项目名称: 基于学习的复杂并行绘制系统负载平衡算法研究
项目编号: No.61472261
项目类型: 面上项目
立项/批准年度: 2015
项目学科: 计算机科学学科
项目作者: 张严辞
作者单位: 四川大学
项目金额: 80万元
中文摘要: 并行绘制系统是解决高端训练仿真系统,基于交互的娱乐系统等大型实时应用对计算能力的要求和单个GPU计算能力之间的矛盾的根本方法。本课题针对包含大量复杂算法和突发绘制任务的并行绘制系统,从系统层面研究其负载平衡算法。与以往负载平衡算法主要依赖检测负载失衡+补偿操作恢复负载平衡这种事后补救的方式不同,本课题尝试从系统的多次反复运行中,通过离线机器学习的方式,积累绘制任务负载估计的知识,最终突破以往负载平衡算法关于在本帧绘制完成之前无法准确知道其绘制开销的假设,实现系统在任意时刻的负载平衡。在实现如上目标的过程中,需要解决两个核心问题:1)如何从所有的可能因素中找出影响负载的主要因素;2)如何确定这些主要因素和负载之间的函数关系。我们试图通过机器学习中的粗糙集技术来探索影响负载的主要因素,并通过人工神经网络技术来表达这些因素和负载之间的函数关系。
中文关键词: 并行绘制;负载平衡;机器学习;粗糙集;人工神经网络
英文摘要: Parallel rendering system is the fundamental way to solve the conflict between the requirements of the computation power of large real-time applications (such as high-end training and simulation system, interactive entertainment system and so on) and the computation power of single GPU. The project is aim to address the load balancing problem on system level for complicated parallel rendering system containing multiple algorithms and unpredictable rendering tasks. Different from the traditional load balancing algorithm which is based on the mechanism of detecting load unbalance + doing something to restore the load balancing, our project tries to accumulate the acknowledge of load estimation from multiple times of system running by an off-line machine learning process. The proposed method can break the assumption that the rendering cost cannot be accurately estimated before the rendering is finished and achieve load-balancing at any time during the system running. Two fundamental issues have to be addressed to achieve the above goal: 1) How to extract the main factors from all possible factors which might influence the workloads of rendering tasks? 2) How to determine the internal relationships between the workload and these main factors? In this project, we attempt to use the rough set theory to address the first problem, and employ the artificial neural network to represent the relationships between the workload and the factors extracted from the first problem.
英文关键词: Parallel Rendering;Load Balance;Machine Learning;Rough Set;Artificial Neural Network