Towards Java-based HPC using the MVAPICH2 Library: Early Experiences

Abstract

There has been sporadic interest in using Java for High Performance Computing (HPC) in the past. These earlier efforts have resulted in several Java Message Passing Interface (MPI) [1] libraries including mpiJava [2], FastMPJ [3], MPJ Express [4], and Java Open MPI [5]. In this paper, we present our efforts in designing and implementing Java bindings for the MVAPICH2 [6] library. The MVAPICH2 Java bindings (MVAPICH2-J) follow the same API as the Java Open MPI library. MVAPICH2-J also provides support for communicating direct New I/O (NIO) ByteBuffers and Java arrays. Direct ByteBuffers reside outside JVM heaps and are not subject to the garbage collection. The library implements and utilizes a buffering layer to explicitly manage memory to avoid creating buffers every time a Java array message is communicated. In order to evaluate the performance of MVAPICH2-J and other Java MPI libraries, we also designed and implemented OMB-J that is a Java extension to the popular OSU Micro-Benchmarks suite (OMB) [7]. OMB-J currently supports a range of bench-marks for evaluating point-to-point and collective communication primitives. We also added support for communicating direct ByteBuffers and Java arrays. Our evaluations reveal that at the OMB-J level, ByteBuffers are superior in performance due to the elimination of extra copying between the Java and the Java Native Interface (JNI) layer. MVAPICH2-J achieves similar performance to Java Open MPI for ByteBuffers in point-to-point communication primitives that is evaluated using latency and bandwidth benchmarks. For Java arrays, there is a slight overhead for MVAPICH2-J due to the use of the buffering layer. For the collective communication benchmarks, we observe good performance for MVAPICH2-J. Where, MVAPICH2-J fairs better than Java Open MPI with ByteBuffers by $a$ factor of 6.2 and 2.76 for broadcast and all reduce, respectively, on average for all messages sizes. And, using Java arrays, $2. 2\times$ and $1. 62\times$ on average for broadcast and allreduce, respectively. The collective communication performance is dictated by the performance of the respective native MPI libraries.