Pyramis is a library for quantitative evaluation of hierarchical UML statecharts with non-Markovian stochastic timing and probabilistic choices. It implements an efficient numerical approach for transient analysis until absorption and steady-state analysis, separately evaluating the Semi-Markov Process (SMP) of each model component. As Pyramis facilitates code reusability, maintainability, and extensibility, it has been easily integrated with the FaultFlow library for dependability analysis of component-based systems, supporting efficient quantitative evaluation of stochastic static fault trees without repeated events. In this paper, we use Pyramis for quantitative evaluation of workflows where activities have non-Markovian stochastic duration and where precedence constraints define a Directed Acyclic Graph (DAG). Workflows have a Service Level Objective (SLO) on their end-to-end (E2E) response time distribution at low workloads of requests. Pyramis efficiently derives the workflow E2E response time distribution, yielding a stochastic upper bound for topologies with non-well-nested precedence DAGs. In our experiments, we consider a workflow with topology derived from a real benchmark and execution times obtained from a known dataset. We report results for workflow variants obtained by increasing the number of sequential, concurrent, or alternative activities of workflow patterns. Results are promising in terms of tradeoff between accuracy and complexity.

Recent advances in Large Language Model-based Multi-Agent Systems (LLM-MAS) are revealing significant potential for streamlining workflows in many areas, including the Internet of Things (IoT), where their integration offers new opportunities to improve decision-making and system interaction. Nevertheless, the proliferation of LLM-based agents worldwide creates significant challenges in identifying and selecting suitable and trusted agents to create Multi-Agent Systems to tackle complex tasks at scale. To help address this issue, this paper presents an innovative paradigm that enables fully decentralized agent and workflow discovery through social relationships among LLM-based agents, as well as the dynamic creation of trusted Social LLM-MAS. A possible reference architecture with its main components is presented, and the basic relationship types that can be established among a global population of LLM agents are illustrated, as well as mechanisms that allow agents to discover peers leveraging navigable social graphs rather than centralized searches. We also discuss possible ways to achieve a trust and reputation system that leverages these social relationships to filter peer-reviewed agents. The resulting Social LLM-MAS paradigm envisioned will represent a mature generational leap over the two-tier Operational/Knowledge-Based frameworks increasingly available worldwide but still disconnected and non-synergistic.

Contact tracing, early diagnosis, and social distancing are crucial to contain outbreaks of infectious diseases, especially in the presence of asymptomatic infectives. Therefore, identifying the individuals at the highest risk of infectious disease and prioritizing them for testing is necessary to ensure both the public health and the cost-effectiveness of screening strategies. In this paper, we present an efficient quantitative approach to predict the spread of infectious diseases within a cluster by exploiting a stochastic model of disease evolution in an individual alongside observations of contacts, symptoms, and results of diagnostic tests. Specifically, we present an iterative solution technique to estimate the probability that a subject is infectious and not isolated over time given the observations acquired up to that time, achieving computational efficiency by disregarding dependencies among observations. We generated multiple synthetic data sets of observations, featuring different social network topologies, numbers of individuals, and densities of internal contacts, while accounting for noisy, erroneous, and missing observations. For each data set, we performed extensive experiments by varying the stochastic parameters of our approach from those used to derive a ground truth by performing stochastic simulation. Experimental results demonstrate that our approach effectively ranks subjects according to their probability of being infectious and not isolated, achieving high accuracy with respect to the ground truth. Notably, our approach outperforms an alternative baseline that accounts for dependencies among observations while maintaining a comparable runtime.

In the simulation of complex systems, quantities of interest are often limited to specific events with extremely low probability. This becomes particularly relevant in various processes of reliability engineering. Rare event simulation (RES) addresses this challenge by focusing the evaluation on a subset of the space of outcomes, reducing estimator variance to achieve an efficient trade-off between estimator accuracy and sample size. We propose an adaptive importance sampling (AIS) technique to estimate top-event failures in static fault trees (SFTs) with self-healing leaf components that alternate two states, operational and failed, with exponential inter-times. To this end, we leverage cross-entropy (CE) adaptation to iteratively refine a parametric importance function by tilting failure and repair rates of individual components in a balanced way. In particular, at each time instant, we decompose a system-level stochastic optimization problem into a set of component-level subproblems, jointly solved by recursive updates. Experimental results on a set of randomly generated models with increasing event rarity show better estimator accuracy and variance reduction with respect to alternative baselines.

