6th-generation (6G)mobile networks are envisioned as AI-native systems, integrating learning and inference across the entire protocol stack. Although 5G’s 3GPP Network Data Analytics Function (NWDAF) introduced analytics-driven automation, it lacks standardised support for model lifecycle control, Data Analytics as a Service (DAaaS), closed-loop feedback, and largescale interoperability. To address these gaps, we propose a protocol-based framework for AI-as-a-Service (AIaaS) management for 6 G, centered on an enhanced NWDAF architecture with four components: Model Lifecycle Orchestrator, Model Registry & Validator, Distributed Execution Engine, and Feedback Aggregator. It introduces two lightweight, service-based interfaces: Model Training and Creation Protocol (MTCP) for intent-based model training and publication, and Model Execution Protocol (MEP) for on-box inference and metric feedback. We validate the framework via formal verification under message loss using a reproducible TLA+ model with three NFs and two model versions. Results show that NWDAF can evolve into a feasible AI lifecycle manager, enabling scalable and stable AI-native deployments in 6G. Complexity modelling confirms linear resource scaling up to 128 network functions (theoretical), with the public TLA+ specification configured for 3 NFs.

Emerging cloud-native applications challenge cloud computing to provide Ultra-Reliable Low Latency Communication (URLLC). To address this challenge, the integration of cloud computing with Time-Sensitive Networks (TSN) has been explored in recent studies. Despite these efforts, existing research lacks a comprehensive analysis of how dynamic cloud computing environments impact the performance of this integration. In this paper, we address research questions related to the performance of TSN in cloud computing, considering different scenarios. Our results demonstrate that TSN can enhance the performance of time-sensitive application (360 -Virtual Reality videos) under certain conditions, but also highlight challenges in maintaining performance as cloud environments change. Furthermore, the study emphasizes the critical role of the convergence of computing and networking resources in meeting the stringent performance requirements of dynamic applications. This work offers valuable insights into the potential and limitations of TSN in cloud environments, providing a foundation for future research in this area.

As network management operations increasingly rely on automation and finer control actions, there is a need for precise telemetry systems. In-band Network Telemetry (INT) methods use data packets to carry telemetry and give real-time insights about network performance. Existing solutions often require specialized hardware or offer limited runtime configuration options. This work presents an INT Framework for heterogeneous private networks, targeting industrial and multimedia applications. The framework is designed to be flexible and runtime-reconfigurable, addressing challenges in real-world applications. We provide implementation details of our elements supporting the configurability and the consolidation of raw telemetry into high-level Quality of Service (QoS) metrics. We evaluated the framework in a testbed with wired and wireless devices. The results show the accuracy in monitoring QoS, as well as an analysis of synchronization requirements, showcasing the feasibility of our framework for solutions requiring precise and flexible QoS monitoring.

Cloud Gaming (CG) has become one of the most important cloud-based services in recent years by providing games to different end-network devices, such as personal computers (wired network) and smartphones/tablets (mobile network). CG services stand challenging for network operators since this service demands rigorous network Quality of Services (QoS). Nevertheless, ensuring proper Quality of Experience (QoE) keeps the end-users engaged in the CG services. However, several factors influence users’ experience, such as context (i.e., game type/players) and the end-network type (wired/mobile). In this case, Machine Learning (ML) models have achieved the state-of-the-art on the end-users’ QoE estimation. Despite that, traditional ML models demand a larger amount of data and assume that the training and test have the same distribution, which can make the ML models hard to generalize to other scenarios from what was trained. This work employs Transfer Learning (TL) techniques to create QoE estimation over different cloud gaming services (wired/mobile) and contexts (game type/players). We improved our previous work by performing a subjective QoE assessment with real users playing new games on a mobile cloud gaming testbed. Results show that transfer learning can decrease the average MSE error by at least 34.7% compared to the source model (wired) performance on the mobile cloud gaming and to 81.5% compared with the model trained from scratch.

Millions of videos are watched per minute on the Internet. Due to real-time performance demands, such as high-quality video streaming, network administrators face new challenges to control the network and cope with the expected quality of experience (QoE). Automatic control is a necessity to reduce the OPEX, because it could reduce the need for resource overprovisioning, as well as the number of human administrators. Dynamic rate in video streaming alleviates the resource usage, but it worsens the video quality when a network bottleneck occurs, lowering the QoE. This paper dynamically adjusts the IEEE 802.11 parameters to improve the network condition and hence maintain a higher QoE. While traditional networks are not aware of the application, in our proposal the controller learns the configuration of the access points (APs) (in terms of transmission power and channel number) that provide the best QoE, using double deep Q-Learning (DDQL). The proposal improves video QoE by 91 % in the best case, when compared to three baselines. It also balances the QoE among clients, improving the fairness up to 115% when compared to the baselines.

