A Feasibility Study for MPI over HDFS

Abstract

With the increasing prominence of integrating highperformance computing (HPC) with big-data (BIGDATA) processing, running MPI over the Hadoop Distributed File System (HDFS) offers a promising approach for delivering better scalability and fault tolerance to traditional HPC applications. However, it comes with challenges that discourage such an approach: (1) two-sided MPI communication to support intermediate data processing, (2) a focus on enabling N-1 writes that is subject to the default HDFS block-placement policy, and (3) a pipelined writing mode in HDFS that cannot fully utilize the underlying HPC hardware. So, while directly integrating MPI with HDFS may deliver better scalability and fault tolerance to MPI applications, it will fall short of delivering competitive performance. Consequently, we present a performance study to evaluate the feasibility of integrating MPI applications to run over HDFS. Specifically, we show that by aggregating and reordering intermediate data and coordinating computation and 110 when running MPI over HDFS, we can deliver up to 1.92x and 1.78x speedup over MPI I/O and HDFS pipelined-write implementations, respectively. Consequently, we present a performance study to evaluate the feasibility of integrating MPI applications to run over HDFS. Specifically, we show that by aggregating and reordering intermediate data and coordinating computation and 110 when running MPI over HDFS, we can deliver up to 1.92x and 1.78x speedup over MPI I/O and HDFS pipelined-write implementations, respectively.