Research on Elastic Extension of Multi Type Resources for OpenMP Program

OpenMP is one of the most popular programming models in the field of parallel computing, which is widely used in multi-threaded programming with shared memory. But with the continuous change of program computing scale, the problem of resource supply has become increasingly prominent. Making use of the cloud computing to dynamically provide resources for OpenMP program is a meaningful work. Based on the internal structure characteristics of OpenMP program, a flexible elastic resource allocation strategy is designed, and a load prediction method combining active and reactive prediction based on machine learning is proposed to realize the load prediction of OpenMP program. Then, the fuzzy controller is used to coordinate the resources and infer the CPU and memory resources needed by the program in the next cycle. On this basis, the middleware OPEC is designed and implemented to realize the dynamic allocation of cloud resources during the operation of OpenMP program. Parsec, a benchmark test suite based on shared memory, is selected for experimental verification. The experimental results show that the solution proposed in this paper can dynamically provide a variety of resources for OpenMP program to meet program running requirements, effectively improve resource utilization efficiency and reduce program running costs.