The Key Technology Research Of Modeling For Parallel Program

With the increasing scale of data that the program can process, single processor can not satisfy the demand for computer resources. Researchers increasingly depend on high performance computing to solve practical problems, especially in scientific computing program. The study focus on the key technology of building model for parallel program, and verify the key technology on the subjects NPB and CGPOP program.Program execution time is divided into two parts, computation and communication. Computation part of the time is considered as the instruction execution time, communication part of the time is considered as the time that communication function called.The factors influencing the computation part time are the machine instruction execution time and the number of instructions. The time that machine instruction execution is related to the platform that the program run on, access platform parameters through benchmark. The research emphasis in this paper is the number of cycles that influencing the number of instructions, designing and implementing Loop Analysis algorithm. The way to identify cycles execution numbers is analyzing the loop structure of IR code, obtaining cycles dependent variables. Obtain the cycles of the experiment subjects with the Loop Analysis algorithm and the Scalar Evolution algorithm provided by GCC compiler, Loop Analysis algorithm can obtain more cycles than that algorithm contrasting experiment results.Communication of scientific computing program is generally based on MPI protocol, the way to achieve the communication time is based on modeling for communication function. Communication function is divided into point-to-point communication and collection communication, the time is associated with program parallelism and communication count. Through the analysis of Log GP model, we use piecewise linear function to fit the test data of point-to-point communication. Collection communication is nonlinear about parallelism and communication count, fit the test data through the neural network model, and adjust the right structure and parameters to fit best. Experiments show neural network model achieve a better fitting effect.Predict basic block execution Frequency, computation time and communication time of the experimental subjects with the key technology mentioned above, and compared with the real running information. Point out the possible reason to cause error and the improvement direction.