A Comprehensive Memory Analysis of Data Intensive Workloads on Server Class Architecture

The emergence of data analytics frameworks requires computational resources and memory subsystems that can naturally scale to manage massive amounts of diverse data. Given the large size and heterogeneity of the data, it is currently unclear whether data analytics frameworks will require high performance and large capacity memory to cope with this change and exactly what role main memory subsystems will play; particularly in terms of energy efficiency. In this paper, we investigate how the choice of DRAM (high-end vs low-end) impacts the performance of Hadoop, Spark, and MPI based Big Data workloads in the presence of different storage types on a local cluster. Our results show that Hadoop workloads do not require high capacity memory. However, Spark and MPI based workloads require large capacity memory. Moreover, Increasing memory bandwidth through the increasing memory frequency or the number of channels does not improve the performance of Hadoop workloads while iterative tasks in Spark and MPI benefits from high bandwidth memory. Among the configurable parameters, our results indicate that increasing the number of DRAM channels reduces DRAM power and improves the energy-efficiency across all applications.