Performance Tools#

Note

New in v4.1.0

Overview#

We use PETSc event logging to instrument the PyLith code for evaluating runtime performance. This allows us to simultaneously evaluate the performance of the PyLith C++ code as well as the PETSc code. In PyLith v2, we logged the floating point operations for the residual and Jacobian evaluation. We have not yet included these in PyLith v3 and later because we now use kernels that are called many, many times and event logging in this methods would introduce overhead and reduce performance.

Implementation of Event Logging#

PyLith includes pylith::utils::EventLogger as a high-level interface for event logging with PETSc. We create local Events classes within the C++ implementation files (*.cc) to create variables that hold the event ids. This streamlines logging events within the C++ code. We identify events using variables rather than strings, which means that typos in names are detected at compile time rather than at runtime. Refer to the _Problem::Events class near the top of libsrc/pylith/problems/Problem.cc for an example of how we implement event logging.

We name events in PyLith C++ code following the template PL:CLASS:METHOD (C++ namespace style but reducing :: to : to reduce the length of the string) in which we are logging method METHOD in C++ class CLASS. In the PyLith Python code, we follow the template PL.CLASS.METHOD.

Evaluating Performance#

Tip

We can get a general ideal of runtime performance bottlenecks with a debugging build of the code, but fine-tuning performance should be done using an optimized build.

Quick View#

Run a PyLith simulation adding the command line argument --petsc.log_view=ascii:FILENAME. This will dump performance information in ASCII format to FILENAME.

Graphical View#

Run a PyLith simulation adding the command line argument --petsc.log_view=:FILENAME_LOGVIEW:ascii_flamegraph. This will dump performance information to FILENAME_LOGVIEW that can be viewed using flame graph tools, such as FlameGraph.

Viewing Performance with FlameGraph#

  1. Run PATH_TO_FLAMEGRAPH/flamegraph.pl FILENAME_LOGVIEW > FILENAME_PLOT.svg

  2. Load FILENAME_PLOT.svg into your favorite svg viewer, such as a web browser.

Note

You can install FlameGraph by cloning the repository or downloading and unpacking a release.