COMPUTING SYSTEMS MULTI-OBJECTIVE OPTIMIZATION USING DOMAIN-KNOWLEDGE

Abstract

The main aim of this short paper is to point out the advanced computing systems’ optimization research developed by The Advanced Computer Architecture and Processing Systems Research Centre (ACAPS) from “Lucian Blaga” University of Sibiu, Romania. Multi-objective optimization (performance, power consumption, temperatures, complexity…) of computing systems having many parameters is a very complex problem. Not only the hardware needs to be simulated and evaluated; frequently we need hardware and software co-optimization (cross-layer optimization). Usually exhaustive search is prohibited due to the enormous design space. The solution consists in developing and implementing some advanced heuristic algorithms in order to solve this NP-hard problem. In our optimization research we used a Pareto-based approach by implementing some multi-objective evolutionary algorithms and bio-inspired algorithms belonging to the Particle Swarm Optimization class. These briefly presented Design Space Exploration (DSE) characteristics were implemented into a dedicated software product entitled Framework for Automatic Design Space Exploration. In order to achieve better convergence speed and solutions’ quality we implemented specific domain-knowledge for each of the target computer architecture to be optimized. Domain-knowledge is represented by a system of complete, non-redundant and noncontradictory rules or other specific restrictions. As far as we know, we were the first ones using fuzzy logic as a method to express computer architecture knowledge into a DSE tool

Keywords: computing systemshardware and software co-optimizationmulti-objective optimizationgenetic algorithmsdomain-knowledgemeta-optimization