3.3.1 Orchid Update

3.3 version with bug fixes and minor improvements.

3.3 Orchid

Release Notes

New features:

• New Jupyter kernel and JupyterLab extension to manage PyCOMPSs in the Jupyter ecosystem (https://github.com/bsc-wdc/jupyter-extension).
• Integration with Energy Aware Runtime (EAR) to obtain energy profiles in python-based applications (https://www.bsc.es/research-and-development/software-and-apps/software-list/ear-energy-management-framework-hpc).
• Support for users-defined dynamic constraints based on of task parameters values.
• GPU cache for PyTorch tensors.

Improvements:

• The support of interactive Python and Jupyter notebooks has been extended to work in non shared disk environment.
• Data transformations are supporting the data conversion to directory types.
• Workflow Provenance: new data persistence feature, new inputs and outputs terms to define data assets by hand, new sources term, improved common paths detection, and minimal YAML support.
• Configuration files for Leonardo and Galileo HPC systems.
• Several Bug fixes.

Known Limitations:

• Dynamic constrains are limited to task parameters declared as IN which are not future objects (generated by previous tasks).
• Issues when using tracing with Java 14+. For Java 17+ require to include this jvm flag "-Dcom.sun.xml.bind.v2.bytecode.ClassTailor.noOptimize=true"
• Collections are not supported in http tasks.
• macOS support is limited to Java and Python without CPU affinity (require to execute with --cpu_affinity=disable). Tracing is not available.
• Reduce operations can consume more disk space than the manually programmed n-ary reduction.
• Objects used as task parameters must be serializable.
• Tasks that invoke Numpy and MKL may experience issues if a different MKL threads count is used in different tasks. This is due to the fact that MKL reuses threads in the different calls and it does not change the number of threads from one call to another. This can be also happen with other libraries implemented with OpenMP. To fix these issues use the DLB option for in the cpu_affinity flag.
• C++ Objects declared as arguments in coarse-grain tasks must be passed as object pointers in order to have proper dependency management.
• Master as worker feature is not working for executions with persistent worker in C++.
• Coherence and concurrent writing in parameters annotated with the "Concurrent" direction must be managed by the underlying distributed storage system.
• Delete file calls for files used as input can produce a significant synchronization of the main code.
• Defining a parameter as OUT is only allowed for files and collections of objects with a default constructor.

For further information, please refer to COMPSs documentation

3.2 Narcissus

Release Notes

New features:

• Support for Containers in MPI and MPMD tasks
• Cache for cuPy Objects in GPU tasks.
• New SSH/SCP Adaptor to submit tasks to remote clusters (GOS Adaptor).

Improvements:

• Workflow Provenance: support the new Workflow Run Crate profile (v0.1), improved the structuring of source files for the application, new term to specify a submitter, more details on the machine that run the workflow (architecture and COMPSs environment variables)
• Configuration files for Jusuf system.
• Several Bug fixes.

Known Limitations:

• Issues when using tracing with Java 14+. For Java 17+ require to include this jvm flag "-Dcom.sun.xml.bind.v2.bytecode.ClassTailor.noOptimize=true"
• Collections are not supported in http tasks.
• macOS support is limited to Java and Python without CPU affinity (require to execute with --cpu_affinity=disable). Tracing is not available.
• Reduce operations can consume more disk space than the manually programmed n-ary reduction.
• Objects used as task parameters must be serializable.
• Tasks that invoke Numpy and MKL may experience issues if a different MKL threads count is used in different tasks. This is due to the fact that MKL reuses threads in the different calls and it does not change the number of threads from one call to another. This can be also happen with other libraries implemented with OpenMP. To fix these issues use the DLB option for in the cpu_affinity flag.
• C++ Objects declared as arguments in coarse-grain tasks must be passed as object pointers in order to have proper dependency management.
• Master as worker feature is not working for executions with persistent worker in C++.
• Coherence and concurrent writing in parameters annotated with the "Concurrent" direction must be managed by the underlying distributed storage system.
• Delete file calls for files used as input can produce a significant synchronization of the main code.
• Defining a parameter as OUT is only allowed for files and collections of objects with a default constructor.

For further information, please refer to COMPSs documentation

3.1 Margarita

Release Notes

New features:

• Support for Julia applications as tasks
• Support for uDocker for container tasks.
• New decorator to apply a transformation to a parameter (@data_transformation)
• Automatic creation of Data Provenance information from Java applications.
• Constraint to force execution in the local agent. (is_local=True)
• Extended external software description (JSON file in Software decorator) to allow the definition of task parameters.
• Enable the specification of master working directory (location where data serializations are stored). Old --master_working_dir flag for indicating the execution dir is now --job_execution_dir.

Improvements:

• Data Provenance: enhanced addition of source files using directories
• Fix issues when wait_on cancelled data. Get the latest version.
• Swap distutils with setuptools in the Python binding installation.
• Improvements in the management of tasks returning a modified parameter (a=task(a)).
• Container decorator allows to define flags to be passed to the container execution commands.
• Improvements in the support for agents and task nesting.
• Improvements in pluggable schedulers.
• Improvements in python cache.
• Improvements in tracing system: Fix assigned gpus events, include python cache events.
• Fix issues in Collection graph generation
• Configuration files for Hawk, Mahti, Dardel and Lenox system.
• Several Bug fixes.

Known Limitations:

• Issues when using tracing with Java 14+. For Java 17+ require to include this jvm flag "-Dcom.sun.xml.bind.v2.bytecode.ClassTailor.noOptimize=true"
• Collections are not supported in http tasks.
• macOS support is limited to Java and Python without CPU affinity (require to execute with --cpu_affinity=disable). Tracing is not available.
• Reduce operations can consume more disk space than the manually programmed n-ary reduction.
• Objects used as task parameters must be serializable.
• Tasks that invoke Numpy and MKL may experience issues if a different MKL threads count is used in different tasks. This is due to the fact that MKL reuses threads in the different calls and it does not change the number of threads from one call to another. This can be also happen with other libraries implemented with OpenMP. To fix these issues use the DLB option for in the cpu_affinity flag.
• C++ Objects declared as arguments in coarse-grain tasks must be passed as object pointers in order to have proper dependency management.
• Master as worker feature is not working for executions with persistent worker in C++.
• Coherence and concurrent writing in parameters annotated with the "Concurrent" direction must be managed by the underlying distributed storage system.
• Delete file calls for files used as input can produce a significant synchronization of the main code.
• Defining a parameter as OUT is only allowed for files and collections of objects with a default constructor.

For further information, please refer to COMPSs documentation

3.0 Lavender

Release Notes

New features:

• CLI to unify executions of application in different environments.
• Automatic creation of Data Provenance information from PyCOMPSs executions.
• Transparent task-based checkpointing support.
• Support for MPMD MPI applications as tasks.
• Support for task epilog and prolog.
• Generic support for reusable descriptions of external software execution inside a COMPSs task (Software decorator).
• Mypy compilation of python binding.
• Integration with DLB DROM for improving affinity in OpenMP tasks.
• RISC-V 64bit support.

Deprecated Features:

• Python 2 support.
• Autoparallel module (requires python2).
• SOAP Service tasks.

Improvements:

• wait_on and wait_on_file API homogenization.
• Improvements in the support for task nesting.
• Improvements in plugable schedulers.
• Improvements in memory profiling reports.
• Improvements in tracing system: Offline tracing generation, and support for changes of working directory.
• Configuration files for Nord3v2 and LaPalma system.
• Several Bug fixes.

Known Limitations:

• Issues when using tracing with Java 14+.
• Collections are not supported in http tasks
• OSX support is limited to Java and Python without CPU affinity (require to execute with --cpu_affinity=disable). We have also detected issues when several python3 versions are installed in the system. Tracing is not available.
• Reduce operations can consume more disk space than the manually programmed n-ary reduction.
• Objects used as task parameters must be serializable.
• Tasks that invoke Numpy and MKL may experience issues if a different MKL threads count is used in different tasks. This is due to the fact that MKL reuses threads in the different calls and it does not change the number of threads from one call to another. To fix these issues use the DLB option for in the cpu_affinity flag.
• C++ Objects declared as arguments in coarse-grain tasks must be passed as object pointers in order to have proper dependency management.
• Master as worker is not working for executions with persistent worker in C++.
• Coherence and concurrent writing in parameters annotated with the "Concurrent" direction must be managed by the underlying distributed storage system.
• Delete file calls for files used as input can produce a significant synchronization of the main code.
• Defining a parameter as OUT is only allowed for files and collection files.

For further information, please refer to COMPSs documentation

2.10 (Kumquat)

Release Notes

New features:

• Runtime:

  • Support for http requests (http decorator).
  • Support for Java versions higher than 8.
  • Enable grouping processes in MPI tasks with the processes_per_node flag.
  • Partial support for OSX systems

• Python:

  • PyArrow object serialization support
  • Cache profiling enabled

Improvements:

• Fixes in managing directory parameters.
• Improvements in asynchronous file system operations.
• Improvements visualizing collections in task dependency graphs.
• Improvements and fixes in the support for different MPI versions (COMPSS_MPIRUN_TYPE)
• Improvements in tracing system. Fixes in events and cfgs
• Configuration files for Karolina, Mahti and CTE-AMD system.
• Several Bug fixes

Known Limitations:

• Collections are not supported in http tasks
• OSX support is limited to Java and Python without CPU affinity (require to execute with --cpu_affinity=disable). We have also detected issues when several python3 versions are installed in the system. Tracing is not available.
• Reduce operations can consume more disk space than the manually programmed n-ary reduction.
• Objects used as task parameters must be serializable.
• Tasks that invoke Numpy and MKL may experience issues if a different MKL threads count is used in different tasks. This is due to the fact that MKL reuses threads in the different calls and it does not change the number of threads from one call to another.
• C++ Objects declared as arguments in coarse-grain tasks must be passed as object pointers in order to have proper dependency management.
• Master as worker is not working for executions with persistent worker in C++.
• Coherence and concurrent writing in parameters annotated with the "Concurrent" direction must be managed by the underlying distributed storage system.
• Delete file calls for files used as input can produce a significant synchronization of the main code.
• Defining a parameter as OUT is only allowed for files and collection files.

For further information, please refer to COMPSs documentation

2.9.2

COMPSs version 2.9.2

Bug fixes in 2.9 branch

2.8.1

COMPSs version 2.8.1

Bug fixes in branch 2.8

2.8 (Iris)

Release Notes

New features

• Runtime:
  • New container task annotation to allow the execution of a task inside a container.
  • New IN_DELETE parameter direction type for one-use parameters. It deletes de element after task execution.
  • Support to Reductions.
  • Data layout for collections in MPI tasks. Allow to group elements of a collection to MPI processes.
  • New cpu affinity and runtime events for a better performance analysis
• Java:
  • CPU Affinity in Java tasks.
• Python:
  • Support for python dictionaries as collection parameters.
  • Dummy implementations for the binary and mpi tasks to allow testing in sequential implementations.
  • Allow tracing events at master user code.
  • New python binding tracing events.
• C++:
  • CPU Affinity in tasks executed with persistent executors.
• DDS-2:
  • New methods and optimizations in DDS class.

Improvements

• Change locality calculation from scheduling to location update.
• Some runtime file system operations removed from task execution critical path.
• Improvements in CPU-Task Executor affinity. Try executor tries to reuse its previous affinity.
• Support for Collection INOUT/OUT in Persistent storage executions.
• Improvements in tracing cfgs.
• Improvements of storage events in traces.
• Improvements in enqueue_compss and supercomputers cfg files semantics for shared and local disks.
• Adding NVRAM mode flag in enqueue_compss to allow initialization of nodes with specific NVRAM mode. Requires support for the cluster's resource manager.
• Configuration files for Irene and A64FX-based systems.
• Several bug fixes.

Known Limitations

• Objects used as task parameters must be serializable.
• Tasks that invoke Numpy and MKL may experience issues if a different MKL threads count is used in different tasks. This is due to the fact that MKL reuses threads in the different calls and it does not change the number of threads from one call to another. This can be also happen with other libraries implemented with OpenMP.
• C++ Objects declared as arguments in a coarse-grain tasks must be passed in the task methods as object pointers in order to have a proper dependency management.
• Master as worker is not working for executions with persistent worker in C++.
• Coherence and concurrent writing in parameters annotated with the "Concurrent" direction must be managed by the underlaying distributed storage system.
• Delete file calls for files used as input can produce a significant synchronization of the main code.
• Defining a parameter as OUT is only allowed for files and collection files.
• There is an issue with hwloc and Docker which could affect to python mpi workers. Fixing it require to upgrade the hwloc version used by the MPI runtime.

More Information can be found at COMPSs documentation

2.6 (Gardenia)

Release Notes

New features:

• Runtime:

  • New Commutative direction type for task parameter. It indicates that the order of modifications done by tasks of the same type to the parameter does not affect the final result, i.e., tasks operating on a given commutative parameter can be executed in any order between them.
  • New "Stream" parameter type, to enable the combination of data-flows and task-based workflows in the same application. The stream parameter type is defined to enable communication of streamed data between tasks.
  • Timeout property for tasks. Tasks lasting more than their timeout will be cancelled and considered as failed. This property can be combined with the "onFailure" mechanism.
  • Enable the definition of task groups.
  • Support for throwing special exceptions (COMPSsExceptions) in tasks and catching them in task groups.
  • Task cancellation management in the occurrence of a task group exception and unexpected application finalization (Exception or exit code different from 0)

• Python:

  • Enable the declaration of a list of strings as a file collection task parameter.
  • Support for Python MPI tasks

• C:

  • Support for tasks with fine-grain parallelization with OmpSs-2 programming model.

Improvements:

• New multi-threaded ready scheduler with better scalability.
• Support for task with "isReplicated" properties and parameters with INOUT/OUT direction.
• Optimization in deletion of python objects to avoid large synchronizations in shared file systems.
• Improved the AutoParallel submodule to define data blocks as collection types.
• Several Bug fixes

Known Limitations:

• Tasks that invoke Numpy and MKL may experience issues if a different MKL threads count is used in different tasks. This is due to the fact that MKL reuses threads in the different calls and it does not change the number of threads from one call to another.
• C++ Objects declared as arguments in coarse-grain tasks must be passed as object pointers in order to have proper dependency management.
• Master as worker is not working for executions with persistent worker in C++.
• Coherence and concurrent writing in parameters annotated with the "Concurrent" direction must be managed by the underlying distributed storage system.
• Delete file calls for files used as input can produce a significant synchronization of the main code.
• Defining a parameter as OUT is only allowed for files and collection files.

For further information, please refer to COMPSs documentation at https://compss-doc.readthedocs.io/en/2.6/