Optimization And Implementation Of Self-adaptive Heterogeneous Parallel Library Based On Multicore Processors

Serial computing era has ended. With the advent of multi-core processors,high-performance parallel computing has appeared as the mainstream now. The TMparallel libraries is to provide users with high-performance parallel computing andhelp developers to develop applications with efficient performance.This paper is the optimization and implementation for parallel libraries, byreference with Intel TBB library. Find the reasons for TM bottlenecks and slowefficiency. According to the problems and data provided by multiple cases todetermine the optimization program. And in accordance with optimization programachieve specific coding work.From the perspective of technical difficulties, related to parallel subtasksdivided technology. Each of the divided sub-task to parallel run in multi-coreprocessors, therefore the size of the sub-division of tasks directly affectsperformance. The idea of thread reuse to reduce the extra time by thread "wakesup". The process of serial and parallel self-adaptive learning, the data withoutlearning through learning to find the pattern in order to determine the serialdirection or the parallel direction. When judged the direction, based on the knownlearning points to curve fitting. The mathematical optimization method of leastsquares curve fitting techniques. The optimization process, one of the key forimproving operating efficiency is improved process thread allocation strategy. Intwo ways with optimizing allocation of threads, one is to reduce the switchingbetween threads and the other hand is full use the current thread.The project using C++language development, is a C++function templatelibrary. TM parallel library includes parallel_for parallel algorithms and pipelineparallel algorithm. Thus improving the efficiency of these two functions, is toimprove the overall efficiency of TM parallel library. The aimed at improvefunction parallel_for and pipeline. The optimization and implementationparallel_for follow ideas to improve the original parallel_for in a small amount ofdata or a small amount of task with low efficiency problems. Make TM morecompetitive. Add the serial and parallel path selection function, providing a promptaction for the subsequent operation. Improve the efficiency of execution the smallamount data.Under multiple testing cases to prove this optimization to improve execution performance of parallel libraries. By implementing function of the desiredoptimization goals completed. TM improves their competitiveness, and for userswith high practical value.