Heterogeneous Hyperthreading

Heterogeneous architectures have been well studied in academia and industry bringing high performance and low power/energy designs with a scheduling algorithm that dynamically schedules different regions of application code on different cores based on the performance and power characteristics of a core and application. However, the thread switching overheads between physically separated cores make it hard to utilize fine-grained scheduling that potentially achieves further enhancement. Single Core Heterogeneous Architecture has reduced the thread-switching overheads dramatically and enabled fine-grained scheduling, and achieved significant power and energy savings with little performance penalty. However, the architecture utilizes only one of its backend pipelines at a time, hence limiting the full potential. In this paper, we propose Heterogeneous Hyperthreading, a novel heterogeneous architecture, extended from the Single Core Heterogeneous architecture. Taking advantage of Hyperthreading, the hardware controller dispatches multiple threads to both backends and switches them, if needed, in a fine-grained manner, realizing performance enhancement. Evaluation results show that the new architecture achieves 18% performance improvements over traditional Heterogeneous Architectures.