Jacopo Panerati

Comparison of Decision-Making Strategies for Self-Optimization in Autonomic Computing Systems

Autonomic computing systems are capable of adapting their behavior and resources thousands of times a second to automatically decide the best way to accomplish a given goal despite changing environmental conditions and demands. Different decision mechanisms are considered in the literature, but in the vast majority of the cases a single technique is applied to a given instance of the problem. This article proposes a comparison of some state of the art approaches for decision making, applied to a self-optimizing autonomic system that allocates resources to a software application. A variety of decision mechanisms, from heuristics to control-theory and machine learning, are investigated. The results obtained with these solutions are compared by means of case studies using standard benchmarks. Our results indicate that the most suitable decision mechanism can vary depending on the specific test case but adaptive and model predictive control systems tend to produce good performance and may work best in a priori unknown situations.

On self-adaptive resource allocation through reinforcement learning

Autonomic computing was proposed as a promising solution to overcome the complexity of modern systems, which is causing management operations to become increasingly difficult for human beings. This work proposes the Adaptation Manager, a comprehensive framework to implement autonomic managers capable of pursuing some of the objectives of autonomic computing (i.e., self-optimization and self-healing). The Adaptation Manager features an active performance monitoring infrastructure and two dynamic knobs to tune the scheduling decisions of an operating system and the working frequency of cores. The Adaptation Manager exploits artificial intelligence and reinforcement learning to close the Monitor-Plan-Analyze-Execute with Knowledge adaptation loop at the very base of every autonomic manager. We evaluate the Adaptation Manager, and especially the adaptation policies it learns by means of reinforcement learning, using a set of representative applications for multicore processors and show the effectiveness of our prototype on commodity computing systems.

A comparative evaluation of multi-objective exploration algorithms for high-level design

This article presents a detailed overview and the experimental comparison of 15 multi-objective design-space exploration (DSE) algorithms for high-level design. These algorithms are collected from recent literature and include heuristic, evolutionary, and statistical methods. To provide a fair comparison, the algorithms are classified according to the approach used and examined against a large set of metrics. In particular, the effectiveness of each algorithm was evaluated for the optimization of a multiprocessor platform, considering initial setup effort, rate of convergence, scalability, and quality of the resulting optimization. Our experiments are performed with statistical rigor, using a set of very diverse benchmark applications (a video converter, a parallel compression algorithm, and a fast Fourier transformation algorithm) to take a large spectrum of realistic workloads into account. Our results provide insights on the effort required to apply each algorithm to a target design space, the number of simulations it requires, its accuracy, and its precision. These insights are used to draw guidelines for the choice of DSE algorithms according to the type and size of design space to be optimized.

Balancing system availability and lifetime with dynamic hidden Markov models

Electronic components in space applications are subject to high levels of ionizing and particle radiation. Their lifetime is reduced by the former (especially at high levels of utilization) and transient errors might be caused by the latter. Transient errors can be detected and corrected using memory scrubbing. However, this causes an overhead that reduces both the availability and the lifetime of the system. In this work, we present a mechanism based on dynamic hidden Markov models (D-HMMs) that balances availability and lifetime of a multi-resource system by estimating the occurrence of permanent faults amid transient faults, and by dynamically migrating the computation on excess resources when failure occurs. The dynamic nature of the model makes it adaptable to different mission profiles and fault rates. Results show that our model is able to lead systems to their desired lifetime, while keeping availability within the 2% of its ideal value, and it outperforms static rule-based and traditional hidden Markov models (HMMs) approaches.

Coordination of Independent Loops in Self-Adaptive Systems

Nowadays, the same piece of code should run on different architectures, providing performance guarantees in a variety of environments and situations. To this end, designers often integrate existing systems with ad-hoc adaptive strategies able to tune specific parameters that impact performance or energy—for example, frequency scaling. However, these strategies interfere with one another and unpredictable performance degradation may occur due to the interaction between different entities. In this article, we propose a software approach to reconfiguration when different strategies, called loops, are encapsulated in the system and are available to be activated. Our solution to loop coordination is based on machine learning and it selects a policy for the activation of loops inside of a system without prior knowledge. We implemented our solution on top of GNU/Linux and evaluated it with a significant subset of the PARSEC benchmark suite.

