can be accessed when wavelength conversions are required. In the paper we evaluate the performances of the proposed architecture when control algorithms with different complexity, are adopted. Under a unicast traffic scenario the obtained performances are compared to the ones of the architectures in which the wavelength converters are shared per node and per output line respectively. The carried out comparison shows that, with respect to the architecture with wavelength converters shared per output line, the proposed architecture allows to obtain a 30% saving of wavelength converters when a simple control algorithm is adopted; the saving can reach 80% if an optimized control algorithm is used. On the contrary the architecture with sharing per node needs fewer wavelength converters than the proposed architecture, but has the drawback to have a switching matrix much complex evaluated in terms of needed number of semiconductor optical amplifier; this complexity is due to the flexibility provided to the arriving packets in accessing the converters bank shared per node.

The objective of this study is investigate the performance difference of packet switching architectures working in a synchronous and asynchronous way respectively; in such architectures the packet contention is resolved in the wavelength domain and the used wavelength converters are shared; we investigate on the saving of the number of converters that the sharing technique allows to obtain in the synchronous and asynchronous architectures and compare the obtained results. These ones show that when a packet loss probability is fixed, in the synchronous case a greater number of converters is saved, in some cases the gain is 40% more than the asynchronous case; furthermore in the asynchronous case a more expensive switching matrix is needed.

In this paper, the optical burst switching with burst dropping (OBS/BD) technique is proposed to be implemented in an all-optical backbone network to support IP traffic. The OBS/BD is based on two main features: i) several IP packets are assembled in a single macro-packet, called burst; ii) the burst contention in an optical switch is handled by the means of two techniques: the wavelength dimension and the "burst dropping". In the optical burst switching (OBS), the entire burst is discarded when all of the output wavelengths are engaged at its arrival instant. Whereas, the OBS/BD technique discards only the initial part of the burst and forwards the final part of the burst beginning at the instant in which one wavelength becomes free. Obviously, the OBS/BD, respect to the OBS, allows to increase the switch throughput, i.e. the number of forwarded IP packets. We develop the analytical models that quantify this increase tacking into account of several system parameters.

In this paper, we present a system for singling out character strings from line drawings. Characters, candidate according to a preliminary selection, are replaced with rectangles preserving information about character size, position and orientation. Rectangles are clustered on the basis of a set of criteria. Each cluster is seen as a graph whose nodes represent rectangles, while arcs represent their possible connections, established on the basis of properly defined visibility criteria. Strings within clusters are searched, in a proper state space, according to a hierarchy derived from the knowledge of the problem, injected by means of models representing the most common configurations of strings in terms of their topological and geometrical properties. The system is implemented in an object-oriented fashion. Experimental results on a large set of documents of different types are reported.

Presents a technique for the detection and separation of symbols connected to graphics in line drawings. The approach is based on a special image representation, that the authors call graph representation. The graph representation is especially convenient for line-like images, since it decomposes the line structure into 'edges' and 'nodes', that formalise the intuitive notions of 'line' and 'crossing point between lines'. In this way, well-known algorithms from graph theory can be used. The graph is searched for characteristic sub-graphs, revealing the presence of symbols or noisy objects overlapping lines. Each symbol candidate is accurately processed, in order to make the area of its connection to lines as regular as possible. Actually, two different algorithms have been designed to perform this task: one for symbols crossing lines and the other for symbols only touching lines. Both of them are rotation invariant.>