Estimating the worst-case execution time in modern architecture is a non-trivial problem. The presence of complex hardware features, necessary to improve the computational power due to the single-core performance barrier, makes this estimation difficult or even impossible if traditional techniques are employed. Probabilistic real-time and, in particular, the measurement-based probabilistic timing analysis have been developed to overcome this issue. Measurement-based approaches are very appealing for the industrial world, because they enable to obtain a statistical estimation of the worst-case execution time by directly observing the execution time of the real system. Unfortunately, the current status of probabilistic real-time is far to be acceptable for certification purposes of safety-critical systems. In this thesis, both theoretical and experimental works, related to the PhD activities, are presented, with a special focus on the uncertainty estimation and the applicability of the probabilistic real-time techniques to real platforms. Both these problems have been studied, providing theoretical tools and experimental evaluations to advance probabilistic approaches towards a safe solution to the worst-case execution time estimation problem. The issues, which still limit the use of such techniques in industrial environment, are discussed and possible future research directions presented. In addition, unconventional exploitations of probabilistic real-time are proposed for mixed-criticality, high performance computing, and energy-constrained systems. During all of these works, some software tools and datasets have been developed, released as open-source and described in this thesis.
Stimare il tempo massimo di esecuzione nelle architetture dei calcolatori moderni è impegnativo. A causa della necessità di andare oltre la potenza di calcolo singolo-core ormai sostanzialmente in stallo, l'introduzione di funzionalità complesse nell'hardware rende il problema difficile o addirittura impossibile da risolvere con gli approcci classici. Il real-time probabilistico, e in particolare le analisi probabilistiche del tempo di esecuzione basate su misure, sono stati recentemente sviluppati per ovviare a questa difficoltà. Avere a disposizione degli approcci basati su misure sarebbe molto interessante per il mondo industriale, perché consentono di ottenere una stima probabilistica del tempo massimo di esecuzione osservando direttamente il tempo di esecuzione sul sistema reale. Sfortunatamente, allo stato attuale, il real-time probabilistico è lontano dall'essere accettabile nel processo di certificazione di un sistema safety-critical. Questa tesi presenta i lavori teorici e sperimentali relativi al dottorato dell'autore, focalizzandosi in particolare sugli aspetti di stima dell'incertezza e applicabilità delle tecniche di real-time probabilistico a piattaforme reali. Entrambi questi problemi sono stati studiati, fornendo sia strumenti teorici che presentando valutazioni sperimentali, allo scopo di migliorare gli approcci probabilistici verso una soluzione atta a calcolare in modo affidabile il tempo massimo di esecuzione. Nel presente elaborato vengono prima discussi i loro limiti, che ancora oggi relegano il loro uso ad ambienti prettamente accademici, e presentate delle possibili linee di ricerca future. Inoltre, sono proposti usi non convenzionali del real-time probabilistico nell'ambito di sistemi a criticalità mista, high-performance computing e con energia limitata. Durante il progresso di queste ricerche, sono stati sviluppati e pubblicati in open-source diversi software e dataset descritti in questa tesi.
Beyond the traditional analyses and resource management in real-time systems
Reghenzani, Federico
2020/2021
Abstract
Estimating the worst-case execution time in modern architecture is a non-trivial problem. The presence of complex hardware features, necessary to improve the computational power due to the single-core performance barrier, makes this estimation difficult or even impossible if traditional techniques are employed. Probabilistic real-time and, in particular, the measurement-based probabilistic timing analysis have been developed to overcome this issue. Measurement-based approaches are very appealing for the industrial world, because they enable to obtain a statistical estimation of the worst-case execution time by directly observing the execution time of the real system. Unfortunately, the current status of probabilistic real-time is far to be acceptable for certification purposes of safety-critical systems. In this thesis, both theoretical and experimental works, related to the PhD activities, are presented, with a special focus on the uncertainty estimation and the applicability of the probabilistic real-time techniques to real platforms. Both these problems have been studied, providing theoretical tools and experimental evaluations to advance probabilistic approaches towards a safe solution to the worst-case execution time estimation problem. The issues, which still limit the use of such techniques in industrial environment, are discussed and possible future research directions presented. In addition, unconventional exploitations of probabilistic real-time are proposed for mixed-criticality, high performance computing, and energy-constrained systems. During all of these works, some software tools and datasets have been developed, released as open-source and described in this thesis.| File | Dimensione | Formato | |
|---|---|---|---|
|
PhD_Thesis-final.pdf
Open Access dal 21/12/2021
Descrizione: Main thesis file
Dimensione
3.32 MB
Formato
Adobe PDF
|
3.32 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/169433