Rapid adoption of machine learning (ML) technologies has led to a surge in power consumption across diverse systems, from tiny IoT devices to massive datacenter clusters. Benchmarking the energy efficiency of these systems is crucial for optimization, but presents novel challenges due to the variety of hardware platforms, workload characteristics, and system-level interactions. This paper introduces MLPerf® Power, a comprehensive benchmarking methodology with capabilities to evaluate the energy efficiency of ML systems at power levels ranging from microwatts to megawatts. Developed by a consortium of industry professionals from more than 20 organizations, coupled with insights from academia, MLPerf Power establishes rules and best practices to ensure comparability across diverse architectures. We use representative workloads from the MLPerf benchmark suite to collect $\mathbf{1, 8 4 1}$ reproducible measurements from 60 systems across the entire range of ML deployment scales. Our analysis reveals trade-offs between performance, complexity, and energy efficiency across this wide range of systems, providing actionable insights for designing optimized ML solutions from the smallest edge devices to the largest cloud infrastructures. This work emphasizes the importance of energy efficiency as a key metric in the evaluation and comparison of the ML system, laying the foundation for future research in this critical area. We discuss the implications for developing sustainable AI solutions and standardizing energy efficiency benchmarking for ML systems.

Modern protocols for wireless sensor networks efficiently support multi-hop upward traffic from many sensors to a collection point, a key functionality enabling monitoring applications. However, the ever-evolving scenarios involving low-power wireless devices increasingly require support also for downward traffic, e.g., enabling a controller to issue actuation commands based on the monitored data. The IETF Routing Protocol for Low-power and Lossy Networks (RPL) is among the few tackling both traffic patterns. Unfortunately, its support for downward traffic is significantly unreliable and inefficient compared to its upward counterpart. We tackle this problem by extending RPL with mechanisms inspired by opposed, yet complementary, principles. At one extreme, we retain the route-based operation of RPL and devise techniques allowed by the standard but commonly neglected by popular implementations. At the other extreme, we rely on flooding as the main networking primitive. Inspired by these principles, we define three base mechanisms, integrate them in a popular RPL implementation, analyze their individual and combined performance, and elicit the resulting tradeoffs in scalability, reliability, and energy consumption. The evaluation relies on simulation, using both real-world topologies from a smart city scenario and synthetic grid ones, as well as on testbed experiments validating our findings from simulation. Results show that the combination of all three mechanisms into a novel protocol, T-RPL (i) yields high reliability, close to the one of flooding, (ii) with a low energy consumption, similar to route-based approaches, and (iii) improves remarkably the scalability of RPL with respect to downward traffic.

The Internet Of Things (IoT) is an emerging paradigm that envisions a networked infrastructure enabling different types of devices to be interconnected. It creates different kinds of artifacts (e.g., services and applications) in various application domains such as health monitoring, sports monitoring, animal monitoring, enhanced retail services, and smart homes. Recommendation technologies can help to more easily identify relevant artifacts and thus will become one of the key technologies in future IoT solutions. In this article, we provide an overview of existing applications of recommendation technologies in the IoT context and present new recommendation techniques on the basis of real-world IoT scenarios.

Fog computing enables use cases where data produced in end devices are stored, processed, and acted on directly at the edges of the network, yet computation can be offloaded to more powerful instances through the edge to cloud continuum. Such offloading mechanism is especially needed in case of modern multi-purpose IoT gateways, where both demand and operation conditions can vary largely between deployments. To facilitate the development and operations of gateways, we implement offloading directly as part of the IoT rapid prototyping process embedded in the software stack, based on Node-RED. We evaluate the implemented method using an image processing example, and compare various offloading strategies based on resource consumption and other system metrics, highlighting the differences in handling demand and service levels reached.

