Releases: multiscale/muscle3
Release 0.7.2
MUSCLE3 0.7.2
MUSCLE3 is the third incarnation of the Multiscale Coupling Library and Environment, and the successor to MUSCLE 2. Its purpose is to make creating coupled multiscale simulations easy, and to then enable efficient Uncertainty Quantification of such models using advanced semi-intrusive algorithms.
MUSCLE3 uses the Multiscale Modelling and Simulation Language (MMSL) to describe the structure of a multiscale model. MMSL can be expressed in the form of a diagram (gMMSL; not yet implemented) or as a YAML file (yMMSL; this is convenient both for people and for software). The MMSL lets one describe which components (submodels, scale bridges, data converters, UQ components, etc.) a multiscale model consist of, how many instances of each we need, and how they are wired together.
MUSCLE3 is intended to scale from your laptop to the exascale. At the low end, it supports a non-distributed but parallel mode in which an entire multiscale simulation, including all component implementations and the MMSL configuration, is in a single (short) Python file. Next is a distributed mode where the manager and component instances run on multiple nodes in a cluster, and communicate directly with one another. Beyond that, additional components and optimisations are envisioned that would allow scaling to huge machines or combinations of multiple machines. Our goal is to make the transitions between these modes as smooth as possible, so that extra compute power can be added gradually, as needed.
Added
- Support for Python 3.12
- Support for NumPy 2 (already working, constraint removed)
Improved
- Fixed
SHUTDOWN_WAIT
profiling event for Python components - Report successful completion on successful completion (instead of an error)
- Fixed crash on SLURM clusters with hyperthreading
- Use psutil instead of netifaces, which is no longer maintained
- Internal improvements and refactoring to support upcoming fixes and features
Removed
- Support for Python 3.7
- Warnings about ruamel.yaml, via ymmsl/YAtiML update
Release 0.7.1
MUSCLE3 0.7.1
MUSCLE3 is the third incarnation of the Multiscale Coupling Library and Environment, and the successor to MUSCLE 2. Its purpose is to make creating coupled multiscale simulations easy, and to then enable efficient Uncertainty Quantification of such models using advanced semi-intrusive algorithms.
MUSCLE3 uses the Multiscale Modelling and Simulation Language (MMSL) to describe the structure of a multiscale model. MMSL can be expressed in the form of a diagram (gMMSL; not yet implemented) or as a YAML file (yMMSL; this is convenient both for people and for software). The MMSL lets one describe which components (submodels, scale bridges, data converters, UQ components, etc.) a multiscale model consist of, how many instances of each we need, and how they are wired together.
MUSCLE3 is intended to scale from your laptop to the exascale. At the low end, it supports a non-distributed but parallel mode in which an entire multiscale simulation, including all component implementations and the MMSL configuration, is in a single (short) Python file. Next is a distributed mode where the manager and component instances run on multiple nodes in a cluster, and communicate directly with one another. Beyond that, additional components and optimisations are envisioned that would allow scaling to huge machines or combinations of multiple machines. Our goal is to make the transitions between these modes as smooth as possible, so that extra compute power can be added gradually, as needed.
Added
- Support for Python 3.11 (working already, now official)
- Enabled type checking support for the libmuscle Python API
Improved
- Easier crash debugging due to improved root cause detection
- Fixed crash in profiling timeline plot
- Better performance of timeline plot
- Better visual quality of timeline plot
- Improved profiling of shutdown process
- Fixed crash in profiler for large simulations
- Fixed several (harmless) compiler warnings
- Small documentation rendering improvements
Thanks
- David for reporting many of these and submitting a fix too!
Release 0.7.0
MUSCLE3 0.7.0
MUSCLE3 is the third incarnation of the Multiscale Coupling Library and Environment, and the successor to MUSCLE 2. Its purpose is to make creating coupled multiscale simulations easy, and to then enable efficient Uncertainty Quantification of such models using advanced semi-intrusive algorithms.
MUSCLE3 uses the Multiscale Modelling and Simulation Language (MMSL) to describe the structure of a multiscale model. MMSL can be expressed in the form of a diagram (gMMSL; not yet implemented) or as a YAML file (yMMSL; this is convenient both for people and for software). The MMSL lets one describe which components (submodels, scale bridges, data converters, UQ components, etc.) a multiscale model consist of, how many instances of each we need, and how they are wired together.
