The substantial memory and computational requirements of large language models (LLMs) hinder their deployment on individual resource-constrained devices. This paper introduces HyperCluster, a framework for fully decentralized collaborative inference over peer-to-peer wireless networks. HyperCluster presents three core innovations: (1) a ring-based pipelined inference protocol where nodes deterministically self-organize into a computational ring based on device capabilities and pass intermediate states directly between peers via QUIC-based direct transport; (2) a generalizable model sharding methodology built on top of the Hugging Face Transformers library that automatically partitions any dense LLM across heterogeneous devices according to available memory; (3) selective layer loading from safetensors files, which only loads the tensor weights required by each node’s assigned shard. We validate HyperCluster on a heterogeneous cluster of consumer-grade devices, demonstrating distributed inference of models up to 3 billion parameters with comprehensive latency and throughput analysis across one to three node configurations.
The amount of data generated each year continues to increase. Some of this data must be preserved, and this becomes a significant issue when critical data are involved. Data storage systems must protect against risks that affect both data and infrastructure (primarily the storage media). To fulfil this requirement, the system should offer both data versioning (to restore previous safe versions) and replication (to mitigate device failures). These two functions should be dimensioned based on risk analysis, which may, however, result in complex software/hardware architectures. In this paper, we propose a simple graph model that accounts for risks to both data and devices and facilitates the design of a data storage system. It is complemented by a matrix model that allows analysis of specific properties. The proposed models enable the detection of suboptimal architectures and facilitate understanding of the operations of the backup system. In addition, this study can also support future work on risk management in backup systems.
Cyber-physical systems (CPS) operate across a computing continuum of heterogeneous devices with varying support for execution formats such as containers, Wasm, and native binaries. While Kubernetes is the de facto orchestration standard, its container-centric model and operational overhead make it unsuitable for resource-constrained embedded devices in CPS deployments. We present an adaptive orchestration system that treats format heterogeneity as a first-class concern, enabling distributed deployment across heterogeneous CPS environments. The system selects and places components based on device capabilities and user-specified non-functional requirements (NFRs). Through a MAPE-K control loop, local device agents continuously monitor and report constraints to a central orchestrator. Upon detecting a constraint violation, the orchestrator directs the agents to locally adapt through execution format transformations, redeployments, or component suspensions. Evaluation on a distributed image processing pipeline of five microservices demonstrates deployment initialization in < 5 min, rapid execution format transformation in < 3 s, and a stable agent memory footprint of 30–40 MB even under active load. These results establish that dynamic, NFR-driven heterogeneous orchestration can be achieved with minimal overhead in resource-constrained CPS environments.
While the Microkernel paradigm is a cornerstone for achieving modularity and extensibility, its application in modern distributed environments is often hampered by a rigid dependency on specific execution contexts. Existing frameworks often entangle business logic with specific communication technologies, necessitating extensive refactoring when transitioning between monolithic and distributed deployments. In this work, we introduce Orbitalis, an open-source framework that extends the Microkernel paradigm to distributed systems. Orbitalis integrates component development, communication, and lifecycle management, enabling seamless execution across both local and distributed environments, enabling applications to operate seamlessly across heterogeneous nodes without requiring a priori knowledge of the deployment topology. At the heart of the framework lies Busline, an asynchronous, protocol-agnostic publish/subscribe backbone that supports dynamic component evolution, replication, and replacement. A DHCP-inspired discovery protocol enables cores and plugins to negotiate, connect, and reconfigure at runtime. Experimental evaluation quantifies the framework’s overhead across various deployment scenarios, demonstrating its practical viability.
