Releases: LLNL/Caliper
Releases · LLNL/Caliper
Caliper v2.12.0
Caliper v2.12.0 brings new features and improvements, including
- A Python annotation API
- Ability to add run metadata through config strings
- Various performance improvements, in particular for writing .cali files
- Support for the new rocprofiler-sdk profiling API for AMD GPUs
- Markers for MPI communication regions and more detailed MPI communication statistics
Caliper v2.11.0
Caliper v2.11.0 brings bug fixes, optimizations, and new features, including
- Support for annotation levels and new
CALI_MARK_PHASE_BEGIN
/CALI_MARK_PHASE_END
macros to mark high-level regions - LDMS service to forward region timing information to the Lightweight Distributed Metrics Service (LDMS, https://github.com/ovis-hpc/ovis)
- Build system and CMake package/target improvements
- More performance optimizations, including much faster .cali reader for
cali-query
Caliper v2.10.0
Caliper v2.10.0 brings optimizations and improvements, including
- Many performance optimizations
- Updated GOTCHA library
- Increased timer resolution
- Retain node visit order in output
- Improved handling of C++ standard requirements in the build system
Caliper v2.9.1
This minor release fixes an issue when building the umpire service.
Caliper v2.9.0
Caliper v2.9.0 brings improvements and new features, including
- Much improved sample profiling support
- New
sample-report
config recipe - Options for filtering regions by name
- Improved event tracing support
cali2traceevent.py
converter script for creating Google TraceEvent files- Region forwarding to CrayPat
- Various internal improvements
Caliper v2.8.0
Caliper v2.8.0 brings improvements and new functionality, including
- AMD ROCm/HIP activity profiling and tracing with the
roctracer
service and therocm-activity-profile
androcm-activity-report
ConfigManager configs - Collecting statistics for the Umpire memory manager with the
umpire.allocators
andumpire.totals
ConfigManager options - New
cuda.gputime
androcm.gputime
ConfigManager options to include GPU time metrics in most profiling configs - MPI message size and count metrics with the
mpi.message.size
andmpi.message.count
ConfigManager options - Various internal optimizations and build system improvements
Caliper v2.7.0
Caliper 2.7.0 brings various improvements, including
Caliper v2.6.0
Caliper 2.6.0 brings improvements and new features, including
- New cuda-activity-profile and cuda-activity-report config recipes for profiling CUDA device activities
- Support for OpenMP profiling with the OpenMP tools interface (ompt), including OpenMP metrics and the openmp-report recipe
- New API for streaming Caliper output into a user-provided C++ stream
- rocTX support for exporting Caliper regions to AMD rocprofiler
- The region.count metric for counting the number of region calls
- Load custom recipes from JSON files in ConfigManager
- Improvements to table and tree formatters
- Improved detection of Python and CUDA toolkit components in the build system
Caliper 2.5.0
Caliper 2.5.0 brings improvements and new features, including
- Ability to run ConfigManager configs without the ConfigManager API via the
CALI_CONFIG
environment variable - Many improvements to sampling and symbol lookup / call-stack unwinding features
- A new callpath-sample-report config
- Support for unified memory event tracing on NVidia GPUs
- Loop profiling for the Spot config
- A Python reader library for .cali files
- Updated documentation
There are some changes to the build and runtime configuration compared to previous versions:
- We now use libdw for symbol lookup. Use the
WITH_LIBDW
cmake option to enable symbol lookup. The dyninst dependency was dropped. - The
WITH_CALLPATH
cmake option was renamed toWITH_LIBUNWIND
- The
WITH_NVPROF
cmake option was renamed toWITH_NVTX
- The nvprof service was renamed to nvtx
Caliper 2.4.0
Caliper 2.4.0 brings improvements and new features, including
- New loop-report and mpi-report ConfigManager configs
- New Fortran wrappers for the annotation API
- Fortran and C wrappers for the ConfigManager API
- Low-frequency performance measurement modes ("monitor modes")
- Easier and more robust linking and initialization