Cloud Gaming renders game data in the cloud and forwards it to players over the network. While this reduces hardware costs for players, it introduces challenges in network management and delivering a good gaming experience. In this context, network providers are encouraged to implement QoE-aware management systems to guarantee a desired Quality of Experience (QoE), in which Machine Learning (ML) models achieve the state-of-the-art on QoE estimation/monitoring. However, it is hard to create ML models that generalize to different contexts, especially since QoE perception is subjective and varies among games and players. This paper employs transfer learning and fine-tuning to adjust a source model to different target domains. First, we performed a subjective QoE assessment with real users playing on a realistic testbed. Based on this, we derived four datasets, one being the source dataset (to create the source model) and three distinct target datasets. Experiments show that transfer learning can decrease the average MSE error by at least 41.6% compared to the source model performance on the target datasets while decreasing the demand for labeled data by at least 81.1%. Furthermore, the improvement is greater when compared to models trained from scratch for each target dataset.

Cloud gaming is a new paradigm that allows more cost-effective gaming for both users and game developers. The market is expected to grow 50-60% annually, reaching 22 billion USD by 2030. Gaming providers and ISPs require models of user satisfaction in order to improve their management of the cloud and network infrastructure. This paper analyses and proposes models that estimate the QoE of cloud gaming. Such models take as features network and game metrics. We assume an information sharing agreement among the cloud gaming platform and the ISP, allowing for a richer dataset. Data collection is performed with real users playing on a realistic testbed using similar protocols of the NVIDIA Geforce Now cloud gaming platform. We use stacking learning in order to improve the accuracy of the models, making a search for the best models and stacking them. We tested various improvements to the models, such as removing users with very low number of matches. Experiments show that models with more experienced players obtained a better precision, achieving 36.08%. When considering a range of plus or minus one within the estimated precision, the hit ratio was 86.56%. We also analyzed the model’ s sensitivity to inputs using feature importance analysis.

Existing Cloud deployments usually perform automated scheduling and rescheduling based on Quality of Service (QoS) objectives. Services are migrating towards Quality of Experience (QoE), which maps the user experience more effectively than QoS. This work proposes extensions to the Kubernetes scheduler in order to employ QoE objectives into the algorithm. For that, we created deep learning models (using LSTM) to estimate user’s QoE that the cloud can offer. The evaluation was performed on a testbed, and considered two QoE-aware applications (live classroom and video on demand). Experimental results in a testbed show that our scheduler improves the average QoE by at least 61.5% compared to other schedulers, while our proposed resource rescheduling improved the QoE by up to 119%, keeping the average QoE closer to the maximum.

With the increasing traffic of Video on Demand (VoD), network providers are seeking to deliver high Quality of Experience (QoE) for their users. Many methods have been proposed to assess VoD-related QoE. Some of them rely on client instrumentation and reporting QoE information to network elements, such as Server and Network Assisted DASH, others are based on statistical methods that make QoE inferences using monitored network conditions, such as throughput and delays. In this article, we present a practical method to estimate QoE for VoD using the widely supported Internet Control Message Protocol (ICMP) probes. Measured network conditions are used as input to a Machine Learning (ML) model that estimates QoE in terms of Mean Opinion Score (MOS), based on the ITU-T P.1203 Recommendation. The estimation encompasses video quality switches and playback stalls. We estimate MOS with an average Root Mean Square Error (RMSE) of 1.05 for a catalog of 25 different videos, training a model with sessions of the shortest video, and evaluating the generalization to the full catalog. We performed experiments using a virtualized setup as well as in a Wide Area Network.