Monitoring glaciers from space using a cubesat

The École Polytechnique de Montréal and the University of Bologna recently collaborated to develop a nanosatellite mission in the context of the 2012-2014 iteration of the Canadian Satellite Design Challenge, an inter-university competition intended for the development of space expertise among graduate and undergraduate students. The mission comprised two different scientific payloads: one aiming at monitoring climate changes in the Arctic, and the other addressing the need to reduce space debris. Here we report the organizational and technical challenges we faced, as well as the lessons we learned.

Effects of online fault detection mechanisms on Probabilistic Timing Analysis

In real time systems, random caches have been proposed as a way to simplify software timing analysis, by avoiding corner cases usually found in deterministic systems. Using this random approach, one can obtain an applications probabilistic Worst Case Execution Time (pWCET) to be used for timing analysis. As with deterministic systems, technology scaling in cache memories is making transient and permanent faults more likely, which in turn affects the systems timing behavior. To mitigate these effects, one can introduce a detection mechanism that classifies a fault as transient or permanent, with the goal of disabling permanently faulty cache blocks to avoid future accesses. In this paper, we compare the effects of two online detection mechanisms for permanent faults, namely rule-based detection and Dynamic Hidden Markov Model (D-HMM) based detection, for the generation of safe pWCET estimates. Experimental results show that different mechanisms can greatly affect safe pWCET margins, and that by using D-HMM the pWCET of the system can be improved compared to rule-based detection.

Probabilistic Resilience in Hidden Markov Models

Originally defined in the context of ecological systems and environmental sciences, resilience has grown to be a property of major interest for the design and analysis of many other complex systems: resilient networks and robotics systems other the desirable capability of absorbing disruption and transforming in response to external shocks, while still providing the services they were designed for. Starting from an existing formalization of resilience for constraint-based systems, we develop a probabilistic framework based on hidden Markov models. In doing so, we introduce two new important features: stochastic evolution and partial observability. Using our framework, we formalize a methodology for the evaluation of probabilities associated with generic properties, we describe an efficient algorithm for the computation of its essential inference step, and show that its complexity is comparable to other state-of-the-art inference algorithms.

Trading off power and fault-tolerance in real-time embedded systems

Reliability and fault-tolerance are essential requirements of critical, autonomous computing systems. In this paper, we propose a methodology to quantify, and maximize, the reliability of computation in the presence of transient errors when considering the mapping of real-time tasks on an homogeneous multiprocessor system with voltage and frequency scaling capabilities. As the likelihood of transient errors due to radiation is environment- and component-specific, we use machine learning to estimate the actual fault-rate of the system. Furthermore, we leverage probability theory to define a trade-off between power consumption and fault-tolerance. If a processing element fails, our methodology is able to re-map the application, establishing whether the real-time requirements will still be met, and how reliable the new, impaired system will be. Results show that the proposed methodology is able to adjust mapping and operating frequencies in order to maintain a fixed level of reliability for different fault-rates.

Assessing the resilience of stochastic dynamic systems under partial observability

Resilience is a property of major interest for the design and analysis of generic complex systems. A system is resilient if it can adjust in response to disruptive shocks, and still provide the services it was designed for, without interruptions. In this work, we adapt a formal definition of resilience for constraint-based systems to a probabilistic framework derived from hidden Markov models. This allows us to more realistically model the stochastic evolution and partial observability of many complex real-world environments. Within this framework, we propose an efficient and exact algorithm for the inference queries required to construct generic property checking. We show that the time complexity of this algorithm is on par with other state-of-the-art inference queries for similar frameworks (that is, linear with respect to the time horizon). We also provide considerations on the specific complexity of the probabilistic checking of resilience and its connected properties, with particular focus on resistance. To demonstrate the flexibility of our approach and to evaluate its performance, we examine it in four qualitative and quantitative example scenarios: (1) disaster management and damage assessment; (2) macroeconomics; (3) self-aware, reconfigurable computing for aerospace applications; and (4) connectivity maintenance in robotic swarms.