11. Profiling

Reference:

1. https://www.firedrakeproject.org/optimising.html

2. https://petsc.org/main/manual/profiling/

3. https://petsc.org/main/manualpages/Sys/PetscInitialize/

4. https://petsc.org/main/manualpages/Profiling/PetscLogView/

11.1. log_view

1. -log_view [:filename] Prints summary of log information

2. -log_view :filename.py:ascii_info_detail Saves logging information from each process as a Python file

3. -log_view :filename.xml:ascii_xml Saves a summary of the logging information in a nested format (see below for how to view it)

4. -log_view :filename.txt:ascii_flamegraph Saves logging information in a format suitable for visualising as a Flame Graph (see below for how to view it)

5. -log_view_memory Also display memory usage in each event

6. -log_view_gpu_time Also display time in each event for GPU kernels (Note this may slow the computation)

7. -log_all Saves a file Log.rank for each MPI rank with details of each step of the computation

8. -log_trace [filename] Displays a trace of what each process is doing

11.1.1. Flame graph

运行代码时加上选现 -log_view :profile.txt:ascii_flamegraph, 将会生成文件 profile.txt, 该文件会记录某些代码块的运行时间, 可用于性能分析. 可以通过在线工具 https://www.speedscope.app/ 可视化该文件. 示例代码如下:

python test.py -log_view :profile.txt:ascii_flamegraph

可以通过 PETSc.Log.Event 和 PETSc.Log.EventDecorator 添加自定义事件, 用于测量代码块的运行事件. 如果如上运行代码, 事件名称和运行时间将会记录在上述 profile.txt 中.

1. PETSc.Log.Event

   from firedrake.petsc import PETSc
   
   with PETSc.Log.Event("foo"):
       do_something_expensive()
   

2. PETSc.Log.EventDecorator

   from firedrake.petsc import PETSc
   
   @PETSc.Log.EventDecorator("foo")
   def do_something_expensive():
       ...
   

11.1.2. -log_view :profile.txt

Ref: https://petsc.org/release/manual/profiling/#interpreting-log-view-output-parallel-performance