A C++ Template Library for Distributed Data Structures with Support for Hierarchical Locality for HPC and Data-Driven Science.
Exascale systems are scheduled to become available in 2018-2020 and will be characterized by extreme scale and a multilevel hierarchical organization.
Efficient and productive programming of these systems will be a challenge, especially in the context of data-intensive applications. Adopting the promising notion of Partitioned Global Address Space (PGAS) programming the DASH project develops a data-structure oriented C++ template library that provides hierarchical PGAS-like abstractions for important data containers (multidimensional arrays, lists, hash tables, etc.) and allows a developer to control (and explicitly take advantage of) the hierarchical data layout of global data structures.
In contrast to other PGAS approaches such as UPC, DASH does not propose a new language or require compiler support to realize global address space semantics. Instead, operator overloading and other advanced C++ features are used to provide the semantics of data residing in a global and hierarchically partitioned address space based on a runtime system with one-sided messaging primitives provided by MPI or GASNet.
As such, DASH can co-exist with parallel programming models already in widespread use (like MPI) and developers can take advantage of DASH by incrementally replacing existing data structures with the implementation provided by DASH. Efficient I/O directly to and from the hierarchical structures and DASH-optimized algorithms such as map-reduce are also part of the project. Two applications from molecular dynamics and geoscience are driving the project and are adapted to use DASH in the course of the project.
DASH is funded by the German Research Foundation (DFG) under the priority programme “Software for Exascale Computing - SPPEXA” (2013-2018).
- HTTPS: [https://github.com/dash-project/dash.git]()
- SSH: [[email protected]:dash-project/dash.git]()
- Slack channel: https://dash-project.slack.com
- Support: [email protected]
- Developer mailing list: [email protected]
See guidelines in CONTRIBUTING.md.
DASH installations are available as Docker containers or build from source using CMake.
For pre-build Docker container images, see the DASH project on Docker Hub.
Building from Source
DASH is build using CMake.
Build scripts are provided for typical DASH configurations and can serve as starting points for custom builds:
|Script file name||Description|
||Standard release build|
||Development / debug build|
||Release build for Intel MIC (Xeon Phi)|
- CMake version 2.8.5 or greater (3.0.0 or greater recommended)
- C compiler supporting the C99 standard
- C++ compiler supporting the C++11 standard
Optional third-party libraries directly supported by DASH:
- BLAS implementation like Intel MKL, ATLAS
Building DASH from Source
To build the DASH project using CMake with default build settings, run:
(dash/)$ cmake --build .
Or, specify a new directory used for the build:
(dash/)$ mkdir build && cd ./build
For a list of available CMake parameters:
(build/)$ cmake .. -L
Parameters can be set using
-D flags. As an example, these parameters
will configure the build process to use icc as C compiler instead of the
(build/)$ cmake -DCMAKE_C_COMPILER=icc ..
To configure build parameters using ccmake:
(build/)$ ccmake ..
1. Choosing a DASH runtime (DART)
DASH provides the following variants:
- MPI: the Message Passing Interface
- CUDA: nNvidia’s Compute Unified Device Architecture (contributor distribution only)
- SHMEM: Symmetric Hierarchical Memory access (contributor distribution only)
The build process creates the following libraries:
By default, DASH is configured to build all variants of the runtime. You can specify which implementations of DART to build using the cmake option
(build/)$ cmake -DDART_IMPLEMENTATIONS=mpi,shmem ...
Programs using DASH select a runtime implementation by linking against the respective library.
Specifying the MPI implementation for the DART-MPI runtime
The most reliable method to build DART-MPI with a specific MPI installation
is to specify the CMake options
(build/) $ cmake -DMPI_C_COMPILER=/path/to/mpi/bin/mpicc \ -DMPI_CXX_COMPILER=/path/to/mpi/bin/mpiCC \ ...
3. Examples and Unit Tests
Source code of usage examples of DASH are located in
Examples each consist of a single executable and are built by default.
Binaries from examples and unit tests are deployed to the build direcory
but will not be installed.
To disable building of examples, specify the cmake option
(build/)$ cmake -DBUILD_EXAMPLES=OFF ...
To disable building of unit tests, specify the cmake option
(build/)$ cmake -DBUILD_TESTS=OFF ...
The example applications are located in the bin/ folder in the build directory.
The default installation path is
$HOME/opt as users on HPC systems
typically have install permissions in their home directory only.
To specify a different installation path, use
(build/)$ cmake -DINSTALL_PREFIX=/your/install/path ../
-DINSTALL_PREFIX=<DASH install path> can also be given in Step 1.
The installation process copies the ‘bin’, ‘lib’, and ‘include’ directories in the build directory to the specified installation path.
(dash/)$ cmake --build . --target install
Or manually using make:
(build/)$ cmake <build options> ../ (build/)$ make (build/)$ make install
Running DASH Applications
With the MPI variant, applications are spawn by MPI:
$ mpirun <MPI args> <app>-mpi
For CUDA and SHMEM, use
$ dartrun-cuda <dartrun-args> <app>-cuda
$ dartrun-shmem <dartrun-args> <app>-shmem
Launch the DASH unit test suite using
(dash/shmem/bin/)$ dartrun-shmem <dartrun args> dash-test-shmem <gtest args>
(dash/mpi/bin/)$ mpirun <MPI args> dash-test-mpi <gtest args>
For example, you would all unit tests of matrix data structures on 4 units using the MPI runtime with:
(dash/mpi/bin/)$ mpirun -n 4 dash-test-mpi --gtest_filter="MatrixTest*"
or all tests except for the Array test suite:
(dash/mpi/bin/)$ mpirun -n 4 dash-test-mpi --gtest_filter="-ArrayTest*"
Profiling DASH Applications using IPM
LD_PRELOAD to run a DASH application built with the DART-MPI backend:
$ LD_PRELOAD=/$IPM_HOME/lib/libipm.so mpirun -n <nproc> <DASH executable>
Available options for IPM are documented in the IPM user guide.
The DASH project homepage: http://www.dash-project.org
The Munich Network Management homepage: http://www.mnm-team.org