An increase of Video on Demand (VoD) consumption has occurred in recent years. Delivering high Quality of Experience (QoE) for users consuming VoD is crucial. Many methods were proposed to estimate QoE based on network metrics, or to obtain direct feedback from video players. Recent proposals usually require monitoring tools installed in multiple network nodes, instrumentation of client devices, updates on existing network elements, among others. We propose a method based on Internet Control Message Protocol (ICMP) probing to estimate QoE for users consuming VoD. The method allows network operators to estimate which QoE level can be delivered to the user according to current network conditions using a Machine Learning (ML) model. Our method does not require installation of software at different network nodes, relying on ICMP probing which is widely supported by existing devices. Our QoE inference model estimates Mean Opinion Score (MOS) with Root Mean Square Error (RMSE) of 0.98, with additional 27 Kbps of traffic during probing. We evaluate the model's generalization capacity when estimating QoE for videos different from the one used for training, which can speed up model's creation process. In those cases MOS was estimated with RMSE of 1.01.

Quality of Experience (QoE) should be the driver for network orchestration in 4G networks. At the same time, the network must be able to cope with high bandwidth requirements from applications such as video streaming, while dealing with a large number of users. This paper proposes a network orchestrator that adjusts network parameters to improve QoE of video streaming. The orchestrator uses Device-to-Device (D2D) communication to improve user's QoE, also reducing the demand on 4G network. The use of D2D is triggered by a machine learning engine. Experiments made in a physical testbed show an improvement on the mean horizontal video resolution from 768 to 1280 pixels, as well as a decrease of around 90% at the impact on the QoE, considering the number of video resolution changes. Finally, the demand on the network backhaul is decreased by around 38%.

Wireless networks are the most common way to access the Internet, with more than 10 billion Wi-Fi devices already sold. Wireless connections suffer from problems related to spectrum overuse, such as transmission errors and loss of information. Intelligent control systems can be used for network management and to improve Quality of Service (QoS). However, the first step towards such systems is a way to correlate current Wi-Fi readings to a QoS value. This work proposes a model using Recurrent Neural Networks (RNN) to infer such relation, based on real Wi-Fi data, and compares two RNN types: Gated Recurrent Unit (GRU) and Long Short Term Memory (LSTM). The model predicts four network metrics: throughput, loss, delay, and jitter, based only in traffic data obtained at the AP. The average Root Mean Square Error is of the order of 10-2for throughput, 10-4for delay, and 10-5for jitter and packet loss using both methods.

Communications using wireless networks nowadays receives an increasing amount of users and devices. This implies in an increasing need for improvements in wireless networks. IEEE 802.11ac standard enables channels wider than 20 MHz bandwidth. But currently, access points allocate channel bandwidth statically, independent of the clients' configurations. This leads to poor network performance. In this work, we design a system that enables clients to choose the link's bandwidth they connect to the APs. We validate the system through realistic experiments. We show the trade-off between latency and throughput. Our results show that our system improves the frequency spectrum usage, and provide better throughput/latency for the user.

Software Defined Networks (SDNs) offer flexibility to current networks, allowing operators to manage network elements using software on an external server. SDNs are founded on a key feature: the separation of the control plane from the data plane. OpenFlow is the most popular SDN southbound interface today. However, does OpenFlow really decouple the data plane from the control plane? This is the leading question in this work. The literature has sought to quantity the impact of OpenFlow commands from control plane on data plane performance. Particularly, we argue that it is possible to damage the date plane by too many flow updates. Attackers, for instance, can use this effect in a cloud environment to reduce the performance of a collocated virtual network. However, it is not clear what is the exact impact of this coupling on production hardware and software switches. We investigate this through experiments, under representative scenarios, and propose a threshold mechanism to mitigate the effect of malicious administrators. We have observed that both hardware and software switches suffer from this limitation, presenting an average RTT degradation of up to 12.35% in the hardware switch, and 25.9% on the software switch. Finally, the proposed mechanism mitigates the lack of decoupling and malicious behavior.

The number of smart devices in home networks is rapidly increasing, making it more complex to manage their faults. In addition, the lack of customer knowledge and tools to automatically diagnose and fix faults aggravate the problem. In this paper, we propose HomeNetRescue, a Software-Defined Network (SDN) service for autonomous management of wireless and wired home networks focused on fault and configuration management. We evaluate HomeNetRescue in a real world prototype, considering throughput, delay, and jitter. Our results show that HomeNetRescue can increase the throughput of the network by up to 131%, reducing wireless transmission delay and jitter by 46% and 24%, respectively.

