Threadguide
Subhasis Banerjee, Kandarp Kaushik, Miodrag Bolić · 2013
Chip Multiprocessors (CMP) are there to stay in everyday computing -- starting from large scale server to mobile handheld devices. The challenge in designing a system with CMP is to utilize the processor resources optimally so that the power and performance goals are met. In this paper we present Threadguide, a framework for application adaptation augmenting mircoarchitectual support in CMP. The hardware assisted methodology collects runtime information from the program to annotate the code in appropriate locations. In subsequent passes compiler generate more refined code (thereby adapting the application) that distributes the workload across available processors and takes decisions to manage total energy to meet the power budget. The hardware is non-intrusive to the processor pipeline and does not hinder the execution of the processor. We show that with minimum hardware real estate (as low as a kilobyte) and energy overhead (less than 1% of dynamic energy), proposed scheme saves up to 18% in energy-delay product (EDP).