Abstract: MPJ Express is a Java messaging system that implements an MPI-like interface. It is used for writing parallel Java applications on High Performance Computing (HPC) hardware including commodity clusters. The software is capable of executing in multicore and cluster mode. In the cluster mode, it currently supports Ethernet and Myrinet based interconnects and provide specialized communication devices for these networks. One recent trend in distributed memory parallel hardware is the emergence of InfiniBand interconnect, which is a high-performance proprietary network and provides low latency and high bandwidth for parallel MPI applications. Currently there is no direct support available in Java (and hence MPJ Express) to exploit the performance benefits of InfiniBand networks. The only option to run distributed Java programs over InfiniBand networks is to rely on TCP/IP emulation layers like IP over InfiniBand (IPoIB) and Sockets Direct Protocol (SDP), which provide poor communication performance. To tackle this issue in the context of MPJ Express, this paper presents a low-level communication device called ibdev that can be used to execute parallel Java applications on InfiniBand clusters. MPJ Express is based on a layered architecture and hence users can opt to use ibdev at runtime on an InfiniBand equipped commodity cluster. ibdev improves Java application performance with access to InfiniBand hardware using native verbs API. Our performance evaluation reveals that MPJ Express achieves much better latency and bandwidth using this new device, compared to IPoIB and SDP. Improvement in communication performance is also evident in NAS parallel benchmark results where ibdev helps MPJ Express achieve better scalability and speedups as compared to IPoIB and SDP. The results show that it is possible to reduce the performance gap between Java and native languages with efficient support for low level communication libraries.

Abstract: Some of the critical hurdles for the widespread adoption of accelerators in high performance computing are portability and programming difficulty. To be an effective HPC platform, these systems need a high level software development environment to facilitate the porting and development of applications, so they can be portable and run efficiently on either accelerators or CPUs. In this paper we present a high level parallel programming environment for accelerator-based systems, which consists of tightly coupled compilers, tools, and libraries that can interoperate and hide the complexity of the system. Ease of use is possible with compilers making it feasible for users to write applications in Fortran, C, or C++ with OpenACC directives, tools to help users port, debug, and optimize for both accelerators and conventional multi-core CPUs, and with auto-tuned scientific libraries.

Abstract: Relative debugging is a useful technique for locating errors that emerge from porting existing code to new programming language or to new computing platform. Recent attention on the UPC programming language has resulted in a number of conventional parallel programs, for example MPI programs, being ported to UPC. This paper gives an overview on the data distribution concepts used in UPC and establishes the challenges in supporting relative debugging technique for UPC programs that run on large supercomputers. The proposed solution is implemented on an existing parallel relative debugger ccdb, and the performance is evaluated on a Cray XE6 system with 16,348 cores.