The development of VTK-m was co-funded by the HBP during the second project phase (SGA1). This page is kept for reference but will no longer be updated.
VTK-m is a scientific visualization and analysis framework that offers a wealth of building blocks to create visualization and analysis applications. VTK-m facilitates scaling those applications to massively parallel shared memory systems, and it will – due to its architecture – most likely also run efficiently on future platforms.
HPX is a task-based programming model. It simplifies the formulation of well-scaling, highly-parallel algorithms. Integrating this programming model into VTK-m streamlines the formulation of its parallel building blocks and thus makes their deployment on present and emerging HPC platforms more efficient. Since neuroscientific applications require more and more compute power as well as memory, harnessing the available resources will become a challenge in itself. By combining VTK-m and HPX into task-based analysis and visualization, we expect to provide suitable tools to effectively face this challenge and facilitate building sophisticated interactive visual analysis tools, tailored to the neuroscientists’ needs.
Parallel primitive algorithms required for VTK-m have been added to HPX along with API support to enable the full range of visualization algorithms developed for VTK-m. A new scheduler has been developed that accepts core/numa placement hints from the programmer such that cache reuse can be maximized and traffic between sockets minimized. High performance tasks that access data shared by application and visualization can use this capability to improve performance. The thread pool management was improved to allow visualization tasks, communication tasks, and application tasks to execute on different cores if necessary, which reduces latency between components and improves the overall throughput of the distributed application. RDMA primitives have been added to the HPX messaging layer. These improvements make it possible to scale HPX applications to very high node/core counts. Respective tests have been successful on 10k nodes using 650k cores.
The development of PyCOMPSs was co-funded by the HBP during the Ramp-up Phase. This page is kept for reference but will no longer be updated, apart from release notes.
PyCOMPSs is the Python binding of COMPSs, (COMP Superscalar) a coarse-grained programming model oriented to distributed environments, with a powerful runtime that leverages low-level APIs (e.g. Amazon EC2) and manages data dependencies (objects and files). From a sequential Python code, it is able to run in parallel and distributed.
COMPSs screenshot
Releases
PyCOMPSs is based on COMPSs. COMPSs version 1.3 was released in November 2015, version 1.4 in May 2016 and version 2.0 in November 2016.
New features in COMPSs v1.3
Runtime
Persistent workers: workers can be deployed on computing nodes and persist during all the application lifetime, thus reducing the runtime overhead. The previous implementation of workers based on a per task process is still supported.
Enhanced logging system
Interoperable communication layer: different inter-nodes communication protocol is supported by implementing the Adaptor interface (JavaGAT and NIO implementations already included)
Simplified cloud connectors interface
JClouds connector
Python/PyCOMPSs
Added constraints support
Enhanced methods support
Lists accepted as a tasks’ parameter type
Support for user decorators
Tools
New monitoring tool: with new views, as workload and possibility of visualizing information about previous runs
Enhanced tracing mechanism
Simplified execution scripts
Simplified installation on supercomputers through better scripts
New features in COMPSs v1.4
Runtime
Added support for Docker
Added support for Chameleon Cloud
Object cache for persistent workers
Improved error management
Added connector for submitting tasks to MN supercomputer from external COMPSs applications
Bug-fixes
Python/PyCOMPSs
General bug-fixes
Tools
Enhanced Tracing mechanism:
Reduced overhead using native Java API
Added support for communications instrumentation added
Added support for PAPI hardware counters
Known Limitations
When executing Python applications with constraints in the cloud the initial VMs must be set to 0
New features in COMPSs v2.0 (released November 2016)
Runtime:
Upgrade to Java 8
Support to remote input files (input files already at workers)
Integration with Persistent Objects
Elasticity with Docker and Mesos
Multi-processor support (CPUs, GPUs, FPGAs)
Dynamic constraints with environment variables
Scheduling taking into account the full tasks graph (not only ready tasks)
Support for SLURM clusters
Initial COMPSs/OmpSs integration
Replicated tasks: Tasks executed in all the workers
Explicit Barrier
Python:
Python user events and HW counters tracing
Improved PyCOMPSs serialization. Added support for lambda and generator parameters.
C:
Constraints support
Tools:
Improved current graph visualization on COMPSs Monitor
Improvements:
Simplified Resource and Project files (NO retrocompatibility)
Improved binding workers execution (use pipes instead of Java Process Builders)
Simplifies cluster job scripts and supercomputers configuration
Several bug fixes
Known Limitations:
When executing python applications with constraints in the cloud the initial VMs must be set to 0
New features in PyCOMPSs/COMPSs v2.1 (released June 2017)
New features:
Runtime:
New annotations to simplify tasks that call external binaries
Integration with other programming models (MPI, OmpSs,..)
