Understanding performance-resource dependency by thread slicing and curve fitting

Various programs or threads, even diverse segments within one thread, differ greatly in their intrinsic performance and resource requirement. Pursuing for the highest performance potential usually leads to over-design and over-supply in most circumstances, consuming more chip size and power. This problem is exacerbated in chip multithreading architectures, where several threads execute on-chip simultaneously, by the fact that inappropriate inter-thread resource competing and sharing probably result in resource misuse. The ultimate reason for the issue lies in the absence of mathematical approach to represent the performance-resource dependency quantitatively. In this paper, we take the rename-register-file, a representative storage resource, for instance to study the performance-resource dependency features and create a straightforward, accurate, and applicable model by thread slicing and curve fitting. The model contains just four parameters but perfectly matches the benchmarks of the SPEC CPU2000 suite. We conclude with a prospective on its application in explicit, on-demand resource allocation and precise performance control.