Deterministic Network Calculus (DNC) is a mathematical framework for the worst-case analysis of networked systems. In less-than-trivial cases, pen-and-paper computation of DNC expressions is not viable, and the support of a software library for the automated computation is instead required. Existing software libraries for DNC focus on computing single operations, lacking the ability to manipulate symbolic expressions. On the other hand, a Computer Algebra System (CAS) for DNC needs means to efficiently construct, visualize, manipulate and simplify complex expressions, to automatically optimize computation runtimes. To address this gap, we developed Nancy.Expressions, an open-source library which can work on arbitrarily complex DNC symbolic expressions. It is designed for ease of adoption, in particular for projects based on the Nancy library. The library is able to analyze an expression and apply equivalences and properties before making any computation, possibly reducing its complexity. It includes a specialized grammar, that enables the user to define new properties to simplify expressions. In this paper we present this library in its goals, architecture and semantics, highlighting with examples its ease of adoption and capabilities for performance optimizations.
Nowadays, data centers are taking a central role in the world of ICT, due to the increase of the computational capabilities needed for modern applications, especially considering the growing importance of artificial intelligence and machine learning. In the context of data centers, that can be defined in a high level manner as a set of servers providing computational capabilities to interested parties, alongside with the increasing demand and performance needed, there is the problem of the environmental impact, especially in terms of power consumption. A technology promising with regards to these problems is asymmetric multiprocessing. Asymmetric Multiprocessors, AMPs, are processors composed by cores with different characteristics, so combining both performance oriented and power efficiency oriented capabilities. This work has the goal of exploring potential application of AMPs in the context of data centers, by implementing Petri Net models for theoretical data centers using AMPs, and analyzing their behaviour with respect to target performance indexes with JMT, a suite of Java tools for the definition and performance evaluation of computing systems, including a Queueing network and Petri net simulator.
Energy control systems need to balance the use of locally produced and consumed power. Their goal is to maximize self-use by charging local storage facilities while maintaining grid-convenience, i.e. avoiding largely different successive feed-in rates to the grid. To avoid over- or undercharging of the battery, control needs to respect certain safety thresholds. Recently, energy control systems have been enhanced with learning components for optimized control. This however, makes it more difficult to ensure that control respects above-mentioned safety thresholds, especially in the presence of incomplete information. This paper proposes an approach for safe learning in energy control systems by combining a shielded reinforcement learning (RL) agent that determines which percentage of the locally produced energy is stored in a battery and which percentage is fed into the grid, with a digital twin of the battery that maintains information on the State of Charge of the battery. To ensure safe learning, we use a formally verified shield, which ensures under certain assumptions that unsafe actions, i.e. overcharging of the battery, are avoided. For formal verification, the RL agent is replaced by a contract and we assume that the digital twin is always available and provides missing information to the RL agent in case of communication losses. This combination then allows us to formally verify the safety of the resulting energy control system with complex discrete and continuous dynamics. We illustrate our approach by developing a Simulink model of a real smart house in Heeten, NL [35]. Our experimental results demonstrate that both self-use and grid-convenience can be achieved while maintaining safe battery use.
Tandem queues with zero buffers find many applications in systems where jobs are processed via a predefined number of sequential processing steps. From an analytic point of view, the analysis of such systems is highly complicated, caused by the phenomenon of after-service blocking (ASB). ASB occurs when a completed job occupies a server while waiting for the next phase. This challenge is addressed by introducing and comparing heuristics to quantify the impact of ASB on the performance of the tandem queue. Our proposed heuristics offer significant advantages over existing methods, delivering more insightful and accurate performance estimations. This allows for faster evaluations and their use in optimization, making it ideal for real-time decision making and capacity allocation. This research presents a valuable tool for queueing system designers and managers to make informed decisions regarding capacity allocation, resource management, and service level optimization, particularly in scenarios with limited budgets.