Support for Singularity containers in Clusters
Extension of the scheduling to support multi-node tasks (MPI apps as tasks)
Support for Grid Engine job scheduler in clusters
Language flag automatically inferred in runcompss script
New schedulers based on tasks’ generation order
Core affinity and over-subscribing thread management in multi-core cluster queue scripts (used with MKL libraries, for example)
Python:
@local annotation to support simpler data synchronizations in master (requires to install guppy)
Support for args and kwargs parameters as task dependencies
Task versioning support in Python (multiple behaviors of the same task)
New Python persistent workers that reduce overhead of Python tasks
Support for task-thread affinity
Tracing extended to support for Python user events and HW counters (with known issues)
C:
Extension of file management API (compss_fopen, compss_ifstream, compss_ofstream, compss_delete_file)
Support for task-thread affinity
Tools:
Visualization of not-running tasks in current graph of the COMPSs Monitor
Improvements
Improved PyCOMPSs serialization
Improvements in cluster job scripts and supercomputers configuration
Several bug fixes
Known Limitations
When executing Python applications with constraints in the cloud the <InitialVMs> property must be set to 0
Tasks that invoke Numpy and MKL may experience issues if tasks use a different number of MKL threads. 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.
New features in PyCOMPSs/COMPSs v2.3 (released June 2018)
Runtime
Persistent storage API implementation based on Redis (distributed as default implementation with COMPSs)
Support for FPGA constraints and reconfiguration scripts
Support for PBS Job Scheduler and the Archer Supercomputer
Java
New API call to delete objects in order to reduce application memory usage
Python
Support for Python 3
Support for Python virtual environments (venv)
Support for running PyCOMPSs as a Python module
Support for tasks returning multiple elements (returns=#)
Automatic import of dummy PyCOMPSs AP
C
Persistent worker with Memory-to-memory transfers
Support for arrays (no serialization required)
Improvements
Distribution with docker images
Source Code and example applications distribution on Github
Automatic inference of task return
Improved obsolete object cleanup
Improved tracing support for applications using persistent memory
Improved finalization process to reduce zombie processes
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.
New features in PyCOMPSs/COMPSs v2.5 (released June 2019)
Runtime:
New Concurrent direction type for task parameter.
Multi-node tasks support for native (Java, Python) tasks. Previously, multi-node tasks were only posible with @mpi or @decaf tasks.
@Compss decorator for executing compss applications as tasks.
New runtime api to synchronize files without opening them.
Customizable task failure management with the “onFailure” task property.
Enabled master node to execute tasks.
Python:
Partial support of numba in tasks.
Support for collection as task parameter.
Supported task inheritance.
New persistent MPI worker mode (alternative to subprocess).
Support to ARM MAP and DDT tools (with MPI worker mode).
C:
Support for task without parameters and applications without src folder.
Improvements:
New task property “targetDirection” to indicate direction of the target object in object methods. Substitutes the “isModifier” task property.
Warnings for deprecated or incorrect task parameters.
Improvements in Jupyter for Supercomputers.
Upgrade of runcompss_docker script to docker stack interface.
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 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.
PyCOMPSs/COMPSs PIP installation package
This is a new feature available since January 2017.
Installation:
Check the dependencies in the PIP section of the PyCOMPSs installation manual (available at the documentation section of compss.bsc.es). Be sure that the target machine satisfies the mentioned dependencies.
The installation can be done in various alternative ways:
Use PIP to install the official PyCOMPSs version from the pypi live repository:
sudo -E python2.7 -m pip install pycompss -v
Use PIP to install PyCOMPSs from a pycompss.tar.gz
The development of OmpSs was co-funded by the HBP during the Ramp-up Phase. This page is kept for reference but will no longer be updated.
OmpSs is a fine-grained programming model oriented to shared memory environments, with a powerful runtime that leverages low-level APIs (e.g. CUDA/OpenCL) and manages data dependencies (memory regions). It exploits task level parallelism and supports asynchronicity, heterogeneity and data movement.
The new version 15.06 provides the following new features as compared to version 15.04 that was part of the HBP-internal Platform Release in M18:
Socket aware (scheduling taking into account processor socket)
Reductions (mechanism to accumulate results of tasks more efficiently)
Work sharing (persistence of data in the worker) mechanisms