It is estimated that Internet traffic will triple in five years, which will increase server response time. One way to reduce such time is to balance the load on replicated servers using SDN switches. However, existing SDN switches have performance limitations that influence the performance of the load balacing. This work compares five load balancing policies: round robin, random, txbytes (transmitted bytes), cpuq-load (CPU usage and number of open connections), and load/load-prev (load forecast with switch statistics). These policies consider limitations such as the cost to retrieve network statistics and to install new rules. The results show that the txbytes and cpuq-load policies outperformed the others. On the other hand, the load-prev policy proved to be promising when adjusted for traffic.

Wireless networks are very dynamic, having a variety of applications with different requirements. This diversity demands more flexible equipment as well as networks that adapt to the context of the applications. This work proposes FS-MAC, a platform that allows more than one MAC protocol to be used on the network. FS-MAC activates each of the MAC protocols when they are most effective. FS-MAC is also extensible, allowing the addition of new protocols. The proposal was tested in a testbed, where we varied the load and the number of connected stations. Results show that FS-MAC has throughput and delay values that are comparable to the best static protocol, having an overhead of around 2%.

Mobile devices are equipped with limited processing power and battery charge. A mobile computation offloading framework is a software that provides better user experience in terms of computation time and energy consumption, also taking profit from edge computing facilities. This article presents User-Level Online Offloading Framework (ULOOF), a lightweight and efficient framework for mobile computation offloading. ULOOF is equipped with a decision engine that minimizes remote execution overhead, while not requiring any modification in the device’s operating system. By means of real experiments with Android systems and simulations using large-scale data from a major cellular network provider, we show that ULOOF can offload up to 73 percent of computations, and improve the execution time by 50 percent while at the same time significantly reducing the energy consumption of mobile devices.

The decoupling of data and control planes of network switches is the main characteristic of Software Defined Networks. The OpenFlow (OF) protocol implements this concept and it is found today in various off-the-shelf equipment. Despite being widely employed in industry and research there is no systematic evaluation of OF data plane performance in the literature. In this paper we evaluate the performance and maturity of the main features of OF 1.0 on nine hardware and software switches. Results show that the performance varies significantly among implementations. For instance, packet delays vary by one order of magnitude among the evaluated switches, while the packet size does not impact the performance of OF switches.

An ever-growing number of embedded devices supports different kinds of applications, such as healthcare, surveillance, gas monitoring, and others, that require an elevated level of communication reliability. However, the expected high density of those embedded devices increases the competition for frequency spectrum, making it difficult to achieve a reliable machine-to-machine (M2M) communication. To overcome these difficulties, the use of link quality estimators (LQE) is crucial to provide a solid communication. In order to provide robust and faster communication under harsh conditions, this paper proposes a new LQE, called PRR2, which uses two metrics and two levels of PRR (Packet Received Ratio). The use of two PRR sliding windows captures link quality variations in the short term and also considers the long-term. PRR2is compared against the state of the art on a prototype using USRPs, and the results show that the proposal reduces the number of retransmissions and increases the delivery rate, which are two important metrics for link layer reliability.

Vehicular Ad Hoc Networks (VANETs) are mobile networks that extend over vast areas and have intense node mobility. These characteristics lead to frequent delays and disruptions. A solution is to employ the Delay Tolerant Network (DTN) paradigm. However, the frequent disruptions as well as the delay and reliability constraints of certain VANET applications hinder the employment of both conventional and DTN-based management architectures. This paper tackles monitoring, one of the tasks of network management. We describe a hierarchical architecture that copes with near real-time as well as non real-time monitoring tasks. The proposed solution is evaluated using simulations, where we measure the delay and delivery rates of the monitoring data. The results show that the proposed solution reduces the delivery delay and increases the chances that a notification will be delivered on time to its destination.

Wireless Sensor Networks (WSNs) employ Industrial, Scientific and Medical (ISM) spectrum bands for communication, which are overloaded due to various technologies such as WLANs and other WSNs. Therefore, such networks must employ intelligent methods such as Cognitive Radio (CR) to coexist with other networks. This study investigates the use of supervised Machine Learning (ML) for channel selection in WSNs. The proposed models were analyzed using ML tools and techniques, and the best algorithms were evaluated on real sensor nodes. The experiments show performance improvements on the delivery rate and delivery delay when the proposed cognitive solutions are employed.