Large language models (LLMs) achieve remarkable performance across domains but remain prone to hallucinations and inconsistencies. Retrieval-augmented generation (RAG) mitigates these issues by augmenting model inputs with relevant documents retrieved from external sources. In many real-world scenarios, relevant knowledge is fragmented across organizations or institutions, motivating the need for federated search mechanisms that can aggregate results from heterogeneous data sources without centralizing the data. We introduce RAGRoute, a lightweight routing mechanism for federated search in RAG systems that dynamically selects relevant data sources at query time using a neural classifier, avoiding indiscriminate querying. This selective routing reduces communication overhead and end-to-end latency while preserving retrieval quality, achieving up to 80.65
Graph-based recommendation models effectively capture high-order collaborative signals from user–item interaction graphs. Federated learning (FL) enables privacy-preserving training across distributed clients. However, directly aggregating graph representations under FL is challenging: locally learned structural embeddings are not globally aligned under non-IID data distributions, and naive parameter averaging fails to recover cross-client relational structure. Existing federated graph-based approaches primarily rely on structural aggregation, yet overlook the global semantic knowledge encoded in large language models (LLMs). In this work, we propose a semantic–structural federated graph recommendation framework that leverages LLM embeddings to guide cross-client alignment. Each client learns user representations from its local interaction graph and summarizes typical interaction patterns into compact semantic vectors using a frozen LLM encoder. These vectors are sent to the server, which identifies semantically related patterns across different clients and combines their structural representations accordingly. The updated representations are then returned to clients to refine subsequent local training. This design enables collaboration guided by shared semantic understanding without exposing raw interaction data, preserving both recommendation accuracy and privacy. Experiments on benchmark datasets demonstrate consistent improvements over existing federated graph-based baselines.
Byzantine fault-tolerant protocols enable applications to run even in the presence of arbitrary faults. Unfortunately, to reach a consensus on the order to execute client requests, these protocols perform costly cryptographic operations that limit their performance. This hinders their adoption for latency-sensitive applications executed in data-centres, such as fault-tolerant machine learning or finance applications. We propose two mechanisms to reduce the impact of digital signature verification: Speculative Execution to decouple consensus from the verification of the digital signature of protocol messages; and Secure Request Broadcast where replicas do not need to wait for the primary before they can verify a request signature. Both mechanisms are combined with a Blocklisting mechanism to ensure robustness against attacks. We implement both mechanisms in SPECULOR, a PBFT-based BFT protocol tailored for a relaxed fault model where clients do not crash and periodically retransmit their requests. Our evaluation shows SPECULOR reduces the tail-latency by at least 34
The widespread use of mobile applications in daily life has raised significant concerns about their energy consumption, as it directly affects user satisfaction, their hosting device longevity, and environmental sustainability. Despite the need for energy-aware mobile application development, developers lack cost-effective, software-based, and intuitive tools to measure and analyze energy usage during the development process. This paper introduces BCProf (Battery Consumption Profiler), a novel framework designed to measure the energy consumption of Android applications at both method and application granularities. BCProf supports both Java and Kotlin languages and is implemented as a plugin for seamless integration into Android Studio. BCProf works by instrumenting the application’s source code with custom log statements that enable tracking Android API calls. During application testing, BCProf monitors the generated logs alongside hardware component usage (e.g., WiFi, GPS, camera, Bluetooth, and screen) to calculate energy consumption. By leveraging this monitored data, BCProf correlates the energy usage of Android API calls and hardware components with specific methods in the application’s source code. The tool supports both online, real-time energy analysis during application execution and offline analysis by saving energy usage data in a JSON file format. Extensive evaluations were conducted across various Android applications to validate the accuracy and versatility of BCProf. The results show that BCProf can achieve up to 98
The convergence of Artificial Intelligence (AI) and the Internet of Things (IoT) has paved the way for the Internet of Robotic Things (IoRT), where autonomous robotic systems leverage AI capabilities to operate seamlessly in interconnected ecosystems. However, traditional IoRT architectures reliant on cloud computing face critical challenges, including data privacy risks, latency, and bandwidth limitations. This paper proposes a novel AI-driven distributed system, AI õ RT, that utilizes edge computing to distribute computational resources closer to robotic devices, enabling reduced response times and enhanced data security. The system employs layered architecture integrating edge AI, heterogeneous edge-split federated learning, and FPGA-accelerated binary compressive sensing (BinCSNet) for efficient model training and data processing. This design optimizes resource usage, supports real-time decision-making, and facilitates scalable deployment across heterogeneous networks. The system also demonstrated its scalability at the robot service level through humanoid robot-based application cases, gait emotion recognition and human depression detection, highlighting its potential to advance autonomous systems in sustainable and secure environments.
