如何通过预测模型探索高效的建筑设计(英文版)(12页).pdfVIP

Efficiently Exploring Architectural Design Spaces via Predictive Modeling ˙ Engin Ipek Sally A. McKee Bronis R. de Supinski Rich Caruana Cornell University Martin Schulz Cornell University {engin,sam}@ Lawrence Livermore National Laboratory caruana@ {bronis,schulzm}@ Abstract of a target machine either to predict performance impacts of ar- Architects use cycle-by-cycle simulation to evaluate design choices chitectural changes, or to find promising design subspaces satisfy- and understand tradeoffs and interactions among design param- ing different performance/cost/complexity/power constraints. Sev- eters. Efficiently exploring exponential-size design spaces with eral factors have unacceptably increased the time and resources re- many interacting parameters remains an open problem: the sheer quired for the latter task, including the desire to model more re- number of experiments renders detailed simulation intractable. We alistic workloads, the increasing complexity of modeled architec- attack this problem via an automated approach that builds accurate, tures, and the exponential design spaces spanned by many indepen- confident predictive design-space models. We simulate sampled dent parameters. Thorough study of even relatively modest design points, using the results to teach our models the function describing spaces becomes challe