| Traditional CPU multi-core processor and cluster sometimes cannot satisfy the computing need of most large-scale applications.The graphic processor is widely used with its unique advantage of high-performance and parallel computing.The CPU/GPU heterogeneous cluster composed of traditional CPU processor and GPU acceleration coprocessor is gradually becoming the hot research point because it has both logic control capability of CPU and calculation intensive advantage of GPU.Heterogeneous parallel architectures match the separation between communication and computing of dataflow programming model.In this system,traditional CPU processor works together with graphic processing unit to discover the parallelism of data flow programs at the greatest extent and reasonable allocate the resources.Clearly,the task partitioning as well as scheduling,software pipeline construction and data communication of previous multicore clusters can't copy to heterogeneous cluster platform.Therefore,considering the multilevel parallel and heterogeneous structure of heterogeneous cluster platforms based on the features of dataflow applications,a data flow programming model oriented multi-core CPU/many-core GPU heterogeneous cluster was proposed and achieved.The dataflow programming model mainly includes four modules,two level task partitioning and scheduling,stage assignment and software pipeline construction,MPI/OpenCL mixed programming for data communication and C++/OpenCL objected code generation.The model has two advantages,maximized performing the potential parallelism of stream applications in multimedia area and improving overall performance through making full use of hardware configuration.The experiment selected X86 multi-core processors and amax mixed schema processor to build heterogeneous cluster as test platform and chose several typical algorithms in digital media area as benchmarks,proved the effectiveness of this CPU/GPU heterogeneous cluster dataflow programming model from multiple perspectives. |
PDF Full Text Request