Archived releases

Build system improvements:

• Add proper support for additional compiler suites via --with-nocross-compiler-suite=<suite>:
  • amdclang: AMD ROCm compilers (amdclang, amdclang++)
  • nvhpc: NVIDIA HPC Toolkit compilers (nvc, nvc++)
  • oneapi: Intel oneAPI compilers (icx, icpx)

Automatic trace analyzer changes & improvements:

• Support OpenMP measurements via OMPT (Score-P ≥ v8.0)
• Fixed a performance regression when analyzing traces including OpenMP with few instrumented user regions.
• Improved metric data aggregation and writing.

Measurement nexus (scan) changes:

• Removed stalled resources counter from default POP multi-run preset and added an enhanced preset pop-with-stalled-resources.

Build system improvements:

• Auto-detect Cray XC platforms with ARM CPUs, supporting Cray, ARM, and GCC compilers
• Added support for Clang and AMD AOCC compilers
• Updated support for Spectrum MPI

Automatic trace analyzer changes & improvements:

• Revised “Early Reduce” wait state definition.
• Added calculation of “Early Reduce” delay costs.
• Fixed various delay cost calculation and propagation issues.
• Fixed various inconsistencies between wait-state and root-cause analysis.
• Made POSIX threads analysis consistent with Score-P by avoiding thread function stub call paths underneath ‘pthread_create’. This also fixes a deadlock when analyzing traces containing “orphaned threads”.

Measurement nexus (scan) changes:

• Added preset mode for multi-run measurements with a preset for POP analysis requirements as an use case.
• Added support for multiple file systems in SCAN_TRACE_FILESYS by using a colon separated list of paths.

Analysis report postprocessing changes:

• Add metric hierarchies for CUDA, OpenCL, and OpenACC. (Note: The trace analysis still only supports host-side events!)
• Renamed ‘-c’ command-line option of ‘square’ to ‘-C’ for running sanity checks on newly created reports.
• Added new ‘-c’ command-line option to ‘square’ to allow specifying the number of counters considered during report scoring (for consistency with ‘scorep-score’).
• Added new ‘-x’ command-line option to ‘square’ to allow passing options directly through to ‘scorep-score’.
• Avoid unnecessary aggregation/postprocessing of reports with multi-run experiments.

Substantial code cleanup.

Automatic trace analyzer changes & improvements:

• Support for virtual process topologies (Score-P ≥ v5.0)
• Various fixes and improvements in timestamp correction algorithm.
• Fixed ‘Late Receiver’ instance tracking.
• Slightly improved analysis report data collation.

Added support for multi-run experiments.

Code refactoring and various bug fixes.

Improved user documentation:

• Revised User Guide including command reference.
• Added man pages.

Build system improvements:

• Fix build issues with compilers defaulting to C++11 or higher (e.g., Intel 2017, PGI 17).
• Fix build issues with PGI 16+ compilers (pgCC no longer available).
• Fix build issues on Cray systems, now also properly taking CRAYPE_LINK_TYPE setting into account.

Automatic trace analyzer changes & improvements:

• Fix rare crash/deadlock in critical-path/delay analysis while analyzing MPI persistent communication.
• Improved memory management.
• Improved handling of OTF2 traces in SIONlib containers.
• Improved trace reading times, especially at scale.
• Fixed detection of wait states in active-target synchronization based on EPIK traces.

Code refactoring and various bug fixes.

• Fixed build issue with GCC 6.1.
• Fixed build issue on the Intel Xeon Phi platform.

Automatic trace analyzer changes & improvements:

• Experimental support for Score-P traces collected using sampling (see OPEN_ISSUES for limitations).

Improved analysis report postprocessing:

• Revised metric hierarchies (organization, metric naming, etc).
• Suppress calculation of performance properties that are only relevant for unused parallel programming models.

Performance property documentation fixes & improvements.

Build system improvements.

Code refactoring and various bug fixes.

Platform support:

• Fixed a build issue on the Intel Xeon Phi platform.
• Improved support for the ‘ibrun’ launcher.

Automatic trace analyzer changes & improvements:

• Worked around rare run-time issue with MVAPICH2.

Platform support:

• Added build system support for Power8/Linux.
• Added build system support for 64-bit ARM/Linux (AArch64).
• Prefer linking static over dynamic Cube/OTF2 libraries on Fujitsu K/FX10/FX100.

Automatic trace analyzer changes & improvements:

• Fixed delay-cost propagation through OpenMP barrier wait states.
• Various algorithmic optimizations reducing overall analysis time for traces of multi-threaded applications:
  • Improved memory management.
  • Improved trace preprocessing.
  • Improved timestamp correction.

Code refactoring and various bug fixes.

Platform support:

• Added support for Intel Xeon Phi, native mode only.
• Added support for Fujitsu FX100 (thanks to T. Nakamura, Fujitsu Ltd).

Automatic trace analyzer changes & improvements:

• Added basic support for POSIX threads.
• Added basic support for OpenMP tasking.
• Added lock contention analysis (OpenMP & POSIX threads).
• Added root-cause/delay analysis (MPI & OpenMP).
• New command-line options ‘–[no-]rootcause’.

Code refactoring and various bug fixes.

Platform support:

• Added support for Fujitsu FX10 & K computer.
• Improved support for Cray systems.

Automatic trace analyzer changes & improvements:

• Added Critical-path analysis.
• Improved Late Receiver detection.
• New command-line options ‘–[no-]critical-path’ and ‘–single-pass’.
• Fixed crash in data collation when number of OpenMP threads varied among MPI processes.

Code refactoring and various small bug fixes.

Initial version of updated User Guide (still work in progress).

New build system based on GNU autotools.

Significant amount of code refactoring.

Automatic trace analyzer changes & improvements:

• Support for arbitrary deep system trees.
• Improved performance of timestamp correction.
• Pattern instance tracking and statistics are now enabled by default.
• New command-line options ‘–verbose’, ‘–[no-]time-correct’, and ‘–[no-]statistics’.
• Limited backward-compatibility support for handling existing traces in EPILOG format generated by Scalasca v1.