Research And Application Of Multi-GPU Parallel Computing Based On OpenCL

Heterogeneous computing, which is regarded as the third era after thesingle-core and multi-core for the development of computer processor, hasrealized the co-computing among the processors with many architectures,alleviated the heat dissipation and energy consumption issues encountered duringthe process of improving CPU clock frequency and kernel number, as well asenhanced the expandability of computing platform.Except for CPU, the heterogeneous computing system usually includes oneor a few co-processors, which are computing devices with the dedicated functionscontaining numerous processing cores and whose parallel computing power canonly be fully utilized by the proper parallel program. OpenCL is a standardframework programmed for the heterogeneous equipment, and although itsappearance has increased the portability of procedures among all platforms, suchproblems as the invocation of many computing devices, load balancing amongdevices and so on have restricted its popularization severely upon thecollaborative computing of many computing devices. Therefore, it is of greatsignificance to research the parallel computing of many devices under theheterogeneous system.Based on the OpenCL standard, this paper has realized the invocation ofmulti-GPU under many platforms by utilizing the multithreading technology,analyzed the influence of different task partitioning models on the calculatedperformance and optimized the data transmission among memorizers. Then, twoalgorithms which can apply multi-GPU for auxiliary acceleration are designed toverify the feasibility of research on the multi-GPU parallel technology. The firstalgorithm, an internal sort algorithm suitable for the larger-scale data, hasdesigned the GPU internal sort algorithm fitting the devices and the merge sortalgorithm based on the loser tree data structure fitting the host. Compared with CPU serial sort algorithm, upon the sort operation of larger-scale data, theperformance of this algorithm has been lifted for about ten times with theassistance of GPU. Another algorithm utilizes the multi-GPU to acceleratesolving the flexible job shop scheduling problem, designs the data structuresuitable for GPU framework and Genetic Algorithm based on the Island Model,as well as improves the evolution of individual in the population and the appraisalefficiency of individual by utilizing GPU. With the acceleration of multi-GPU,compared with the Genetic Algorithm to solve the flexible job shop schedulingproblem based on CPU, this algorithm has obtained a better solution within ashorter time on the premise of processing a larger population size.