The use of mobile devices for work and for entertainment is growing. However, some applications may consume too much battery or be too CPU intensive for those devices. One alternative is to dynamically offload parts of the computation of those applications to the Cloud, reducing the battery usage of the devices and improve the responsiveness of the applications. One of the biggest challenges in offloading is to decide whether offloading a certain computation should be performed or not, since the time and energy required to upload the code as well as the input parameters may be less advantageous than executing it locally. This paper proposes a location aware decision engine for mobile offloading. The proposed decision engine can be easily plugged into any offloading framework. Further, we propose a black-box approach to estimate a method's execution time using simple code annotations. An experimental evaluation indicates that location-aware offloading can reduce the energy usage by 50%, with a CPU overhead of approximately 10%.

Wireless Networks have become ubiquitous and dense to support the growing demand from mobile users. To improve the performance of these networks, new approaches are required, such as context and service aware control algorithms, which are not possible on today's closed proprietary WLAN controllers. In this work, we propose Ethanol, a software-defined networking architecture for 802.11 dense WLANs. This paper describes the benefits of programmable APs, and proposes Ethanol, an architecture for network-wide control of QoS, user mobility, AP virtualization, and security on 802.11 APs. The proposal is evaluated on a prototype using off-the-shelf APs over three use cases.

In the recent years, the research community proposed several protocols and applications for opportunistic networking. A common assumption is that all nodes have pro social behavior and are willing to cooperate with the network. However, in opportunistic networking applications, this assumption can lead to degradations in the network performance. People can be selfish and this behavior affects the operation of the network. In this work, we propose an incentive mechanism to improve routing, called MINEIRO, which aims to detect and avoid selfish nodes based on the source of the messages. We demonstrate under which constraints our algorithm leads to Bayesian equilibrium. Moreover, we show that without an incentive mechanism the network supports up to 60% of nodes with selfish behavior without performance degradation in a random mobility scenario. Meanwhile, in a scenario with social-based mobility, the performance decreases linearly for more than 20% of selfish nodes. Our proposal, on the other hand, improves the performance with any amount of selfish nodes by encouraging users to relay messages from third-parties.

With the increasing number of users subscribing to mobile Internet such as 3G and 4G networks, Wireless Internet Services providers (WISP) aim to provide a good service for customers while elevating the number of clients. Several proposals to offload the traffic of 3G networks were made in the last few years, including the use of femtocells, WiFi offloading and more recently mobile-to-mobile opportunistic offloading. In this paper, we propose a multi-criteria decision-making framework to manage the offload of data from 3G networks using mobile-to-mobile opportunistic communications. Primarily, we focus on building a decision framework that employs only user knowledge to select which users should handover from infrastructure to mobile-to-mobile network, avoiding changes in the infrastructure. Next, we evaluate our proposal and demonstrate its feasibility through trace-driven simulations, achieving 6% of data offload when there is no delay tolerance in the application and up to 36% when application can tolerate 20 minutes of delay.

Vehicular Disruption Tolerant Networks appear due to the search of information by drivers, which build mobile ad hoc networks that may suffer from long interruptions. This paper proposes a certification mechanism by means of social networks, enabling cars to share keys through direct contacts between two acquaintances that warrant their identity, so they sign a reciprocal certificate. Certificates signed by a third party can be validated if the user's public key is available to the other party. Further, the reputation mechanism can identify certificates of trusted users. The evaluation shows the behavior of a vehicular network that uses certificates in social networks as cryptographic security mechanism to establish a degree of trust between users.

The growth of mobile computing has increased the demand for wireless communication, causing a higher demand for the wireless medium as well as spectrum pollution. Smart radios, also called cognitive radios, monitor the network to identify the best available channel, in order to avoid interference. This paper proposes a framework for the development and testing of protocols for wireless sensor networks that employ cognitive radios (CRSN). We also developed two spectrum decision protocols for CRSN, which provide distributed mechanisms to select the best wireless channel based on the application's QoS requirements. Simulations of low, medium and high noise scenarios have shown that the protocols improve the delivery rate by up to 69%, while keeping the delay and energy consumption unaltered.

Several applications of Wireless Sensor Networks (WSN) demand information only from sub-regions of the monitored area. The user establishes the region of interest (RoI) and requires the WSN to collect data only inside this region. This kind of query is called spatial query. The state of the art of spatial query processing considers, in general, that nodes do not fail. However, nodes can be destroyed, interference can deny the communication and nodes can go to sleep mode (turn off the radio in duty cycles) in order to save energy. This work proposes a fault-tolerant energy efficient in-network spatial query processing mechanism for WSN. The proposed mechanism employs failure detection algorithms in order to avoid failed neighbors when forwarding and processing spatial queries. Our contributions are lightweight and resilient because they operate on demand and avoid the use of routing tables.

