Research On Dynamic Fine-Grained Optimal Power Allocation Method In Multi-Core System

With the dual driven of rapid development of integrated circuit and emerging applications, multi-core processor with many advantages becomes the mainstream of the Times. Because the multi-core processor can meet the computing performance requirements of cloud computing, high performance computing, data center, and other fields, it has been widely researched and used today. Multi-core processor has brought a giant promotion on the performance for us, while its power consumption also increases greatly and has been exceeded the chip power budget. Therefore, multi-core processor energy management is becoming the research direction and hot topic. Meanwhile, it is becoming more and more important and complicated in a large-scale multi-core system. Different from the traditional low power design of multicore, now we are facing a problem that is how to allocate the power among the multicore system resources under a strict energy constraint and obtains the optimal overall system performance. Due to the resources integrated to system numerous, applications various and input power change greatly, multi-core systems of energy management is also becoming more and more difficulty. Computing module, communication module and cache module in the system have the different influence to the performance of the system, and each component (such as each core or each router of NoC) in module also has difference influence on the performance of the system. In existing study of multi-core system energy management, it mainly focused on a single module under the condition of power constraint to get the optimal system performance or coarse-grained manage the energy for several modules. In this paper, basis on existing researches, I aim at three key modules of computing, communication and cache, and propose a two-level real-time dynamic fine-grained power allocation method under the condition of power constraint. Through building the corresponding power-performance model, I implement a top-to-down hierarchical structure for energy management to overcome disadvantage of centralized management method. This method can perceive the variation of power and the workload of each component and allocate power fine-grained on-line to optimize the energy management. The time complexity and the hardware or software costs are low in this method. Most of all, it has a good adaptive ability on the scale of the problem. The experimental results show that compared with other state-of-the-art multi-core system power allocation methods, our method can reduce 25.6% of the average execution time and at least 30.4% of the average network latency. Therefore, the proposed approach in this paper is more practical for the many-core Time in the future and it can manage the power of multicore system better.