With the increase in the number and scale of data centers, more attention is being focused on their energy consumption and, above all, enabling a better balance between performance and power consumption. To address the trade-off problem, we propose a multi-server setup queueing model with a power-saving policy, where a server switches to a power-saving mode when it becomes idle. Upon the arrival of a job, a server will be set up to return to normal mode and process the job in the power-saving mode simultaneously. Under a Markovian setting, we obtain the steady-state probabilities and system performance. The analytical results allow us to evaluate the trade-off between performance and energy using an existing speed-based cost function. This paper presents several numerical experiments on performance evaluation and derives optimal values for speed parameters.
Malware poses a growing threat to computer security, with potentially devastating consequences for individuals and enterprises. Understanding the strategies employed by these cyber threat programs and their interactions when competing is essential for developing effective countermeasures and safeguarding digital infrastructures. In this article, we investigate a population of digital nodes (such as phones, computers, … ) under attack from two competing malware, each aiming to maximize the number of compromised devices in the network. These malware compromise nodes to harness their computational resources for targeted activities such as cryptocurrency mining, cloud computing, and other resource-intensive tasks. We model the network using a susceptible-infected-resistant (S2IR) epidemiological framework, where entities move between susceptible, infected (active, those whose resources are exploited or passive, those used to propagate the malware), or resistant states. Furthermore, we consider that malware developers set a resource consumption rate from their controlled hosts, affecting the detection and recovery rate of these nodes. Therefore, instead of focusing on optimizing resource usage, the objective of each malware is to maximize the peak of active devices. We formulate this as a non-cooperative game and analyze infection rate strategies at Nash equilibrium using a function approximating the count of active hosts at each instant. Numerical results validate our analysis, allowing us to assess the maximum number of active hosts for each malware. Our work provides insights into the dynamics of competition among malware, solely motivated by the goal of maximizing the share of the active population.
In this paper, we address the challenge of state space explosion in the analysis of large stochastic models by advancing the lumpability approach, a state aggregation technique that exploits structural regularities in Markov chains to efficiently compute stationary performance indices. We generalize the concept of proportional lumpability, which extends the well-known notion of lumpability and allows for the exact computation of stationary performance indices, in contrast to quasi-lumpability, which only provides bounds. Proportional lumpability is achieved through a perturbation of the original Markov chain’s transition rates, guided by a proportionality function. We further explore the idea of perturbing Markov chains through left and right multiplications by square matrices, introducing the concepts of left and right-perturbed Markov chains, which preserve the original model’s topology. For left-perturbed Markov chains, the steady-state distribution of the original chain can be derived by multiplying the probability vector of the perturbed chain by the square matrix used to define the perturbation. In contrast, for right-perturbed Markov chains, the steady-state probability distribution remains unchanged.
The complexity of contemporary computational tasks necessitates the development of efficient optimization techniques to ensure optimal performance, resource utilization, and satisfaction of constraints. In scenarios where parallel computational power is available, careful selection of the software/hardware configurations to test is even more crucial. This work tackles computationally expensive, constrained black-box optimization problems within cloud computing and High-Performance Computing settings. We propose PAK-MAN (PArallel Knowledge with MAchiNe learning), a novel parallel optimization algorithm leveraging Bayesian Optimization (BO) and Machine Learning (ML) models. PAK-MAN combines the iteration efficiency of BO methods and the ability of ML to predict constrained resources with remarkable accuracy. This integration manages the exploration-exploitation trade-off effectively while avoiding unfeasible configurations. We propose synchronous and asynchronous versions, both with their strengths. We evaluate the algorithm in various scenarios, including cloud resource management, edge computing, and HPC. Comparative analysis with an established state-of-the-art method demonstrates superior performance, improving simple regret by up to 21
As a popular language running on High-Level Language Virtual Machines (HLLVMs), Java is an important programming language both in industry and academia. The Java Virtual Machine is the execution platform for applications implemented with Java. One of the most significant performance bottlenecks on an x86 Instruction Set Architecture is stalling. Therefore, understanding the effect of stalls on Java applications executed on the x86 platforms is important to application development and compiler design. In this paper, a survey of stalls in Java applications on x86 platforms is carried out to gather critical information about the influence of stalls on various Java applications and the frequencies of different types of stalls in Java workloads. Based on such information, a stall-focused benchmark suite is proposed and validated.