Stringent latency requirements in advanced Internet of Things (IoT) applications as well as an increased load on cloud data centers have prompted a move towards a more decentralized approach, bringing storage and processing of IoT data closer to the end-devices through the deployment of multi-purpose IoT gateways. However, the resource constrained nature and diversity of these gateways pose a challenge in developing applications that can be deployed widely. This challenge can be overcome with containerization, a form of lightweight virtualization, bringing support for a wide range of hardware architectures and operating system agnostic deployment of applications on IoT gateways. This paper discusses the architectural aspects of containerization, and studies the suitability of available containerization tools for multi-container deployment in the context of IoT gateways. We present containerization in the context of AGILE, a multi-container and micro-service based open source framework for IoT gateways, developed as part of a Horizon 2020 project. Our study of containerized services to perform common gateway functions like device discovery, data management and cloud integration among others, reveal the advantages of having a containerized environment for IoT gateways with regard to use of base image hierarchies and image layering for in-container and cross-container performance optimizations. We illustrate these results in a set of benchmark experiments in this paper.

RPL, the IPv6 Routing Protocol for Low-Power and Lossy Networks, supports both upward and downward traffic. The latter is fundamental for actuation, for queries, and for any bidirectional protocol such as TCP, yet its support is compromised by memory limitation in the nodes. In RPL storing mode, nodes store routing entries for each destination in their sub-graph, limiting the size of the network, and often leading to unreachable nodes and protocol failures. We propose here D-RPL, a mechanism that overcomes the scalability limitation by mending storing mode forwarding with multicast-based dissemination. Our modification has minimal impact on code size and memory usage. D-RPL is activated only when memory limits are reached, and affects only the portion of the traffic and the segments of the network that have exceeded memory limits. We evaluate our solution using Cooja emulation over different synthetic topologies, showing a six-fold improvement in scalability.

Summary This paper describes an approach to perform reproducible performance tests on virtual routers, comparing different virtual routing architectures, different software versions and configurations. The presented approach is based on VRKit, a software tool that allows to build pre‐configured bootable virtual router images with the desired characteristics. Design decisions are described, and some usage examples are presented, showing how the tool can be used for research on virtual routing, facilitating performance comparison of various virtual router implementations. Copyright © 2013 John Wiley & Sons, Ltd.

Peer-to-peer live-streaming (P2P-TV) systems' goal is disseminating real-time video content using peer-to-peer technology. Their performance is driven by the overlay topology, i.e., the virtual topology that peers use to exchange video chunks. Several proposals have been made in the past to optimize it, yet few experimental studies have corroborated results. The aim of this paper is to provide a comprehensive experimental comparison based on PeerStreamer in order to benchmark different strategies for the construction and maintenance of the overlay topology in P2P-TV systems. We present only experimental results in which fully distributed strategies are evaluated in both controlled experiments and the Internet using thousands of peers. Results confirm that the topological properties of the overlay have a deep impact on both user quality of experience and network load. Strategies based solely on random peer selection are greatly outperformed by smart yet simple and actually implementable strategies. The most performing strategy we devise guarantees to deliver almost all chunks to all peers with a playout delay as low as 6 s even when system load approaches 1, and in almost adversarial network scenarios. PeerStreamer is open-source to make results reproducible and allow further research by the community.

This paper shows how to use an open virtualisation architecture to analyse and improve the forwarding performance of a virtual router. In particular, the forwarding performance of the Linux kernel running inside a KVM virtual machine and the performance of some more advanced architectures based on virtual routers aggregation are analysed, showing how increasing the number of used CPU core can improve performance and how properly setting the CPU affinity of the various virtualisation activities affects virtual router throughput.

Whether used for file distribution, for video streaming, or for other services, every large scale peer-to-peer distributed system should build and maintain a topology. Apart from cases where neighborhood relations are handled by a centralized tracker, peers need information about the existence and routable ID of other peers in order to build their neighborhood.

P2P-TV systems performance are driven by the overlay topology that peers form. Several proposals have been made in the past to optimize it, yet little experimental studies have corroborated results. The aim of this work is to provide a comprehensive experimental comparison of different strategies for the construction and maintenance of the overlay topology in P2P-TV systems. To this goal, we have implemented different fully-distributed strategies in a P2P-TV application, called Peer-Streamer, that we use to run extensive experimental campaigns in a completely controlled set-up which involves thousands of peers, spanning very different networking scenarios. Results show that the topological properties of the overlay have a deep impact on both user quality of experience and network load. Strategies based solely on random peer selection are greatly outperformed by smart, yet simple strategies that can be implemented with negligible overhead. Even with different and complex scenarios, the neighborhood filtering strategy we devised as most performing guarantees to deliver almost all chunks to all peers with a play-out delay as low as only 6s even with system loads close to 1.0. Results are confirmed by running experiments on PlanetLab. PeerStreamer is open-source to make results reproducible and allow further research by the community.

