PoF: Performance-Oriented Framework SupportingModel-Driven Multicore Software Development

• Main Authors: Chao Sheng Lin, 林朝聖
• Other Authors: Pao Ann Hsiung
• Format: Others
• Language: en_US
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/38557938682119551430

博士 === 國立中正大學 === 資訊工程研究所 === 102 === The framework PoF is a performance-oriented framework which supports the modeldriven development for multicore software. The framework contains four man components, including the model parser, parallel code generator, monitor integrator and the performance tuner. In the Thesis, the concepts of three main parallelism, i.e., task parallelism, data parallelism, and data flow parallelism, are described by SysML activity diagram. The multicore software developers implement the three concept by specifying the implementation or concept code and additional configuration information in the specific components of activity diagram. The model parser parses the implementation or concept code and the configuration information from the models, and then passes to the parallel code generator to generate the corresponding skeleton code and interface for each realization of parallelism, and put the specified user code in the position where it should be. In order to tune the performance of multicore software, the multicore monitoring system is designed and integrated into the framework. Once the parallel code is generated, the monitor integrator inserts the monitoring code in the parallel code, and the monitor system can access performance data through the monitor code. The performance then evaluates the performance data obtained from the monitor code, and tunes the configuration, such as the number of threads, of the multicore software to achieve the optimal performance with fewest computing resources, the processing cores of multicore architectures. The main contributions of the Thesis include 1) the model-driven design of three parallelism is proposed based on SysML activity diagram and the components of the multicore software models can be annotated into the corresponding parallel code, 2) the multicore monitoring system is proposed and through the framework the monitoring code is integrated in the generated parallel code, 3) the performance tuner for general-purpose multicore applications is proposed and the experimental results shows its excellent work than the other performance tuner.