Efficient Task Migration in Heterogeneous Multi-Core Systems
In modern processors where power efficiency inhibits performance and vice versa, the need for sensible decisions during execution is a necessity. In single core designs focus is given in voltage/frequency throttling techniques, while in heterogeneous multi-core architectures the degrees of freedom also include optimal core selection. The former case has been diligently studied and recent work indicates that schedulers approach near-optimal behaviour. In contrast, the latter case is very much an open subject in terms of understanding optimal heterogeneous architecture design and operation.
In this research project, we are investigating the field of heterogeneous multiprocessor performance and power-efficiency, with particular focus on methods to develop alternate heterogeneous architectures and policies that can improve performance and efficiency.
We devised a simulation-based methodology to better dissect the effects of migrations. Our proposal identifies the primary bottlenecks and allows for a quantitative comparison of different types of heterogeneous systems. Using our methodology, we propose a different architecture that, with minimal sharing, can achieve near-optimal results.