Defesa de Exame de Qualificação – Luís Fernando Arcaro – 10/03/2017
10/03/2017 16:38
| Defesa de Exame de Qualificação | |
| Aluno | Luís Fernando Arcaro |
| Orientador | Prof. Rômulo Silva de Oliveira, Dr. – DAS/UFSC |
| Data
Local |
10/03/2017 14h30 (sexta-feira)
Sala PPGEAS I (piso superior) |
| Prof. Luiz Cláudio Villar dos Santos, Dr. – PPGEAS/UFSC (orientador)
Prof. Sandro Rigo, Dr. – IC/UNICAMP Prof. Cesar Albenes Zeferino, Dr. – CTTMar/UNIVALI
|
|
| Título
|
Time-Randomized Hardware Elements for Increasing the Applicability of Measurement-Based Probabilistic Timing Analysis |
| Abstract: Recently proposed Measurement-Based Probabilistic Timing Analysis (MBPTA) approaches employ Extreme Value Theory (EVT), a statistical framework that provides foundation to the probabilistic analysis of extreme events, for deriving bounds on Worst-Case Execution Times (WCETs) for tasks that compose critical Real-Time Systems (RTSs). Under conditions that guarantee its applicability requirements are met, EVT is claimed to allow deriving WCET bounds with associated exceedance probabilities which can be low enough even for certification purposes, based solely on measurements of the analyzed tasks’ execution times. To be analyzed through MBPTA, the execution times of a task must be modellable as a set of independent and identically distributed random variables, and their maximum values must present frequencies that fit the probability distributions expected by EVT (e.g. Gumbel, Fréchet or Weibull). Perfect fitting to these distributions is not likely to be obtained in practice due at least to (1) the lack of probabilistic behavior of the underlying hardware, which is typically observed in traditional platforms, and (2) the discreteness of the yielded execution time distributions due to hardware construction characteristics, e.g. if variability sources are coarse-grained, which is especially harmful for short code segments (e.g. basic blocks). The introduction of time-randomized hardware elements is capable of both (A) enabling processors to present true probabilistic timing behavior, by making latencies dependent mainly on (pseudo-)random events, and (B) reducing the granularity of the yielded execution times by inducing fine-grained jitter, potentially improving the fitting quality. In this work we expect to improve MBPTA applicability by proposing (A) an interleaved multi-threaded pipeline that randomly chooses the thread that owns the next instruction to be processed, (B) a mesh Network-on-Chip (NoC) that employs randomized routing and switching policies, and (C) other time-randomized hardware elements, such as a randomly pre-charged and refreshed Dynamic Random Access Memory (DRAM) controller and a randomly-forwarded bus bridge. Since traditional fitting evaluation tools (e.g. quantile plots) do not allow the quantitative comparison of time-randomized hardware platforms in relation to their adequacy to MBPTA, we also propose developing proper metrics to assess the achieved results. With that, we expect to develop a processor capable of increasing the chances of successfully applying MBPTA for deriving Probabilistic Worst-Case Execution Time (pWCET) estimates even for small code segments, thus making it an even more attractive solution for RTSs’ timing analysis in industry. | |
