Skip to content

Profiling and benchmarking

This page describes a practical approach to investigating JURASSIC performance. For user-facing configuration choices and scaling considerations, see Performance tuning. For build options, see Building from source.

Choose a representative workload

Use a reproducible input configuration and record the code revision, compiler, optimization flags, hardware, thread/rank counts, lookup tables, and input size. Compare runs only when these conditions are understood. For retrieval workloads, record iteration counts and convergence behavior as well as elapsed time.

Identify the expensive work

The relevant cost centers depend on the workflow:

  • Forward calculations: radiative-transfer evaluations and lookup-table access.
  • Kernel calculations: repeated evaluations used to construct Jacobians.
  • Retrievals: forward-model and Jacobian calls over multiple iterations.
  • Parallel runs: work distribution, I/O, and synchronization overhead.

Measure before optimizing. A smaller elapsed time is not sufficient evidence of an equivalent scientific result; compare the corresponding outputs and relevant regression tests.

Compiler and profiling options

The src/Makefile documents the build options OPT, INFO, and PROF. For example, from src/:

make clean
make COMPILER=gcc OPT="-O2 -g"

The PROF=1 option adds -pg for profiling builds, and INFO=1 enables compiler optimization reports where supported. The available profiling tools and their output depend on the selected compiler and platform. Do not assume that one profiler works with every CPU, GPU, or MPI configuration.

CPU and parallel benchmarking

Start with a serial or single-process baseline, then vary OpenMP threads or MPI ranks only for executables and modes that support them. Keep input data and scientific settings fixed. Record wall-clock time, memory usage where available, and numerical differences.

For HPC batch runs, see HPC workflows and Parallelization. The current MPI retrieval implementation and workflow-level parallelism have different scaling characteristics; do not interpret them as interchangeable.

Reporting results

A useful benchmark record includes:

  • Git commit or release and whether the checkout was modified.
  • Compiler, libraries, build flags, and execution environment.
  • Workload and relevant dimensions (rays, channels, retrieval cases).
  • Runtime, rank/thread configuration, and repeatability.
  • Regression-test status and any numerical differences.

For optimization changes, report both performance and correctness relative to a documented baseline.