Smart hardware designs for probabilistically-analyzable processor architectures
Future Critical Real-Time Embedded Systems (CRTES), like those is planes, cars or trains, require more and more guaranteed performance in order to satisfy the increasing performance demands of advanced complex software features. While increased performance can be achieved by deploying processor tech...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2022 |
| País: | España |
| Institución: | CBUC, CESCA |
| Repositorio: | TDR. Tesis Doctorales en Red |
| OAI Identifier: | oai:www.tdx.cat:10803/674151 |
| Acceso en línea: | http://hdl.handle.net/10803/674151 https://dx.doi.org/10.5821/dissertation-2117-366909 |
| Access Level: | acceso abierto |
| Palabra clave: | Àrees temàtiques de la UPC::Informàtica 004 |
| Sumario: | Future Critical Real-Time Embedded Systems (CRTES), like those is planes, cars or trains, require more and more guaranteed performance in order to satisfy the increasing performance demands of advanced complex software features. While increased performance can be achieved by deploying processor techniques currently used in High-Performance Computing (HPC) and mainstream domains, their use challenges the software timing analysis, a necessary step in CRTES' verification and validation. Cache memories are known to have high impact in performance, and in fact, current CRTES include multicores usually with several levels of cache. In this line, this Thesis aims at increasing the guaranteed performance of CRTES by using techniques for caches building upon time randomization and providing probabilistic guarantees of tasks' execution time. In this Thesis, we first focus on on improving cache placement and replacement to improve guaranteed performance. For placement, different existing policies are explored in a multi-level cache setup, and a solution is reached in which different of those policies are combined. For cache replacement, we analyze a pathological scenario that no cache policy so far accounts and propose several policies that fix this pathological scenario. For shared caches in multicore we observe that contention is mainly caused by private writes that go through to the shared cache, yet using a pure write-back policy also has its drawbacks. We propose a hybrid approach to mitigate this contention. Building on this solution, the next contribution tackles a problem caused by the need of some reliability mechanisms in CRTES. Implementing reliability close to the processor's core has a significant impact in performance. A look-ahead error detection solution is proposed to greatly mitigate the performance impact. The next contribution proposes the first hardware prefetcher for CRTES with arbitrary cache hierarchies. Given its speculative nature, prefetchers that have a guaranteed positive impact on performance are difficult to design. We present a framework that provides execution time guarantees and obtains a performance benefit. Finally, we focus on the impact of timing anomalies in CRTES with caches. For the first time, a definition and taxonomy of timing anomalies is given for Measurement-Based Timing Analysis. Then, we focus on a specific timing anomaly that can happen with caches and provide a solution to account for it in the execution time estimates. |
|---|