A collaboration between the University of Florence and BioMérieux developed an efficient scheduling algorithm for an automated immunodiagnostic system. The challenge was coordinating many parallel biological analyses sharing a single pipettor while respecting strict timing constraints. Using Difference Bound Matrices with A* search and Floyd–Warshall normalization, the solution achieved millisecond computation times on limited hardware, leading to a patented and widely used industrial system.

The empirical cumulative distribution function (ECDF) is an efficient estimator of the distribution of sampled data. Yet, it has discontinuities at each step (where probability mass is concentrated) which result in a probability density function (PDF) with discrete spikes, preventing its use in models that require continuous PDFs. Bernstein approximants bridge this gap, yielding smooth estimators for bounded ECDFs and their PDFs. In this paper, we consider a strategy from the literature to select the order of Bernstein polynomials (BP) over bounded support, and we extend the approach to Bernstein phase types (BPH) in order to approximate distributions with unbounded support. Preliminary experiments confirm the effectiveness of this approach in balancing the bias-variance tradeoff, giving insight into the relation between the sample size and the Bernstein order. Comparative experiments with Gaussian kernel density estimators show that the BPH approximant achieves better or comparable performance on a benchmark of Erlang PDFs with unit mean, while guaranteeing a low memory footprint thanks to its closed form and allowing integration into continuous-time Markov chains. We provide a replication package to support reproducibility of experiments.

Context. The rapid growth of Deep Learning (DL) has significantly increased its environmental footprint due to the high computational demands of training and inference. While prior work has investigated energy-efficient algorithms and hardware-aware optimizations, the impact of practitioner-level ecosystem choices (language bindings, frameworks, runtimes, toolchains) remains largely unexplored.Objectives. This paper investigates how the choice of language–framework ecosystems affects DL energy efficiency. We adopt a practitioner-centric perspective to quantify operational energy–time trade-offs for both training and inference, focusing on computer vision as a representative and energy-intensive workload.Methods. We conduct an empirical case study comparing ResNet-18 and VGG-16 across six ecosystems (Python, C++, Java, R, MATLAB, and Rust), multiple frameworks (PyTorch, TensorFlow, and JAX within Python) and three datasets (Fashion-MNIST, CIFAR-100, and Tiny ImageNet).Experiments are executed on a dedicated NVIDIA L40S GPU server, using the CodeCarbon toolkit under a uniform experimental protocol.Results. Our experiments reveal substantial differences across ecosystems. Training energy varies by up to 4.6x, while execution time shows an 11x gap. For inference, the C++ ecosystem is the most efficient, whereas the R ecosystem is the slowest and most energy-intensive. Overall, the results suggest that execution time is an unreliable proxy for energy usage, and that training- and inference-phase rankings may diverge.Conclusions. Our findings indicate that language–framework ecosystem selection can materially affect the energy profile of DL applications. This work provides actionable evidence for incorporating energy efficiency into DL research and practice, offering an extensive replication package to enable reproducibility and extension of our study.

