FRiDA: uma ferramenta de predição para rápida exploração do espaço de projeto de processadores combinados com aceleradores reconfiguráveis
Every year the demand of embedded applications for computational resources increases. To meet this demand, the embedded system designs have made use of the combination of diversified components, resulting in heterogeneous platforms that aim to balance the processing power with the energy consumpt...
Na minha lista:
Autor principal: | |
---|---|
Outros Autores: | |
Formato: | doctoralThesis |
Idioma: | pt_BR |
Publicado em: |
Universidade Federal do Rio Grande do Norte
|
Assuntos: | |
Endereço do item: | https://repositorio.ufrn.br/handle/123456789/32220 |
Tags: |
Adicionar Tag
Sem tags, seja o primeiro a adicionar uma tag!
|
Resumo: | Every year the demand of embedded applications for computational resources increases.
To meet this demand, the embedded system designs have made use of the combination
of diversified components, resulting in heterogeneous platforms that aim to balance the
processing power with the energy consumption. Reconfigurable hardware accelerators
emerge as an alternative to face this demand. However, a key question in the design of
general purpose processors (GPPs) coupled to reconfigurable accelerators (RAs) is which
components to combine to meet the expected performance at the cost of additional area
and power. To perform a vast design space exploration (DSE) allows the designer to
estimate the cost of these platforms before the manufacturing phase. However, the number
of possible solutions to be evaluated grows exponentially and evaluating the cost of all
those solutions is an infeasible task. In this work, one presents FRiDA, a Predictive tool
for Fast DSE of Processors combined with Reconfigurable Accelerators. Using FRiDA, one
is able to speed up the design space exploration of heterogeneous systems combining GPPs
and RAs. One employs an approach based on machine learning algorithms. By simulating
a subset of the design space using a high level simulator, regression models are trained to
fast predict the cost of unsimulated architectural configurations. Several regression models
were evaluated to be used by FRiDA and the regression ensembles ones, in particular the
textit Gradient Boosting model, show the best tradeoff when aspects as accuracy and
throughput were considered. In the case study used to validate the tool, it was possible to
reach prediction error rates below 3.5% when the results were compared to a high-level
simulator. Using the machine learning based techinique, one was able to perform more
than 6,000 predictions per second. This fact allowed to go through the investigated design
space with more than 100,000 architectural configurations in less than 30 seconds. FRiDA
allows the designer to define which design aspects should be optimized in the project and
also allows to include new aspects to be optimized. FRiDA helps the designer to explore
thousands of configurations and to find architectural solutions with high efficiency with
a very low error prediction rate. Using a multiobjective algorithm, FRiDA quickly finds
efficient solutions that satisfy one or multiple conflicting objectives of the project. |
---|