The SENSEI project takes aim at a set of research challenges for enabling scientific knowledge discovery within the context of in situ processing at extreme-scale concurrency. This work is motivated by a widening gap between FLOPs and I/O capacity which will make full-resolution, I/O-intensive post hoc analysis prohibitively expensive, if not impossible.
We focus on new algorithms for analysis, and visualization - topological, geometric, statistical analysis, flow field analysis, pattern detection and matching - suitable for use in an in situ context aimed specifically at enabling scientific knowledge discovery in several exemplar application areas of importance to DOE. Complementary to the in situ algorithmic work, we focus on several leading in situ infrastructures, and tackle research questions germane to enabling new algorithms to run at scale across a diversity of existing in situ implementations.
Our intent is to move the field of in situ processing in a direction where it may ultimately be possible to write an algorithm once, then have it execute in one of several different in situ software implementations. The combination of algorithmic and infrastructure work is grounded in direct interactions with specific application code teams, all of which are engaged in their own R&D aimed at evolving to the exascale.
| Quick links |
|---|
| Project Organization |
| Build and Install |
| Using the SENSEI Library |
The SENSEI library contains core base classes that declare the AnalysisAdaptor API which is used to interface to in situ infrastructures and implement custom analyses; the DataAdaptor API which AnalysisAdaptors use to access simulation data in a consistent way; and a number of implementations of both. For more information see our SC16 paper.
SENSEI is open source and freely available on github at https://github.com/SENSEI-insitu/SENSEI
SENSEI makes use of a heavily stripped down and mangled version of the VTK 9.0.0 data model. The best source of documentation for SENSEI's data model is VTK itself (VTK doxygen).
Instrumenting a simulation typically involves creating and initializing an instance of sensei::ConfigurableAnalysis with an XML file and passing a simulation specific sensei::DataAdaptor when in situ processing is invoked.
| Class | Description |
|---|---|
| sensei::ConfigurableAnalysis | uses a run time provided XML file to slect and confgure one or more library specific data consumers or in transit transports |
| sensei::DataAdaptor | simulations implement an instance that packages simulation data into SENSEI's data model |
| sensei::SVTKDataAdaptor | An adaptor that can manage and serve SVTK data objects. Use this to return data from an analysis |
SENSEI comes with a number of ready to use in situ processing options. These include:
| Class | Description |
|---|---|
| sensei::AnalysisAdaptor | base class for in situ data processors |
| sensei::DataAdaptor | defines the API by which data processors fetch data from the simulation |
| sensei::Histogram | Computes histograms |
| sensei::AscentAnalysisAdaptor | Processes simulation data using Ascent |
| sensei::CatalystAnalysisAdaptor | Processes simulation data using ParaView Catalyst |
| sensei::LibsimAnalysisAdaptor | Processes simulation data using VisIt Libsim |
| sensei::Autocorrelation | Compute autocorrelation of simulation data over time |
| sensei::VTKPosthocIO | Writes simulation data to disk in a VTK format |
| sensei::VTKAmrWriter | Writes simulation data to disk in a VTK format |
| sensei::PythonAnalysis | Invokes user provided Pythons scripts that process simulation data |
| sensei::SliceExtract | Computes planar slices and iso-surfaces on simulation data |
| sensei::DataBinning | Transforms point based samples onto a Cartesian mesh |
A unique feature of SENSEI is the ability to invoke user provided code written in Python or C++ on a SENSEI instrumented simulation. This makes SENSEI much easier to extend and customize than other solutions.
| Class | Description |
|---|---|
| sensei::AnalysisAdaptor | used to invoke user defined C++ code. Override the sensei::AnalysisAdaptor::Execute method. |
| sensei::PythonAnalysis | used to invoke user defined Python code Implement an Execute function in Python |
For more information see our ISAV 2018 paper.