This paper introduces recent advances in EnOSlib, a Python library that aims at facilitating the design and execution of reproducible experiments across distributed computing infrastructures. Originally developed to simplify experimentation on testbeds such as Grid'5000 and Chameleon Cloud, EnOSlib now incorporates support multi-provider deployments, including access to edge resources, as well as advanced services. Key contributions include integration with Kwollect for fine-grained energy measurements, a planning service for executing timed events, and enhanced network emulation functionalities. These features enable users to model and study complex, realistic scenarios such as latency-sensitive edge-to-cloud applications. A major new capability is the support for synchronized multiinfrastructure experiments, allowing simultaneous resource reservation and deployment across diverse testbeds. The paper illustrates these capabilities through a distributed video processing use case spanning edge and cloud platforms. This paper is the companion paper of the tutorial presented in DAIS 2025.
Air pollution remains a critical environmental and public health challenge in urban areas, which requires accurate and efficient predictive models to mitigate its impact. This study introduces a novel spatiotemporal model, Graph Temporal LSTM (GT-LSTM), which integrates Machine Learning (ML) techniques to forecast air pollution levels, with a focus on Particulate Matter (PM) 2.5 concentrations. By combining Graph Convolutional Network (GCN) to capture spatial dependencies and Long Short-Term Memory (LSTM) to model temporal patterns, the proposed framework provides precise and localized predictions in urban and suburban regions. Our analysis demonstrates the competitive predictive capabilities of the model, achieving high Coefficient of determination (R2) and low error values, highlighting its robustness in correlating predicted and observed pollutant levels. The GT-LSTM model effectively incorporates historical data, neighboring influences, and local pollution sources, allowing reliable short- and long-term forecasts, even in data-short environments. In addition to its predictive accuracy, the model prioritizes computational efficiency and scalability, using cost-effective sensor networks to expand coverage and reduce the dependence on traditional data sources. By offering fine-grained insights into air quality patterns, this approach supports real-time monitoring, long-term planning, and proactive decision-making, benefiting policymakers and urban residents alike. This study underscores the transformative potential of spatiotemporal modeling and ML techniques in enhancing air pollution monitoring systems, ultimately contributing to improved air quality management and public health outcomes.
Distributed Stream Processing (DSP) engines analyze continuous data via queries expressed as a graph of operators. Auto-scalers adjust the number of parallel instances of these operators to support a target rate. Current auto-scalers couple CPU and memory scaling, allocating resources as one-size-fits-all packages. This contrasts with operators' high diversity of requirements. We present Justin, an auto-scaler that enables hybrid CPU and memory scaling of DSP operators. Justin monitors both CPU usage and the performance of operators' storage operations. Its mechanisms enable finegrain memory allocation for tasks upon a query reconfiguration. The Justin policy identifies individual operators' memory pressure and decides between adjusting parallelism and/or memory assignment. We implement Justin in Apache Flink, extending the Flink Kubernetes Operator and the DS2 CPU-only auto-scaler. Using the Nexmark benchmark, our evaluation shows that Justin identifies suitable resource allocation in as many or fewer reconfiguration steps as DS2 and supports a target rate with significantly fewer CPU and memory resources.
Since the advent of mobile devices, both end-users and the IT industry have been longing for direct device-to-device (D2D) communication capabilities, expecting new kinds of interactive, personalized, and collaborative services. Fifteen years later, many D2D solutions have been implemented and deployed, but their availability and functionality are underwhelming. Arguably, the most widely-adopted D2D use case covers the pairing of accessories with smartphones; however, many other use cases-such as mobile media sharing-did not progress. Pervasive computing and cyber-physical convergence need local communication paradigms to scale. For inherently local use cases, they are even more appealing than ever: eschewing third-parties simultaneously fosters environmental sustainability, privacy and network resiliency. This paper proposes a survey on D2D communication, investigates its deployment and adoption, with the objective of easing the creation and adoption of modern D2D frameworks. We present the results of an online poll that estimates end-users' utilisation of D2D processes, and review enabling technologies and security models.