The paper focuses on a particular polling system known as the cyclic Bernoulli polling (CBP) system, where a server moves cyclically between the stations and serves the queue at a station with a certain probability when polled. Each station follows either a gated or partially exhaustive service discipline. In the steady state of such a system, we study a new game-theoretic aspect, where, the stations strategically choose the probability of accepting or rejecting the service from the server when polled. We examine three variants of non-cooperative games among stations: (i) each station selfishly minimizes its expected waiting time, (ii) a team game where each station minimizes the expected workload of the system, and (iii) stations act with partial cooperation, incurring an additional linear cost. We begin by presenting a new result for the CBP system regarding the continuity of expected waiting times in relation to the probabilities selected by the stations. For each game, we then investigate the existence and uniqueness of the Nash equilibrium (NE). In some cases, the NE is explicitly derived, while in others, characterizing the NE remains challenging due to the complex dependence of waiting times on the non-trivial buffer occupancy equations. Nonetheless, we analyze the NE and its properties through numerical experiments. Notably, in many instances, stations opt to accept service with a probability less than 1—a trend observed even among selfish stations.
New and evolving Artificial Intelligence (AI) applications span across the full spectrum of computing resources, seamlessly integrating across Edge and Cloud platforms. Deploying applications at the network edge reduces latency, while cloud computing ensures a higher processing power. Managing resources in such a dynamic and diverse environment requires a strategic approach to satisfy Quality of Service requirements and minimize costs. To address this challenge, we propose FIGARO (reinForcement learnInG mAnagement acRoss computing cOntinuum), which exploits offline training and imitation learning to speed up the training of reinforcement learning-based agents able to control resources in the full cloud continuum stack. By extending our framework, we designed a hierarchical system structure and tested agents that only need to manage one computational layer at a time. This approach enables the system to efficiently manage multiple application components in complex AI pipelines. The results demonstrate the effectiveness of the hierarchical method, as the local agents dynamically scale computational resources, limiting QoS constraint violations to a maximum of 1.4
With network neutrality regulation imposed worldwide at different extents, there is a need for monitoring tools to verify if Internet service providers (among other actors) comply with the existing rules. Among the very few maintained tools, Wehe has been developed by Northeastern University and has been publicized by the French regulator, ARCEP. The tool runs traces and produces a positive or negative response about service differentiation, but users have to trust that result without any visual perception of the degradation. Our goal in this paper is to describe a demonstrator of Wehe for video streaming, from which we can simultaneously see a baseline video and its manually-differentiated counterpart, as well as run Wehe to evaluate its response with respect to the perceived quality degradation. The demonstrator allows to study the efficiency of Wehe in order to propose potential improvements, if needed.
A method for parameter estimation of biochemical networks has recently been proposed using dynamic Bayesian network (DBN) approximations. Although it has proven to be more efficient than ODE-based methods, it still suffers from the curse of dimensionality. To maximize the effectiveness of the learning method, we suggest using model reduction as a preliminary step to simplify the biochemical networks before performing the actual parameter estimation. In this paper, we discuss the most common reduction methods and highlight their limitations when applied to this framework, and propose adaptions that aim to significantly improve the parameter estimation method.
In this paper, we analyze a vaccination control in a large population of agents for the Susceptible-Exposed-Infected-Recovered (SEIR) model. We model the problem of selfish vaccination as a mean field game in continuous time with a finite number of states. Our findings indicate that this game yields a unique and pure equilibrium characterized by a bang-bang strategy. Furthermore, we evaluate the overall cost of the mean field equilibrium and contrast it with the cost of the socially optimal vaccination strategy. The results of the study show that when individuals are left to decide whether to vaccinate or not, the mean field equilibrium stops vaccinating at an earlier time than the socially optimal strategy. This implies that promoting vaccination may be necessary to incentivize individuals to adopt the optimal behavior and reduce the overall cost of the mean field equilibrium.