The increasing complexity of next-generation services demands efficient orchestration across the edge-to-cloud continuum to balance computational intensity, latency constraints, and resource availability. These services are typically decomposed into interdependent sub-tasks, requiring careful synchronization to meet stringent completion time requirements. The challenge is further amplified in heterogeneous and resource-constrained edge environments, where multiple providers dynamically compete for sub-task execution. This paper introduces a game-theoretic stochastic framework that optimizes system welfare from both users' and providers' perspectives, ensuring efficient task allocation across distributed computing resources. We propose a Cumulative Distribution Function (CDF)-driven game, where edge nodes serve as intermediaries between users and cloud/edge service providers. The framework is structured as a two-level mechanism: (i) a matching game governing the user-to-edge node association, and (ii) a nested Vickrey-Clarke-Groves auction selecting the optimal provider, based on a CDF-driven assessment of service completion times. To enhance feasibility in decentralized edge computing environments, provider bids are represented as uniform CDFs, establishing a dominance relation that mitigates strategic manipulation. We theoretically analyze cheating strategies, showing that truthful bidding is a rational provider behavior and that the resulting user-edge matching satisfies a suitable stability notion. Extensive simulations compare the proposed approach against a full-knowledge-based allocation, conventional game-theoretic models, and a heuristic recently proposed in the literature, evaluating the price of anarchy, system welfare, and outage probability. The results demonstrate the effectiveness of our framework in achieving resilient, cost-efficient, and low-latency orchestration across the edge-to-cloud continuum in heterogeneous edge deployments.

The Omnibus Java library efficiently derives optimal signal schedules for multimodal intersections. Specifically, intersections among tram lines with right of way and vehicle flows are considered, minimizing the maximum expected percentage of queued vehicles of each flow. Trams are modeled by Stochastic Time Petri Nets (STPNs), capturing periodic tram departures and bounded delays and travel times with general (i.e., non-Exponential) distribution. Vehicles are modeled by finite-capacity vacation queues with general vacation times determined by the intersection availability. For each vehicle flow, the expected queue size over time is derived, as well as the steady-state distribution of the expected queue size at multiples of the hyperperiod (resulting from nominal tram arrival times and vehicle traffic signals). Then, the behavior of each vehicle flow can be studied over intervals of arbitrary duration by just performing transient analysis for the hyperperiod duration, starting from the steady-state distribution of the expected queue size. Omnibus is notably designed to facilitate code usability, maintainability, and extensibility. It is available open source under the AGPLv3 licence. In particular, Omnibus leverages the SIRIO Library of the ORIS tool to model duration distributions and to specify and analyze STPNs.

Software rejuvenation is a proactive maintenance technique that counteracts software aging by restarting a system, making selection of rejuvenation times critical to improve reliability without incurring excessive downtime costs. Various stochastic models of Software Aging and Rejuvenation (SAR) have been developed, mostly having an underlying stochastic process in the class of Continuous Time Markov Chains (CTMCs), Semi-Markov Processes (SMPs), and Markov Regenerative Processes (MRGPs) under the enabling restriction, requiring that at most one general (GEN), i.e., non-Exponential, timer be enabled in each state. We present a SAR model with an underlying MRGP under the bounded regeneration restriction, allowing for multiple GEN timers to be concurrently enabled in each state. This expressivity gain not only supports more accurate fitting of duration distributions from observed statistics, but also enables the definition of mixed rejuvenation strategies combining time-based and inspection-based policies, where the time to the next inspection or rejuvenation depends on the outcomes of diagnostic tests. Experimental results show that replacing GEN timers with Exponential timers with the same mean (to satisfy the enabling restriction) yields inaccurate rejuvenation policies, and that mixed rejuvenation outperforms time-based rejuvenation in maximizing reliability, though at the cost of an acceptable decrease in availability.

As a new and disruptive technology, the introduction of large language models (LLMs) may be the first step into a paradigm shift of how we develop and deploy software-intensive systems. While the capabilities of LLM agents for software engineering and architecture tasks are currently explored, how to architect LLM-based systems appears to be to date an uncharted territory. Software architectures based on a single LLM agent face inherent challenges, such as lack of task customization, lack of memory, and limited access to ground truth. These challenges become especially pressing in real-world contexts that demand persistent context, validated information, and task-specific flexibility. As a potential solution to overcome these challenges, multiple LLM-agents can be adopted for specialized tasks within a single software-intensive system. In this contribution, we open the discourse on architecting LLM-intensive software products by presenting SALLMA, a Software Architecture for LLMbased Multi-Agent systems. SALLMA leverages two core layers, namely (i) the Operational Layer, responsible for request intent management, handling real-time task execution and dynamic orchestration of agents, and (ii) the Knowledge Layer, used to to store and manage metamodels and configurations for workflows and agents. To primarily assess the viability of SALLMA, we develop a proof of concept leveraging as key technologies Docker, Kubernetes, Python, LangChain, Hugging Face, Mistral, LLaMA, and SQL and NoSQL databases. Currently, SALLMA is deployed to provide information on behalf of public administration offices, and is currently utilized in a business simulation scenario.

