大数据负载的体系结构特征分析

针对大数据离线分析类和交互式查询类负载,首先对这些负载的一些共性进行分析,提取出公共操作集,并对它们进行分组整理；然后在大数据平台上测试这些负载运行过程中的微体系结构特征,采用PCA和SimpleKMeans算法对这些体系结构特征参数进行降维和聚类处理。实验分析结果表明负载之间有公共的操作集,如Join和Cross Production；有些负载有相似的属性,如Difference和Projection共享相同的微体系结构特征。实验结果对于 处理器等硬件平台的设计以及应用程序的优化具有指导性的意义,并且为大数据基准测试平台的设计提供了参考。

Analysis of Architecture Characteristics of Big Data Workloads

Aiming at the big data workloads with off-line analysis and interactive queries,first we analyzed some common features of these workloads,extracted the common set of operations and arranged the workloads in groups.Then,we tested the big data workloads on the BigDataBench platform and got the micro-architecture characteristics using PCA and SimpleKMeans algorithm for dimensionality reduction and clustering analysis.Our study revealed that big data workloads share a common set of operations such as the Join and Cross Production.We also observed that some of the big data workloads have many similar features.For example,the Difference and Projection operations share micro-architectural characteristics.The result of our experiment has a guiding significance for the design of hardware platforms like processors and the optimization of applications.Meanwhile,it also provides valuable insights into the implementation of the big data benchmark platform.