为了充分利用多核处理器提供的多级并行和解决多核资源分配问题,提出了一种将多任务并行程序映射到多核处理器平台上的模型驱动的方法。该方法首先创建一个三维优化空间来表示资源分配配置,而后通过搜索该空间为应用生成多种并行机制,最后对各种并行机制进行静态评估从而找出最优的机制。该方法同时考虑了任务并行、数据并行以及通讯开销。在异构多核处理器Cell上利用一个图像处理应用对此方法进行了测试。实验表明.这种模型驱动的方法能够很好地评估性能并为应用确定有效的并行机制。
To fully utilize the multi-level parallelism delivered by heterogeneous multi-core processors and solve the problem of multi-core resource allocation, this paper presents a model-driven technique for mapping multi-task parallel programs onto multi-core platforms. When using the technique, the resource allocation configurations are expressed within a three-di- mensional optimization space which affects the task-level parallelism and the data-level parallelism, and comJnunication. Then, the performance of each valid parallelization scheme is estimated statically, taking both the computation cost and the communication cost into account. The experiment on the approach on the Cell BE using an image processing application. The experimental results show that the proposed model-driven method can correctly predict the overall performance and highlight the most efficient parallelization schemes.