分布式流体系结构及其编程模型与资源管理

利用互联网资源提供大数据计算服务面临着资源异构性、动态性与通信长延迟等方面的挑战,现有分布式计算模型仍存在一些不足。运用流计算模型提出分布式流体系结构,包括分布式流编程模型与资源管理等,能够高效支持多种并行执行模式。在10个CPU-GPU异构结点上实现了原型系统,仿真实验验证了7个不同的测试用例。实验结果表明,与本地串行计算相比,分布式流体系结构可以平均提高39倍计算性能,具有较大的应用潜力。

The Programming Model and Resource Management of Distributed Stream Architecture

While providing big data computing services using Internet resources, there remains a big challenge to researchers, including heterogeneity of Internet resources, dynamics of Internet resources and long latency of the Internet communication. However, current influent distributed computing models still have some shortage. Therefore, a novel distributed stream computing model is proposed based on the traditional stream computing model, including the distributed stream programming model and resource management. They can efficiently support multiple parallel execution modes. The prototype system has been implemented on the 10 CPU-GPU heterogeneous nodes. Seven different benchmarks are used in the simulation experiment. The experimental result shows that the distributed stream architecture can achieve the speedup of at least 39x on average over the local serial computing, with significant potential for applications.