It is often advantageous to move data onto a seperate set of compute nodes for concurrent processing in a job seperate from the simulation. SENSEI supports this through run time configurable transports and the SENSEIEndPoint an application written in C++ that can be configured to receive and process data while simulation is running.
| Class | Description |
|---|---|
| sensei::AnalysisAdaptor | base class for the write side of the transport |
| sensei::DataAdaptor | base class for the read side of the transport |
| sensei::ConfigurableAnalysis | used to select and configure the write side of the transport at run time from XML |
| sensei::ConfigurableInTransitDataAdaptor | used to configure the read side of the transport at run time from XML |
| sensei::ADIOS2AnalysisAdaptor | The write side of the ADIOS 2 transport |
| sensei::ADIOS2DataAdaptor | The read side of the ADIOS 2 transport |
| sensei::HDF5AnalysisAdaptor | The write side of the HDF5 transport |
| sensei::HDF5DataAdaptor | The read side of the HDF5 transport |
| sensei::Partitioner | base class for data partitioner which maps data to MPI ranks as it is moved |
| sensei::ConfigurablePartitioner | Selects and configures one of the partitioners at run time from XML |
| sensei::BlockPartitioner | maps blocks to ranks such that consecutive blocks share a rank |
| sensei::PlanarPartitioner | Maps blocks to MPI ranks in a round robbin fassion |
| sensei::MappedPartitioner | Maps blocks to MPI ranks using a run time user provided mapping |
| sensei::IsoSurfacePartitioner | Maps blocks to MPI ranks such that blocks not intersecting the iso surface are excluded |
| sensei::PlanarSlicePartitioner | Maps blocks to MPI ranks such that blocks not intersecting the slice are excluded |
For more information see our EGPGV 2020 and ISAV 2018 papers.
End points are programs that receive and analyze simulation data through transport layers such as ADIOS and LibIS. The SENSEI end point uses the transport's data adaptor to read data being serialized by the transport's analysis adaptor and pass it back into a SENSEI analysis for further processing.
SENSEI ships with a number of mini-apps that demonstrate use of the SENSEI library with custom analyses and the supported in situ infrastructures. When the SENSEI library is enabled the mini-apps will make use of any of the supported in situ infrastructures that are enabled. When the SENSEI library is disabled mini-apps are restricted to the custom analysis such as histogram and autocorrelation.
More information on each mini-app is provided in the coresponding README in the mini-app's source directory.
-
Oscillators The miniapp from year II generates time varying data on a uniform mesh and demonstrates usage with in situ infrasturctures, histogram, and autocorrelation analyses.
-
Newton++ A complete re-write of the original in C++ using OpenMP target offload for platform portable acceleration. This mini app demonstrates zero-copy data transfer from GPUs and accelerators.
-
Newton Python This Python n-body miniapp demonstrates usage of in situ infrastructures and custom analyses from Python. This has been replaced by the C++ port Newton++.
SENSEI makes use of a fork of VTK 9.0.0 that has been mangled and minified. Minification has removed all source code from our fork except for the data model and its dependencies. This substantially reduces the overheads associated with VTK providing only the features we need for our data model. VTK filters and I/O can be used by pointing the build to an install of standard VTK as released by Kitware. Mangling changed the character sequences VTK and vtk to SVTK and svtk this allows for interoberability with VTK as released by Kitware.
Extensions to SENSEI's execution and data model have been introduced to support in situ on heterogeneous architectures. Extensions to the data model make zero-copy transfer of accelerator backed memory possible. Simply use svtkHAMRDataArray when passing array based data in your data adaptor and initialize it using one of the zero-copy constructors. Extensions to SENSEI's execution model provide placement controls as well as control over synchronous or asynchronous execution method. The controls for our execution model extensions are accessed through the APIs defined in sensei::AnalysisAdaptor or via XML attributes defined in sensei::ConfigurableAnalysis.
For more details including source code examples, XML, and a demonstration at scale on NERSC's Cray NVIDIA system Perlmutter please see our SC23 paper.
The SENSEI project uses CMake 3.0 or later. The CMake build options allow you to choose which of the mini-apps to build as well as which frameworks to enable. It is fine to enable multiple infrastructures, however note that Catalyst and Libsim are currently mutually exclusive options due to their respective use of different versions of VTK.
