论文部分内容阅读
互联网和物联网技术的飞速发展开启了“大数据”时代。目前，硬件的高速发展催生了许多异构芯片，它们越来越多地出现在大规模数据中心里，支持不同的应用程序，在提升性能的同时降低整体功耗。文章重点介绍了基于 MapReduce编程模型的 Hadoop+框架的设计与实现，它允许用户在单个任务中调用 CUDA/OpenCL的并行实现，并能通过异构任务模型帮助用户。在我们的实验平台上，五种常见机器学习算法使用 Hadoop+框架相对于 Hadoop能达到1.4×～16.1×的加速比，在 Hadoop+框架中使用异构任务模型指导其资源分配策略，对单个应用负载上最高达到36.0%的性能提升；对多应用的混合负载，最多能减少36.9%，平均17.6%的应用执行时间。“,”The rapid development of Internet and Internet of Things opens the era of big data. Currently, heterogeneous architectures are being widely adopted in large-scale datacenters, for the sake of performance improvement and reduction of energy consumption. This paper presents the design and implementation of Hadoop+, a programming framework that implements MapReduce and enables invocation of parallelized CUDA/OpenCL within a map/reduce task, and helps the user by taking advantage of a heterogeneous task model. Experimental result shows that Hadoop+ attains 1.4× to 16.1× speedups over Hadoop for five commonly used machine learning algorithms. Coupled with a heterogeneous task model that helps allocate computing resouces, Hadoop+ brings a 36.0% improvement in data processing speed for single-application workloads, and for mixed workloads of multiple applications, the execution time is reduced by up to 36.9%with an average 17.6%.