Full text
| |
| Author(s): |
João Batista Corrêa Gomes Moreira
Total Authors: 1
|
| Document type: | Master's Dissertation |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2010-10-26 |
| Examining board members: |
Nahri Balesdent Moreano;
Mario Lúcio Cortês
|
| Advisor: | Sandro Rigo |
| Abstract | |
The advent of the contemporary multiprocessor architectures has challenged software development. In order to overcome the hurdle of properly ordering the execution and data flows, new synchronization methods with simplified abstraction have been proposed. In this context, Transactional Memories have emerged as an alternative to traditional synchronization methods. Little is known about the effects on power consumption due to the use of ransactional memories since it is a recently proposed alternative. This work compares the Power consumption of the STAMP benchmark execution when using a STM system and a lockbased implementation. The results show that the STM implementation presented a worse performance, consuming three times more energy in avarage. In addition, the penalties deriving from the employment of locks in power consumption were assessed, indicating that, in systems where a failure in lock acquisition costs more than ten thousand cycles, the use of STMs becomes a competitive approach. The experiments with Transactional Memories executed during the first stage of this research indicated that faster simulation tools for hardware design and software testing are needed. Hence, this work describes an implementation of a simulation platform, built using hybrid abstraction level, that is able to estimate power consumption. The platform is the result of integrating functional processors described in the ArchC language with the MPARM platform, which is cycle-based. The implementation displays an average performance speedup of 2.1 and a maximum of 2.9. Inaccuracies due to power consumption estimation could be statistically adjusted by applying corrections based on linear regression. The model carries an average error of 5.85% with a maximum of 19.6% and minimum of 0.86% (AU) | |
| FAPESP's process: | 09/04707-6 - Energy consumption analysis in STMs and a multicore simulation platform with hybrid abstraction |
| Grantee: | João Batista Correa Gomes Moreira |
| Support Opportunities: | Scholarships in Brazil - Master |