$ mkdir build
$ cd build
$ ccmake .. # set one or more -D options as needed
$ make
$ make install| Build Option | Default | Description |
|---|---|---|
SENSEI_ENABLE_CUDA |
OFF | Enables CUDA accelerated codes. Requires compute capability 7.5 and CUDA 11 or later. |
SENSEI_ENABLE_PYTHON |
OFF | Enables Python bindings. Requires VTK, Python, Numpy, mpi4py, and SWIG. |
SENSEI_ENABLE_CATALYST |
OFF | Enables the Catalyst analysis adaptor. Depends on ParaView Catalyst. Set ParaView_DIR. |
SENSEI_ENABLE_CATALYST_PYTHON |
OFF | Enables Python features of the Catalyst analysis adaptor. |
SENSEI_ENABLE_ASCENT |
OFF | Enables the Ascent analysis adaptor. |
SENSEI_ENABLE_ADIOS1 |
OFF | Enables ADIOS 1 adaptors and endpoints. Set ADIOS_DIR. |
SENSEI_ENABLE_HDF5 |
OFF | Enables HDF5 adaptors and endpoints. Set HDF5_DIR. |
SENSEI_ENABLE_LIBSIM |
OFF | Enables Libsim data and analysis adaptors. Requires Libsim. Set VTK_DIR and LIBSIM_DIR. |
SENSEI_ENABLE_VTK_IO |
OFF | Enables adaptors to write to VTK XML format. |
SENSEI_ENABLE_VTK_MPI |
OFF | Enables MPI parallel VTK filters, such as parallel I/O. |
SENSEI_ENABLE_VTKM |
ON | Enables analyses that use VTKm directly instead of via VTK. |
SENSEI_ENABLE_OSCILLATORS |
ON | Enables the oscillators mini-app. |
VTK_DIR |
Set to the directory containing VTKConfig.cmake. | |
ParaView_DIR |
Set to the directory containing ParaViewConfig.cmake. | |
ADIOS_DIR |
Set to the directory containing ADIOSConfig.cmake | |
LIBSIM_DIR |
Path to libsim install. |
cmake -DENABLE_ASCENT=ON -DVTKM_DIR=[your path] -DVTKH_DIR=[your path] \
-DCONDUIT_DIR=[your path] -DAscent_DIR=[your path] -DVTK_DIR=[your path] \
..Note that the VTK build needs to explicitly disable use of VTK-m as this will conflict with the version required by Ascent. We used the instructions for building Ascent and its dependencies (VTK-m, VTK-h, Conduit, etc) manually as described in the Ascent documentation.
cmake -DENABLE_SENSEI=ON -DENABLE_LIBSIM=ON -DVTK_DIR=[your path] -DLIBSIM_DIR=[your path] ..VTK_DIR should point to the VTK used by Libsim.
cmake -DENABLE_SENSEI=ON -DENABLE_CATALYST=ON -DParaView_DIR=[your path] ..Optionally, -DENABLE_CATALYST_PYTHON=ON will enable Catalyst Python scripts.
Note that a development version of ParaView is required when building with
both SENSEI_ENABLE_CATALYST and SENSEI_ENABLE_VTKM are enabled as released versions of
ParaView (5.5.2 and earlier) do not include a modern-enough version of vtk-m.
cmake -DENABLE_SENSEI=ON -DENABLE_ADIOS1=ON -DVTK_DIR=[your path] -DADIOS_DIR=[your path] ..Can be used with either ParaView_DIR when configuring in conjunction with
Catalyst, or VTK_DIR otherwise.
In essence this is as simple as adding -DENABLE_PYTHON=ON -DSENSEI_PYTHON_VERSION=3
However, VTK (or ParaView when used with Catalyst) needs to be built with
Python enabled and the SENSEI build needs to use the same version; and NumPy,
mpi4py, and SWIG are required. Note that there are some caveats when used with
Catalyst and Libsim. These are described in more detail in the Newton mini app
README.
cmake -DENABLE_SENSEI=ON -DENABLE_VTK_IO=ON -DVTK_DIR=[your path] ..Can be used with either ParaView_DIR or VTK_DIR.
cmake -DENABLE_SENSEI=ON -DENABLE_VTKM=ON -DVTK_DIR=[your path] ..Note that a development version of VTK is required when building with
both SENSEI_ENABLE_SENSEI and SENSEI_ENABLE_VTKM are enabled as released versions of
VTK (8.1.1 and earlier) do not include a modern-enough version of vtk-m.
To use SENSEI from your CMake based project include the SENSEI CMake config in your CMakeLists.txt.
find_package(SENSEI REQUIRED)
add_executable(myexec ...)
target_link_libraries(myexec sensei ...)Additionally, your source code may need to include senseiConfig.h to capture
compile time configuration.