This paper presents an approach to address the complex scheduling problem faced by sales representatives. Many scheduling and routing problems have been explored before in research but real-life applications often feature complex constraints and requirements. The problem we propose is a variant of the travelling salesman problem, in which customers must be visited multiple times over a long planning period at different intervals, with daily workload constraints, regularity and varying flexibility in the requirements, with the goal of minimizing the total travel distance. To tackle this challenge, which deviates significantly from classical routing problems, a genetic algorithm-based solution has been developed and integrated into a user-friendly web application. The algorithm effectively generates feasible and efficient schedules, considering several and diverse factors. Experimental results demonstrate the algorithm’s ability to produce high-quality solutions, outperforming manual planning methods and the other tested approaches. The web application provides an intuitive interface for problem definition and solution visualization, facilitating user interaction and optional refinement.
Parameter synthesis for Collective Adaptive Systems (CAS) is computationally intensive, often rendering the analysis of large-scale stochastic models impractical. This paper addresses this challenge by introducing a parameter synthesis framework integrated into the Sibilla tool. Our approach utilises machine learning surrogate models to reduce computational overhead, enabling efficient exploration of complex parameter spaces. The framework supports various synthesis tasks, including optimal feasibility, and introduces a new technique: the Variant Model Adjustment approach. This technique allows for systematic comparison and fine-tuning of model variants while adhering to specified temporal logic properties. Several case studies are provided to demonstrate the effectiveness of the framework in identifying optimal parameter configurations.
We provide an example of a strictly subcritical multiclass queueing network which is unstable under the longest job first (LJF) service protocol. It is a reentrant line with two servers and four customer classes. The customer arrivals in our system are modeled by a renewal process with a lattice interarrival time distribution, while the corresponding service times are deterministic. We also present examples of deterministic and stochastic strictly subcritical linear resource sharing networks which are unstable under the LJF policy. Finally, we give an example of a deterministic, strictly subcritical two-node linear network with shared resources which is unstable under the longest remaining time first (LRTF) scheduling algorithm.
Given the significant advances in machine learning techniques on mobile devices, particularly in the domain of computer vision, in this work we quantitatively study the performance characteristics of 190 real-world vision transformers (ViTs) on mobile devices. Through a comparison with 102 real-world convolutional neural networks (CNNs), we provide insights into the factors that influence the latency of ViT architectures on mobile devices. Based on these insights, we develop a dataset including measured latencies of 1000 synthetic ViTs with representative building blocks and state-of-the-art architectures from two machine learning frameworks and six mobile platforms. Using this dataset, we show that inference latency of new ViTs can be predicted with sufficient accuracy for real-world applications.
In order to obtain a stochastic model that accounts for the stochastic aspects of the dynamics of a business process, usually the following steps are taken. Given an event log, a process tree is obtained through a process discovery algorithm, i.e., a process tree that is aimed at reproducing, as accurately as possible, the language of the log. The process tree is then transformed into a Petri net that generates the same set of sequences as the process tree. In order to capture the frequency of the sequences in the event log, weights are assigned to the transitions of the Petri net, resulting in a stochastic Petri net with a stochastic language in which each sequence is associated with a probability. In this paper we show that this procedure has unfavorable properties. First, the weights assigned to the transitions of the Petri net have an unclear role in the resulting stochastic language. We will show that a weight can have multiple, ambiguous impact on the probability of the sequences generated by the Petri net. Second, a number of different Petri nets with different number of transitions can correspond to the same process tree. This means that the number of parameters (the number of weights) that determines the stochastic language is not well-defined. In order to avoid these ambiguities, in this paper, we propose to add stochasticity directly to process trees. The result is a new formalism, called stochastic process trees, in which the number of parameters and their role in the associated stochastic language is clear and well-defined.
We study content fetching and delivery in an edge caching system consisting of a sensor, a set of users, and an aggregator. The users occasionally request the aggregator for the dynamically varying sensor measurements also called the content. The aggregator may fetch the fresh content from the sensor and serve, may serve the locally cached versions, or may not serve at all. Content fetching and delivery incur fetching and transmission costs, respectively, and serving stale content also incurs age cost. We study optimal content fetching and delivery problem, aiming to minimize the time-averaged content fetching, transmission, and age costs. The problem lends itself to Markov decision problem framework albeit with a high dimensional state space. We first consider a single-user problem and derive the optimal policy. We then propose a heuristic for the multi-user problem, which is obtained by combining solutions to certain fictitious versions of the single-user problems. We numerically illustrate the derived content fetching and delivery policies and their properties.
