Defesa de Tese de Doutorado – Luís Fernando Arcaro – 8/4/2019
17/05/2019 15:58
| Defesa de Tese de Doutorado | |
| Aluno | Luís Fernando Arcaro |
| Orientador | Prof. Rômulo Silva de Oliveira, Dr. – DAS/UFSC |
| Data
|
8/4/2019 8h00 (segunda-feira)
Sala PPGEAS I (piso superior) |
|
Banca |
Prof. Rômulo Silva de Oliveira, Dr. – DAS/UFSC (presidente);
Prof. Sandro Rigo, Dr. – IC/UNICAMP; Prof. Luiz Cláudio Villar dos Santos, Dr. – INE/UFSC; Prof. Eduardo Augusto Bezerra, Dr. – DEEL/UFSC. |
| Título
|
Increasing the Reliability and Applicability of Measurement-Based Probabilistic Timing Analysis |
| Abstract: As the complexity of computer architectures grows in order to improve performance and/or to reduce costs, the use of modern processors in the design of Real-Time Systems (RTSs) is increasingly hampered by the emergence of timing effects that challenge determining reliable and tight bounds for tasks’ Worst-Case Execution Times (WCETs). The Measurement-Based Probabilistic Timing Analysis (MBPTA) technique aims determining probabilistic WCET bounds (i.e. pWCETs) by applying Extreme Value Theory (EVT) on tasks’ execution time measurements, and is hence promising in handling hardware complexity issues within RTSs’ design. Hardware-level time-randomized processors were recently proposed as a means to cause computing systems’ timing behaviour to become more easily analysable through probabilistic tools, and are designed replacing deterministic or speculative internal information with (pseudo-)random numbers. The scientific research whose outcomes are presented in this thesis produced contributions on two distinct fronts. In first place, we proposed and applied methods for evaluating the reliability of pWCET estimates produced using MBPTA, based on collecting large execution time samples and then comparing (1) the pWCETs against the largest observed execution times, and (2) pWCETs’ exceedance densities against their expected values. These evaluations led us to conclude that EVT probabilistic models intended to yield more precise bounds may often lead to pWCET underestimations, and we hence recommended that upper-bounding models should instead be used for deriving pWCETs with increased reliability. In second place, we evaluated the hypothesis that randomized scheduling techniques can benefit the timing analysis of tasks executed on multithread pipelines through MBPTA, by causing the yielded execution times to meet the technique’s basic application requirements. For that, we considered both (A) a scheduler that employs a purely random policy, and (B) a randomized scheduler capable of limiting the timing effects of inter-thread interference, without compromising analysability, by using a credit-based eligibility regulation mechanism. | |