MUSCLE3 is intended to scale from your laptop to the exascale. At the low end, it supports a non-distributed but parallel mode in which an entire multiscale simulation, including all component implementations and the MMSL configuration, is in a single (short) Python file. Next is a distributed mode where the manager and component instances run on multiple nodes in a cluster, and communicate directly with one another. Beyond that, additional components and optimisations are envisioned that would allow scaling to huge machines or combinations of multiple machines. Our goal is to make the transitions between these modes as smooth as possible, so that extra compute power can be added gradually, as needed.
Added
- Checkpointing is now supported in C++ and Fortran as well
- Added built-in profiling feature
- New object-oriented Fortran API (existing API also still available)
- New Instance.list_settings() function
- Build support for macOS with CLang and G++/GFortran
- Build support for Cray machines and compilers
Improved
- Compiling with MPI and linking without or vice versa is now impossible
- Fixed MessagePack build failure on old OSes
- Fixed resource allocation for instance sets
- Planner now detects F_INIT -> O_F loops and gives an error
- Manager correctly handles instances that never register
- Last lines of log now printer to screen on error for smoother problem solving
- Various small fixes and improvements
Backwards Incompatible changes
-
Instance.reuse_instance
no longer acceptsapply_overlay
argument. UseInstanceFlags.DONT_APPLY_OVERLAY
when creating the instance instead. -
LIBMUSCLE_Instance_create
signature has changed, this might lead to errors like:30 | instance = LIBMUSCLE_Instance_create(ports, MPI_COMM_WORLD, root_rank) | 1 Error: Type mismatch in argument ‘flags’ at (1); passed INTEGER(4) to TYPE(libmuscle_instanceflags)
You may provide an explicit
InstanceFlags()
argument, or use named arguments:instance = LIBMUSCLE_Instance_create(ports, LIBMUSCLE_InstanceFlags(), MPI_COMM_WORLD, root_rank) instance = LIBMUSCLE_Instance_create(ports, communicator=MPI_COMM_WORLD, root=root_rank)
-
DataConstRef
items can no longer be added to aData
containing a list or dict. The newly addedDataConstRef::list
andDataConstRef::dict
should be used instead.
Thanks
- Maarten at Ignition Computing for implementing much of the above
- Peter for debugging the MessagePack build issue
- Peter, Jon and Gavin for ARCHER2 access and support
- Koen for testing macOS build support
- Everyone who reported issues and contributed feature ideas!
Release 0.6.0
MUSCLE3 0.6.0
MUSCLE3 is the third incarnation of the Multiscale Coupling Library and Environment, and the successor to MUSCLE 2. Its purpose is to make creating coupled multiscale simulations easy, and to then enable efficient Uncertainty Quantification of such models using advanced semi-intrusive algorithms.
MUSCLE3 uses the Multiscale Modelling and Simulation Language (MMSL) to describe the structure of a multiscale model. MMSL can be expressed in the form of a diagram (gMMSL; not yet implemented) or as a YAML file (yMMSL; this is convenient both for people and for software). The MMSL lets one describe which components (submodels, scale bridges, data converters, UQ components, etc.) a multiscale model consist of, how many instances of each we need, and how they are wired together.
MUSCLE3 is intended to scale from your laptop to the exascale. At the low end, it supports a non-distributed but parallel mode in which an entire multiscale simulation, including all component implementations and the MMSL configuration, is in a single (short) Python file. Next is a distributed mode where the manager and component instances run on multiple nodes in a cluster, and communicate directly with one another. Beyond that, additional components and optimisations are envisioned that would allow scaling to huge machines or combinations of multiple machines. Our goal is to make the transitions between these modes as smooth as possible, so that extra compute power can be added gradually, as needed.
Added
- Connecting multiple conduits to outgoing ports
- Checkpointing (preview, not fully reliable and open to change)
- Clang support
- Intel® compiler support
- Error in case different versions of MUSCLE3 are used
Improved
- TCP latency (performance)
- More helpful messages for configuration errors
- Small documentation improvements
Removed
- Python 3.6 support
Thanks
- Maarten at Ignition Computing for implementing most of the above
- The ITER Organisation for funding most of this work
Release 0.5.0
MUSCLE3 0.5.0
MUSCLE3 is the third incarnation of the Multiscale Coupling Library and Environment, and the successor to MUSCLE 2. Its purpose is to make creating coupled multiscale simulations easy, and to then enable efficient Uncertainty Quantification of such models using advanced semi-intrusive algorithms.
MUSCLE3 uses the Multiscale Modelling and Simulation Language (MMSL) to describe the structure of a multiscale model. MMSL can be expressed in the form of a diagram (gMMSL; not yet implemented) or as a YAML file (yMMSL; this is convenient both for people and for software). The MMSL lets one describe which components (submodels, scale bridges, data converters, UQ components, etc.) a multiscale model consist of, how many instances of each we need, and how they are wired together.
MUSCLE3 is intended to scale from your laptop to the exascale. At the low end, it supports a non-distributed but parallel mode in which an entire multiscale simulation, including all component implementations and the MMSL configuration, is in a single (short) Python file. Next is a distributed mode where the manager and component instances are started on multiple nodes in a cluster, and communicate directly with one another. Beyond that, additional components and optimisations are envisioned that would allow scaling to huge machines or combinations of multiple machines. Our goal is to make the transitions between these modes as smooth as possible, so that extra compute power can be added gradually, as needed.
Added
- MUSCLE3 now starts submodels and other components (using QCG-PilotJob)
- Automatic resource management for components on HPC
Improved
- Build and installation process now even easier
- Improved error messages and reliability
- Cleaner and more informative logging output
- TCP performance and scalability improvements
Fixed
- Various issues when building and running on HPC clusters
- Many small fixes
Removed
- Python 3.5 support
- Removed gRPC for faster and more reliable builds
- Pipe-based networking, as it had no benefits and some issues
Thanks
- Stefan, Merijn and Maarten for reporting issues
- Piotr and Bartek for creating and supporting QCG-PilotJob
Release 0.4.0
MUSCLE 3
MUSCLE 3 is the third incarnation of the Multiscale Coupling Library and Environment.
With MUSCLE 3, you can connect multiple simulation models together into a multiscale simulation. Simulation models can be as simple as a single Python file, or as complex as a combination of multiple separate simulation codes written in C++ or Fortran, and running on an HPC machine.
Please refer to the MUSCLE 3 documentation to get started.
Release 0.4.0
Incompatible changes
compute_elements
are now calledcomponents
in .ymmsl files
Improved
- Use latest OpenSSL library when installing it automatically
Fixed
- Handling of non-contiguous and F-order numpy arrays
- C++ memory usage for large dicts/lists now more reasonable
- Improved shutdown when Python submodel crashes
- Logging warning message
Release 0.3.2
Improved
- Accessing settings from C++ now more flexible
- Python produces more detailed logs to aid in debugging
- Improved pkg-config set-up
- Improved build system output to help find problems
- Documentation on logging in Python
- Protobuf dependency build now more compatible
Fixed
- C++ list/dict building functions
- C++ use-after-free when receiving grids
Thanks
- Pavel for testing and reporting issues
- Dongwei for testing and reporting issue
Release 0.3.1
Added
- Support for sending and receiving multidimensional grids/arrays
- Support for Python 3.8
Improved
- Python 3.5.1 support
- Build compatibility on more operating systems
Thanks
- Olivier for testing, reporting and fixing build issues
- Pavel for testing and reporting build issues
- Hamid for testing and reporting build issues
- Ben for testing and reporting build issues
Release 0.3.0
Incompatible changes
- Data::key() now returns std::string instead of DataConstRef.
- Data::value() now return Data rather than DataConstRef
Added
- Support for Fortran, including MPI
Improved
- Fixes to examples
- Small documentation improvements
- Improved compatibility with other packages using gRPC
Thanks
- Pavel for reporting documentation/examples issues
- Derek for testing on Eagle
- Dongwei for reporting the gRPC issue
Release 0.2.0
Added
- Support for C++
- Support for MPI in C++
Improved
- Cluster/HPC networking
Incompatible Changes
- Fatal logic errors now throw instead of exiting, so that you have a chance to shut down the model cleanly before exiting.
- Instance.exit_error() was replaced by Instance.error_shutdown(), which no longer exits the process, it just shuts down the Instance.
- Central MUSCLE 3-managed settings are called settings everywhere now, not parameters. As a result, the API has changed in several places.