Profiling Profiling - Easy to use, low overhead, small files
- Statistical information
Event Tracing - High overhead
- Application perturbation
- Trace file storage
- Fine detail
- Time-based effects
- Flexibility
Developed by the Center for Information Services and High Performance Computing (ZIH) of TU Dresden Developed by the Center for Information Services and High Performance Computing (ZIH) of TU Dresden Open Source library (BSD license) LLNL-funded OTF trace format - MPI
- Threads: OpenMP and pthreads
- Application function entry/exit
- Performance Counters
- Memory and I/O Tracing
Profiling mode
Proprietary trace analysis and visualization tool Includes - Summary
- Timelines
- Charts
- Communication grid
Re-build application with VampirTrace wrapper scripts - C : vtcc –vt:cc mpicc
- Fortran : vtf77 –vt:f77 mpif77
- C++ : vtCC –vtcxx mpiCC
- Options : vtcc –vt:help
- No application instrumentation : -vt:inst manual
Run application Examine trace data with Vampir or VampirServer - vampir test_HPCCG.otf
- vng
- srun –n 4 –ppdebug vngd
Trace generation (VampirTrace) Trace generation (VampirTrace) - Tracing limits
- 32MB trace buffer
- Default is to deactivate tracing when buffer is full.
- Event filter
- Run-time filter file
- vtfilter command
Trace Visualization (VampirServer)
Vampir Vampir VampirTrace - http://www.tu-dresden.de/zih/vampirtrace
Dostları ilə paylaş: |