Skip to content

Service flux.daemon_flux does not benefit from Numba just-in-time compilation #929

Closed

Description

@thehrh

It would seem both @jit(forceobj=True)-decorated functions in the daemon_flux service are similar to the second type of example (pandas) from https://numba.readthedocs.io/en/0.65.0/user/5minguide.html#will-numba-work-for-my-code.

As the example notebook demonstrates, a PISA user currently faces the following warning when outputs of a pipeline containing the service are computed for the first time:

NumbaWarning: Compilation is falling back to object mode WITHOUT looplifting enabled because Function "make_2d_flux_map" failed type inference due to: non-precise type pyobject

To confirm, I have performed timings with and without just-in-time compilation of both functions by repeatedly generating outputs with randomised parameter values for the IceCube_3y_neutrinos_daemon.cfg pipeline, on both single- and multi-core targets:

  • with just-in-time compilation, the first compute call (emitting the warning) takes longer
  • all subsequent compute calls take equally long (roughly 0.73s on the benchmark machine, independent of number of cores)

Metadata

Metadata

Assignees

Type

Projects

No projects

    Milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions