All notable changes to the OmpSs-2 programming model, the Nanos6 runtime system, the Mercurium source-to-source compiler, and the OmpSs-2 compiler based on LLVM will be documented in this file.
The OmpSs-2 2020.06 release introduces several features that improve the general performance of OmpSs-2 applications. It adds a new variant to extract execution traces with a lightweight internal tracer. It also improves the support for CUDA and provides support for OpenACC tasks. Additionally, it introduces a new compiler for OmpSs-2 programs in beta development based on the open-source LLVM infrastructure.
- Use jemalloc as a scalable multi-threading memory allocator
- Add
turbo
variant enabling floating-point optimizations and the discrete dependency system - Refactor of CPU Manager and DLB support improvements
- Add new OmpSs-2 compiler based on LLVM in beta development supporting several OmpSs-2 features
- Add support for non-blocking TAMPI in the LLVM OpenMP runtime system
- Bugfixes, performance and code improvements
- Improve taskfor distribution policy
- Improve scheduling performance and code infrastructure
- Implement the discrete dependency system with lock-free techniques
- Add support for weak dependencies in discrete
- Add support for commutative and concurrent dependencies in discrete
- Refactor the hardware counters infrastructure and support both PAPI and PQoS counters
- Add
ctf
variant to extract execution traces in CTF format using a lightweight internal tracer - Provide the
ctf2prv
tool to convert CTF traces to Paraver traces - Avoid Extrae trace desynchronizations in hybrid MPI+OmpSs-2 executions
- Remove the
stats-papi
instrumentation variant
- Refactor of the devices' infrastructure
- Perform transparent CUDA Unified Memory prefetching
- Add support for cuBLAS and similar CUDA APIs
- Add support for OpenACC tasks
The OmpSs-2 2019.11.2 release introduces some bug fixes.
- Fix important error at the Nanos6 initialization
- Fix in discrete dependency system
- Several fixes for OmpSs-2@Cluster
- Fix minor issues of Mercurium
The OmpSs-2 2019.11.1 release introduces some bugfixes and performance improvements.
- Fix execution of CUDA tasks
- Fix
dmalloc
in OmpSs-2@Cluster - Add missing calls to CPU Manager
- Improve taskfor performance
- Improve general performance by using a reasonable cache line size padding
- Add tests checking the execution of CUDA tasks
- Fix minor issues of Mercurium
The OmpSs-2 2019.11 release introduces a new optimized data dependency implementation. It improves the usability,
performance and code of the scheduling infrastructure and the task for
feature. It also adds support for DLB
and OmpSs-2@Linter.
- Data dependency implementation can be decided at run-time through
NANOS6_DEPENDENCIES
variable - Performance and code improvements on the
task for
feature - Add support for Dynamic Load Balancing (DLB) tool
- Add compiler and runtime support for OmpSs-2@Linter
- Important bugfix in memory allocator (used by OmpSs-2@Cluster)
- Bugfixes, performance and code improvements
- Add new optimized discrete dependency system implementation; enabled by
NANOS6_DEPENDENCIES=discrete
- Usability, performance and code improvements on the scheduling infrastructure
- Remove profile instrumentation variant
- Remove interception mechanism of memory allocation functions
The OmpSs-2 2019.06.2 release introduces some bugfixes.
- Compiling extrae variant with high optimization flags
- Removing backtrace sampling from the extrae variant
The OmpSs-2 2019.06.1 release mainly introduces bugfixes and code improvements.
- Renaming loop directive to task for
- Tasks can leverage reductions and external events at the same time (over distinct data regions)
- OmpSs-2@Cluster bugfixes
- Fixing binding information reported by nanos6-info binary
- Support for the TAGASPI library
- Other bugfixes and code improvements
The OmpSs-2 2019.06 release mainly introduces the new support for OmpSs-2@Cluster. It also includes some improvements and optimizations for array task reductions and general bugfixes.
- Support for OmpSs-2@Cluster
- Bugfixes and performance improvements
- Bugfixes and optimization for array reductions
- Delete obsolete task data dependency implementations
- Delete obsolete schedulers
The OmpSs-2 2018.11 release provides full support for the TAMPI library. It also includes general bugfixes and performance improvements.
- Full support for TAMPI
- Bugfixes and performance improvements
- Bugfixes in task external events API
The OmpSs-2 2018.06.2 release introduces some bugfixes.
- Bugfixes in HWLOC support
The OmpSs-2 2018.06.1 release introduces some bugfixes.
- Bugfixes in task reductions
The OmpSs-2 2018.06 release introduces support for OmpSs-2@CUDA in Unified Memory NVIDIA devices. It also supports array task reductions in C/C++ and task priorities. Additionally, it provides two new APIs used by the TAMPI library.
- Support for OmpSs-2@CUDA Unified Memory
- Bugfixes and performance improvements
- Support for array task reductions in C/C++
- Support for task priorities
- Add priority scheduler
- Add polling services API
- Add task external events API
- Rename taskloop construct to loop
The OmpSs-2 2017.11.1 release provides general bugfixes.
- Fixes for the building system
- Fixes for the loading system
The OmpSs-2 2017.11 release is the first release of the OmpSs-2 parallel task-based programming model, which comprises the Nanos6 runtime system and the Mercurium source-to-source compiler. This version provides the essential infrastructure to manage the parallelism of user tasks (task creation, task scheduling, etc.) and their data dependencies. The task dependency system supports array section dependencies, the nested dependency domain connection, and both early release and weak dependency models.
- General infrastructure of the runtime system and the compiler
- Support for user tasks and nesting of tasks
- Implement different schedulers: FIFO, LIFO, etc
- Implementation of a task data dependency system
- Support for array section dependencies
- Support for nested dependency domain connection
- Support for early release of task dependencies
- Support for weak task dependencies
- Support for reductions
- Taskloop construct with dependencies
- Task pause/resume API