In this paper we consider mesh based P2P streaming systems focusing on the problem of regulating peer upload rate to match the system demand while not overloading each peer upload link capacity. We propose Hose Rate Control (HRC), a novel scheme to control the speed at which peers offer chunks to other peers, ultimately controlling peer uplink capacity utilization. This is of critical importance for heterogeneous scenarios like the one faced in the Internet, where peer upload capacity is unknown and varies widely. HRC nicely adapts to the actual peer available upload bandwidth and system demand, so that users' Quality of Experience is greatly enhanced. Both simulations and actual experiments involving up to 1000 peers are presented to assess performance in real scenarios. Results show that HRC consistently outperforms the Quality of Experience achieved by non-adaptive schemes.

Splitting a P2P video distribution in multiple media flows with different priorities is an interesting approach for developing flexible and adaptive streaming systems, ranging from VoD to TV. Such an approach can both yield satisfactory quality to all end users and be light in network resources usage, because low-priority flows can be discarded a-priori when target peers do not have enough resources to receive them. This paper focuses on chunk-based video distribution in unstructured meshes, adopting a push strategy (the sender takes the scheduling decision) based on buffer map exchange to avoid sending duplicated chunks. A deadline-based scheduling algorithm is proposed, where different flows of chunks are prioritized using different deadline postponing parameters for each flow. Some experimental results show good differentiation properties and streaming performance much better than with strict priority enforcement. Also, PSNR measures on real video streams show improvements compared to both strict priority and single stream distribution.

P2P TV and video streaming are among the most bandwidth-hungry applications running over the Internet. One of the main reasons is that the scheduling of information transfer between peers is extremely aggressive and does not take network characteristics into account. Moreover, schedulers are not designed to be robust and configurable, so that their performance is greatly affected by networking conditions. This work first analyzes the impact of network heterogeneity on the streaming performance and then proposes a novel, robust, configurable, network-aware scheduler that outperforms the other schedulers in all networking scenarios.

P2P TV distribution is going commercial, and the video quality delivered to users becomes of the utmost importance. However, the impact of P2P distribution on the video quality is not completely understood yet, especially in live streaming situations. This work addresses the impact of P2P distribution when the delay of the playout is limited, as it must be in any true live TV service. A methodology for the evaluation (using standard objective video quality metrics) is proposed, showing that different ways of grouping frames in chunks for the distribution can lead to very different quality when the system is overloaded.

P2P-TV systems have become part of the Internet landscape (See for instance http://www.pplive.com, http://www.soapcast.com, http://www.tvants.com, and many others). The architecture of these (normally proprietary) applications is generally receiver-driven, in that receivers actively search for suitable peers to download from, trying to maximize their performance. This results in very aggressive applications that generate huge and non optimized traffic loads. The demo summarized in this short paper shows the impact of various P2P streaming options and the efficiency of the Napa-Wine approach (compared to more traditional approaches) by running real streaming clients in realistic conditions. To make this comparison possible, the software developed in Napa-Wine is highly modular and configurable, allowing the user to test different topology management and chunk trading techniques developed within the Napa-Wine project, as well as to configure it to mimic other chunk/peer selection strategies known from literature.

Protecting users' privacy is becoming one of the rising issues for the success of future communications. The Internet in particular, with its open architecture, presents several threats to the right of protecting personal and sensitive data. One fundamental building block of privacy-respectful communications is protecting the communication parties identities, or, as it is commonly called within the research community anonymous networks (ANs). An AN prevents external observers as well as the network to have access to communicating partners identities and addresses. In this paper we propose a novel architecture to realize ANs, as an extension to IPsec. After explaining the rationale and discussing possible alternatives, we present a working prototype implementation and its experimental performance comparison with application level solutions.