skip to: onlinetools | mainnavigation | content | footer

XPRESS

Ron Brightwell, Sandia
Coordinating PI

Thomas Sterling, IU
Chief Scientist

Kelsey Shephard, IU
Project Manager

XPRESS

Software

APEX

We have developed a software of the APEX prototype with initial versions of OpenX. It is engineered to work with both HPX-3 and HPX-5, and integrates RCR Toolkit and TAU. Both RCR Toolkit and TAU were updated to work with APEX and support certain capabilities that APEX required. APEX has been ported to several platforms, including Edison system at NERSC. APEX is included in the XPRESS software releases. Ongoing work with APEX will interface with XPI, enhance LXK introspection, and implement more sophisticated policies, especially for multi-objective control. APEX is currently being evaluated with XPRESS project applications.

HPX-3

HPX-3 is a general purpose C++ runtime system for parallel and distributed applications of any scale. The most recent version can be downloaded here.

HPX-5

The HPX-5 runtime system directly reflects the semantic constructs of the underlying ParalleX execution model, being comprised of an active global address space, lightweight user thread management, parcels and message-driven computation, distributed processes, local control objects, and continuations. HPX-5 has focused on understanding low-level details of efficient and scalable exascale runtime implementation, including interfaces and interactions with operating systems and with hardware architecture. The initial implementation of HPX-5 relied on a two-sided, parcel-based model of communication. However, a fast, low-latency, one-sided communication mode is the current default. HPX-5 supports dynamic, distributed processes with determination detection and a flat, byte-addressable, global address space using native RDMA transport. Our implementation also supports a software based active global address space (AGAS) that allows explicit movement of blocks within the global address space. HPX-5 can be downloaded here.

Kitten Lightweight Kernel

The Kitten lightweight kernel1 is the basis for the development of the Lightweight eXascale Kernel (LXK) in the XPRESS project. Kitten continues to be actively developed as part of the XPRESS and Hobbes projects. As a research prototype, no formal releases of Kitten are made. The latest Kitten source can be obtained from Github here.

Open MPI

University of Houston's research team are long-time contributors to the Open MPI software project. For the XPRESS project, they have added support to use HPX as a runtime layer for Open MPI. It currently uses the HPX runtime environment for startup and communication through the TCP/IP transport module, although some details with respect to the integration of memory allocators and the internal utilization of threads in Open MPI still have to be worked out. Support for other network transports, specifically InfiniBand and shared memory, are currently being actively worked on. Open MPI is also able to use the native SLURM startup through HPX. Open MPI can be dowloaded here.

OMPTX

The University of Houston team developed OMPTX, an OpenMP runtime based on HPX which shares the same ABI as the newly open-sourced Intel OpenMP runtime. This runtime has been adopted in existing compilers from Intel, GNU, PathScale, and Clang, and support within OpenUH is underway. The ability to use OMPTX as a drop-in replacement for the Intel OpenMP runtime facilitates experimentation in running OpenMP programs with HPX. The Intel OpenMP Runtime interface and a number of the compilers which use it support OpenMP 4.0, allowing us to experiment with implementation of newer OpenMP features within the proposed software infrastructure.

Profugus

The Profugus mini-application (mini-app)2 is a set of computational kernels extracted from the Exnihilo particle transport code package at Oak Ridge National Laboratory (ORNL)3. Exnihilo provides advanced transport solvers and associated front-ends that are applicable to a wide-range of nuclear technology applications including reactor neutronics, shielding, criticality safety, dosimetry, fusion component analysis, and others. The resulting problem-specific complexity in the various front-ends, the dependence on a wide range of third-party libraries (TPL), and, most imporantly, export control makes Exnihilo a difficult platform to share among computer scientists and hardware engineers. To overcome these limitations and to provide a means to collaborate across fields, we have developed the Profugus mini-app. The goal of Profugus is to provide open-source kernels that effectively capture the algorithmic features of the full Exnihilo applications. In this manner more effective connections can be established between applications developers and Xpress researchers.

TAUdb

We use the TAUdb performance database to store performance results from experiments. However, the databases are primarily for internal use only. TAUdb is part of the TAU project that can be downloaded from here.

Web site contact: rbbrigh@sandia.gov