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.

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.

Source code

SENSEI is open source and freely available on github at https://github.com/SENSEI-insitu/SENSEI

Data model

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

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.

In situ data processing

SENSEI comes with a number of ready to use in situ processing options. These include:

User defined in situ processing

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.

For more information see our ISAV 2018 paper.

In transit data processing

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.

For more information see our EGPGV 2020 and ISAV 2018 papers.

The SENSEI End point

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.

Mini-apps

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++.

Data Model

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.

Support for Heterogeneous Architectures

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.

Build and Install

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.

Typical build procedure

$ mkdir build
$ cd build
$ ccmake .. # set one or more -D options as needed
$ make
$ make install

Build Options

For use with Ascent

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.

For use the Libsim

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.

For use with Catalyst

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.

For use with ADIOS 1

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.

For use with Python

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.

Enable writing to Visit ".visit" format or ParaView ".pvd" format

cmake -DENABLE_SENSEI=ON -DENABLE_VTK_IO=ON  -DVTK_DIR=[your path] ..

Can be used with either ParaView_DIR or VTK_DIR.

For use with experimental VTK-m analyses

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.

Using the SENSEI library from another project

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.