Magellan: A Framework for Fast Multi-core Design Space Exploration and Optimization Using Search and Machine Learning

As multi-core processor architectures with tens or even hundreds of cores, not all of them necessarily identical, become common, the current processor design methodology, that relies on large-scale simulations, is not going to scale well because of the number of possibilities to be considered. We need intelligent/efficient techniques to navigate through the processor design space. In this paper, we propose to treat processor design space exploration as a classical search problem. We adapt several well known (and some less known) search/optimization techniques that have been used very successfully in other domains to the problem of efficiently exploring the processor design space. We observe that these techniques result in multi-core processors whose performance is comparable (within 1%) to a processor design that requires an exhaustive exploration of the design space. These techniques often take orders of magnitude (a factor of 3800 at the minimum) less time for coming up with these processors. We also show that machine learning-based techniques can be applied on top of these search/optimization-based techniques to prune the search space even further. We leverage the knowledge gained in this research to develop Magellan – a framework for accelerating multi-core design space exploration and optimization. Magellan can be used to find the highest throughput processors of a given type for a given area, power, or time budget. It can be used to aid even experienced processor designers that prefer to rely on intuition by allowing fast refinements to an input design.