This paper examines the equilibrium between user transaction fees and miner profitability within proof-of-work-based blockchains, specifically focusing on Bitcoin. We analyze the dependency of mining profit on factors such as transaction fee adjustments and operational costs, particularly electricity. By applying a multidimensional profitability model and performing a sensitivity analysis, we evaluate the potential for profit maximization through operational cost reduction versus fee increases. Our model integrates variable electricity costs, market-driven Bitcoin prices, mining hardware efficiency, network hash rate, and transaction fee elasticity. We show that mining strategies aimed at reducing electricity expenses are far more profitable than pursuing transactions with higher fees.
In this paper we propose an alternative task scheduling mechanism for stream processing systems such as Apache Flink, that targets resource efficiency in a multi-tenant stream processing environment with several resource heterogeneous tasks being executed in parallel. The task scheduler we propose doesn’t limit the amount of tasks that can run on each machine, instead, it adapts tasks’ allocation based on their runtime metrics, scheduling tasks to the machines with more available resources. At the same time, we explore load shedding in stream processing applications, as a mechanism to solve the tasks’ resource starvation problem that may appear due to bad decisions performed by the scheduler, because of its optimistic approach and due to the dynamic workloads of the applications. We implemented a proof-of-concept of such system in Apache Flink and tested it against scenarios that show the different aspects and advantages of the developed mechanism in action.
Many offline-first applications use an embedded relational database, such as SQLite, to manage their data. The replication of the database eases the addition of collaborative features to its applications. Most of the approaches for replicating a relational database require coordination at some extent. A few approaches propose a coordination-less replication to allow offline work. These approaches are limited in two ways: (i) They do not respect Strong Eventual Consistency that states that two replicas converge as soon as they integrate the same set of modifications; (ii) They fail to preserve the combined effect of operations’ intent in complex scenarios. We propose Synql, an approach based on Conflict-free Replicated Data Types (CRDTs) that addresses these two limitations. Synql relies on a replicated state defined by the composition of CRDT primitives. The state of the database is computed over the replicated state. The user modifications are compensated so that the computed state corresponds to what the users saw and changed.
Due to the rise of Internet of Things networks, targeting vulnerabilities related to the limitation of resources in devices has increased. Therefore, it is necessary to delegate encryption services to cloud and fog platforms. Encryption as a Service (EaaS) provides all cryptographic services to end-users to help them cope with their limited resources and processing capabilities. This paper reviews the existing research on EaaS platforms and categorizes them based on their underlying encryption algorithm types. We also introduce different EaaS architectures based on the location of the main components. To our knowledge, none of the existing surveys in this field have covered the aforementioned features.
Data streams produced by mobile devices, such as smartphones, offer highly valuable sources of information to build ubiquitous services. However, the diversity of embedded sensors and the resulting data deluge makes it impractical to provision such services directly on mobiles, due to their constrained storage capacity, communication bandwidth and processing power. Unfortunately, the improving hardware capabilities of devices are unlikely to resolve these structural issues. We, therefore, believe that mobile data management systems should, instead, handle data streams efficiently and compactly, to provision services directly at the edge, while accounting for the limits of existing assets and network infrastructures. This paper introduces the FLI framework, which leverages a piece-wise linear approximation technique to capture compact representations of data streams in mobile devices. Our experiments, performed on Android and iOS devices, show that FLI outperforms the state of the art both in memory footprint and I/O throughput. Our Flutter implementation of FLI can store stream datasets in mobile devices, which is a prerequisite to processing big data from ubiquitous devices in situ.
Deep neural networks used for computer vision tasks are typically trained on datasets consisting of thousands of images, called examples. Recent studies have shown that examples in a dataset are not of equal importance for model training and can be categorized based on quantifiable measures reflecting a notion of “hardness” or “importance”. In this work, we conduct an empirical study of the impact of importance-aware partitioning of the dataset examples across workers on the performance of data-parallel training of deep neural networks. Our experiments with CIFAR-10 and CIFAR-100 image datasets show that data-parallel training with importance-aware partitioning can perform better than vanilla data-parallel training, which is oblivious to the importance of examples. More specifically, the proper choice of the importance measure, partitioning heuristic, and the number of intervals for dataset repartitioning can improve the best accuracy of the model trained for a fixed number of epochs. We conclude that the parameters related to importance-aware data-parallel training, including the importance measure, number of warmup training epochs, and others defined in the paper, may be considered as hyperparameters of data-parallel model training.
