面向大数据复杂应用的GPU协同计算模型

大数据计算中存在流计算、内存计算、批计算和图计算等不同模式，各种计算模式有不同的访存、通信和资源利用等特征。GPU异构集群在大数据分析处理中得到广泛应用，然而缺少研究GPU异构集群在大数据分析中的计算模型。多核CPU与GPU协同计算时不仅增加了计算资源的密度，而且提高节点间和节点内的通信复杂度。为了从理论上研究GPU与多核CPU协同计算问题，面向多种计算模式建立一个多阶段的协同计算模型（p-DCOT）。p-DCOT以BSP大同步并行模型为核心，将协同计算过程分成数据层、计算层和通信层三个层次，并且延用DOT模型的矩阵来形式化描述计算和通信行为。通过扩展p-DOT模型描述节点内和节点间的协同计算行为，细化了负载均衡的参数并证明时间成本函数，最后用典型计算作业验证模型及参数分析的有效性。该协同计算模型可成为揭示大数据分析处理中协同计算行为的工具。

GPU collaborative computing model for complex applications in large-scale data processing

The large-scale data computig process includes different modes which are streaming computing, memory computing, batching computing and graph computing mode. Each mode has different access memory, communication and resource utilization. GPU clusters are widely used in large-scale data processing. However, there is lack of computing model for GPU cluster in large-scale data process. The collaborative computing between GPU and CPU not only increases density of computing resources but also improves communication complexity between inter-nodes and intra-nodes. To explore rule of collaborative computing process, this paper built a novel multi-stage collaborative computing model(p-DCOT) for multi-computing modes. This model was based on BSP model, which was large synchronous parallel model. It divided the collaborative computing process into three levels: data layer, computing layer and communication layer. This model also used matrix of DOT model to describe computing and communication behaviors. Then it extendd the p-DOT model to describe collaborative computing behavior within computing nodes and refined the parameters of workloads balancie. This model could prove the time cost function of collaborative computing model. Finally, the typical computing tasks verified the model validity and parameter analysis. This collaborative computing model will be a tool to reveal collaborative computing behaviors in large-scale data analysis processing.