Microservices are typically deployed as replica sets to serve high workloads of requests while guaranteeing high availability and reliability. Being designed for continuous longterm operation, replica-based systems suffer from software aging, leading to progressive performance degradation and possibly to failures. This problem is exacerbated by the heterogeneity of the endpoints of a microservice in terms of workload characteristics, software aging effects produced on the replicas, and QoS requirements. Thus, configuration of replica sets is challenging. In this work, we address quantitative modeling and evaluation of software aging and rejuvenation in replica sets. Specifically, we define Generalized Stochastic Petri Nets (GSPNs) that model the workload, service process, and aging effects characterizing each endpoint, as well as software rejuvenation and repair of the replicas. Steady-state analysis of these models enables deriving quantitative measures of interest, including the expected number of requests yielding a replica failure and the expected number of rejected requests. The analysis results show the effectiveness of our approach in determining the convenience of different replica set configurations, i.e., joined configurations, where two endpoints share a set of replicas, versus separate configurations, where each endpoint has exclusive access to a subset of replicas.

The rapid advancement of Large Language Model (LLM) based agents is enabling sophisticated multi-agent systems for complex task automation. However, current frameworks for orchestrating multiple LLM agents lack formalized mechanisms for defining inter-agent relationships and responsibilities. This limitation hinders, among other properties, scalability, maintainability, and governance in production deployments. We propose the implementation of an accountability-based architectural tactic that adapts the accountability pattern to LLM-based multi-agent systems, introducing the architectural components Agent Types, Agent Relationships, and a clear separation between Operational and Knowledge levels. Our architectural tactic is conceived with the threefold goal of enabling (i) declarative specification of agent capabilities and constraints, (ii) dynamic reconfiguration of agents and their relationships without code changes, and (iii) systematic tracking of responsibilities and audit trails. We validate our architectural tactic through a proof-of-concept implementation that extends SALLMA (a Software Architecture for LLM-based Multi-Agent systems) with accountability mechanisms. A comparative case study with LLM-based agents acting as a software development team showcases that our accountability-based architectural tactic reduces configuration burden, enables agent addition without orchestration code changes, and provides explicit governance mechanisms compared to non-accountability baseline orchestration.

Context: Microservices are gaining significant traction in academic research and industry due to their advantages, and technical debt has long been a heavily researched metric in software quality context. However, to date, no study has attempted to understand how code technical debt evolves in such architectures. Aim: This research aims to understand how technical debt evolves over time in microservice architectures by investigating its trends, patterns, and potential relations with microservices number. Method: We analyze the technical debt evolution of 13 open-source projects. We collect data from systems through automated source code analysis, statistically analyze results to identify technical debt trends and correlations with microservices number, and conduct a subsequent manual commit inspection. Results: Technical debt increases over time, with periods of stability. The growth is related to microservices number, but its rate is not. The analysis revealed trend differences during initial development phases and later stages. Different activities can introduce technical debt, while its removal relies mainly on refactoring. Conclusions: Microservices independence is fundamental to maintain the technical debt under control, keeping it compartmentalized. The findings underscore the importance of technical debt management strategies to support the long-term success of microservices.

