Can Non-volatile Memory Benefit MapReduce Applications on HPC Clusters?

Abstract

Modern High-Performance Computing (HPC) clusters are equipped with advanced technological resources that need to be properly utilized to achieve supreme performance for end applications. One such example, Non-Volatile Memory (NVM), provides the opportunity for fast scalable performance through its DRAM-like performance characteristics. On the other hand, distributed processing engines, such as MapReduce, are continuously being enhanced with features enabling high-performance technologies. In this paper, we present a novel MapReduce framework with NVRAM-assisted map output spill approach. We have designed our framework on top of the existing RDMA-enhanced Hadoop MapReduce to ensure both map and reduce phase performance enhancements to be present for end applications. Our proposed approach significantly enhances map phase performance proven by a wide variety of MapReduce benchmarks and workloads from Intel HiBench [9] and PUMA [18] suites. Our performance evaluation illustrates that NVRAM-based spill approach can improve map execution performance by 2.73x which contributes to the overall execution improvement of 55% for Sort. Our design also guarantees significant performance benefits for other workloads: 54% for TeraSort, 21% for PageRank, 58% for SelfJoin, etc. To the best of our knowledge, this is the first approach towards leveraging NVRAM in MapReduce execution frameworks for applications on HPC clusters.