The extreme dynamics of wireless multi-hop networks (WMN) - those that use multiple wireless links to forward data - require self-configuration solutions in every communication layer. Self-configuration demands mechanisms for information sharing, as well as specific decision-making algorithms for each service or application. Self-configuration must also be embedded in the communication layers in order to cope with changes in the wireless environment. This article is a digest of the three contributions of, which investigated the self-configuration tasks of programmability, sharing of information and specialized control loops for WMN. First, we designed a middleware, based on the Information Plane (InP) concept, to ease the programmability and sharing of information among nodes. Next, we proposed two InP-based control loops for WMN: one for data rate and transmission power aware routing, and another for mobile context-aware peer-to-peer services.

Due to the emergence of multimedia context-rich applications and services over wireless networks, networking protocols and services are becoming more and more integrated, thus relying on context and application information to support their operation. Further, wireless protocols and services now employ information from several network layers and the environment, breaking the layering paradigm. In order to cope with this increasing reliance on information, we have proposed MANIP, a middleware for MANETs that instantiates a new networking plane. The Information Plane (InP) is a distributed entity to store and disseminate information concerning the network, its services and the environment, orchestrating the collaboration among cross-layer protocols, autonomic management solutions and context-aware services. We use MANIP to support the autonomic reconfiguration of a P2P network over MANETs. Simulation results show that the MANIP-enabled solutions reduce the response time and increase the number of solved P2P queries when compared to classic, cross-layer implementations of the same protocols.

In mobile wireless ad hoc networks (MANETs), layering is frequently broken to cope with changes on the medium. Layering is also violated in order to implement autonomic behavior, which depends on correlating data from various layers to identify relevant events. This paper proposes a networking plane, called MANET knowledge plane (MANKOP), that stores information concerning all protocol layers. This plane improves network performance, as protocols may employ a broader range of inputs on their algorithms. Moreover, being an information repository, MANKOP eases the deployment of self-optimizing and self-configuring mechanisms. To showcase our solution, we use MANKOP to build transmission power control aware protocols, which increases the throughput of the network by up to 280%.

Wireless sensor networks are ad hoc networks with severe resource constraints. These constraints preclude the use of traditional ad hoc protocols, and demand optimizations that incur in solutions specific to a class of applications. This article presents PROC, a protocol designed for continuous data dissemination networks, that interacts with the application to establish routes, allowing the application to reconfigure PROC on runtime. A performance evaluation in topologies varying from 50 to 200 nodes showed that PROC increases network lifetime around 7% to 12%, and has higher throughput than EAD and TinyOS Beaconing. Furthermore, PROC presents a softer performance degradation when the number of nodes in the network increases.

Medium access control (MAC) protocols manage energy consumption on the network element during communication, which is the most energy-consuming event on Wireless Sensor Networks (WSNs). One method to mitigate energy consumption is to adjust transmission power. This paper presents two approaches to adjust transmission power in WSNs. The first approach employs dynamic adjustments by exchange of information among nodes, and the second one calculates the ideal transmission power according to signal attenuation in the link. The proposed algorithms were implemented and evaluated with experiments, comparing their results with B-MAC, the standard MAC protocol in the Mica Motes 2 platform. Results show that transmission power control is an effective method to decrease energy consumption, and incurs in a negligible loss in packet delivery rates. For node distances of 5m, the proposed transmission power control techniques decrease energy consumption by 27% over B-MAC.

Both mobile ad hoc networks (MANETs) and peer-to-peer (P2P) networks are decentralized and self-organizing networks with dynamic topology and are responsible for routing queries in a distributed environment. Because MANETs are composed of resource-constrained devices susceptible to faults, whereas P2P networks are fault-tolerant, P2P networks are the ideal data sharing system for MANETs. We have conducted an evaluation of two approaches for P2P content discovery running over a MANET. The first, based on unstructured P2P networks, relies on controlled flooding, while the second, based on structured P2P networks, uses distributed indexing to optimize searches. We use simulations to evaluate the effect of network size, mobility, channel error rates, network workload, and application dynamics in the performance of P2P protocols over MANETs. Results show that unstructured protocols are the most resilient, although at higher energy and delay costs. Structured protocols, conversely, consume less energy and are more appropriate for MANETs where topology is mostly static.