In performance engineering of composed services, coordinated provisioning can reduce the amount of resources required to meet end-to-end response time objectives. To this aim, various intertwined aspects of the application architecture need to be taken into account, notably including precedence constraints in the composition of elementary services, along with their durations and sensitivity to the scaling of provisioned resources. We address coordinated provisioning of resources for elementary services with stochastic durations with general distributions (i.e., including non-exponential distributions). We compose services in a workflow where precedence constraints define a Directed Acyclic Graph (DAG) and the distribution of the end-to-end (E2E) response time is subject to a Service Level Objective (SLO). We leverage a surrogate model of service performance, assuming a low workload of workflow requests (i.e., a single-request scenario) and service durations inversely proportional to provisioned resources. Given the total amount of resources, our approach derives the service provisioning that optimizes the workflow E2E response time distribution, by exploiting a compositional approach and by using stochastically ordered approximations to manage dependencies in non-well-nested precedence DAGs. Then, the approach scales provisioned resources up or down to determine the minimum amount of resources needed to satisfy the SLO, while leaving the remaining resources for horizontal scaling in order to manage multiple workflow requests at high workloads. Experiments consider low-workload and high-workload scenarios, different relations between elementary service durations and provisioned resources, and workflow topologies taken from benchmarks or randomly generated with controlled statistics, using elementary service durations from a dataset of the literature. Results show that the technique is feasible also for workflows with a thousand of services and that it outperforms other provisioning methods in fitting the SLO using the same resource amount and in minimizing the resource amount needed to fit the SLO.

The ever-increasing complexity of modern software architectures has exacerbated the need for advanced software tools able to track software execution traces to improve software reliability. In this paper, we present OREO, a tool for offline and run-time monitoring and fault localization. The tool implements a novel method enabling to trace software executions to discover the run-time status, dependencies, and interactions among software components. OREO is based on a timeline extractor, i.e., an abstraction of component lifecycles and their interactions. The timeline extractor enables the tool to perform a runtime health state examination of the software under analysis. The profiler is then used to analyze the error propagation originated during the running states among software components. In so doing, the possible fault-error-failure chains are identified. To showcase the capabilities of OREO and its flexibility, we report the execution of the tool on three software projects of different nature, sizes, and architectures. The analysis results in the localization of fault-error-failure chains and safe components of the three software projects. A discussion of the versatility, scalability, and applicability of the proposed tool to a rich variety of application contexts is provided.

Tramways decrease time, cost, and environmental impact of urban transport, while requiring multimodal intersections where trams arriving with nominal periodic timetables may have right of way over road vehicles. Quantitative evaluation of stochastic models enables early exploration and online adaptation of design choices, identifying operational parameters that mitigate impact on road transport performance.We present an efficient analytical approach for offline scheduling of traffic signals at multimodal intersections among road traffic flows and tram lines with right of way, minimizing the maximum expected percentage of queued vehicles of each flow with respect to sequence and duration of phases. To this end, we compute the expected queue size over time of each vehicle flow through a compositional approach, decoupling analyses of tram and road traffic. On the one hand, we define microscopic models of tram traffic, capturing periodic tram departures, bounded delays, and travel times with general (i.e., non-Exponential) distribution with bounded support, open to represent arrival and travel processes estimated from operational data. On the other hand, we define macroscopic models of road transport flows as finite-capacity vacation queues, with general vacation times determined by the transient probability that the intersection is available for vehicles, efficiently evaluating the exact expected queue size over time. We show that the distribution of the expected queue size of each flow at multiples of the hyperperiod, resulting from temporization of nominal tram arrivals and vehicle traffic signals, reaches a steady state within few hyper-periods. Therefore, transient analysis starting from this steady-state distribution and lasting for the hyper-period duration turns out to be sufficient to characterize road transport behavior over time intervals of arbitrary duration.We implemented the proposed approach in the novel OMNIBUS Java library, and we compared against Simulation of Urban MObility (SUMO). Experimental results on case studies of real complexity with time-varying parameters show the approach effectiveness at identifying optimal traffic signal schedules, notably exploring in few minutes hundreds of schedules requiring tens of hours in SUMO.

To improve performance or reliability, systems frequently include multiple components that operate in parallel or with limited interaction, e.g., replicated components for triple modular redundancy. We consider components modeled by independent and possibly different continuous-time Markov chains and propose an approach to estimate the distribution of first passage times for a combination of component states (e.g., a system state where all components have failed) without generating the joint state space of the underlying Markov chain nor evaluating probabilities for each of its states. Our results highlight that the approach leads to accurate approximations with significant reductions of computational complexity.