Research And Application Of Load-Prediction Scheduling On CPU-GPU Heterogeneous High Performance Computing

With high price-quality ratio and energy-efficiency ratio, the Multicore CPU-GPU computing platform is extensively used, and it also makes the existence of two heterogeneous computing resources in the stand-alone. However, it is the efficient scheduling algorithm that brings the Multicore CPU-GPU computing platform into full play. Therefore, how to make full use of heterogeneous computing resources and how to achieve load balancing become hotspot of research.The traditional scheduling methods of round-robin scheduling are static scheduling and dynamic scheduling. Static scheduling costs little, however, it is easy to lead to the imbalance of load and the reduction of computing resources’ use ratio; dynamic scheduling can achieve load balancing better with more scheduling cost. And, if we combine the two referred scheduling methods, we can largely reduce the cost of scheduling and achieve load balancing. While GPU is based on SIMD structure and suitable for computing tasks with high parallelism and large calculation, and GPU’s calculated performance is much better than that of CPU, however, there is no scheduling algorithm for now which can assign the computing tasks appropriately according to the characteristic of the hardware in the CPU-GPU heterogeneous computing platform.Aiming at the above questions, this article comes up with a new scheduling method--Load-Prediction Scheduling(LPS) which can bring the heterogeneous Multicore CPU-GPU’s computing ability into full use and achieve the valid combination of static and dynamic scheduling. These main jobs completed are included:1. This article puts forward LPS, and these characteristics of the algorithm are included:1) Assign the tasks according to the characteristics of GPU’s hardware, and bring GPU’s computing performances into full play.2) Efficiently combine dynamic scheduling and static scheduling, and achieve load balancing and reduce the cost of scheduling so as to be suitable to be applied in the heterogeneous environment.3) Bring multicore CPU’s computing performance into full play.2. Apply the LPS to Computer Simulation of Electrocardiogram and achieve the above characteristics. In addition, by load-prediction the branches in computing are removed, and the GPU’s computing granular is improved, therefore, the computing scope of GPU is broaden and computing efficiency is further increased.3. Apply the LPS to N-Body problem.