Long Running Hadronic Processes

[Laurie Nevay]

I'm making this as a holding ground for working on this recent problem.

Jobs involving hadronic physics can run very very very long due to non-conservation of energy and repeatedly resampling the interaction. This results in an event taking at least an order of magnitude longer than a typical event.

This could be related to geant4.10.2.

QGSP_BERT is worse than FTFP_BERT.

Example: Expected job time from 100 events - 8hrs Typical job time - 16hrs ~10% (of ~300 jobs) at 65hrs and still going.

[Laurie Nevay]

Stephen's job experience.

[Laurie Nevay]

It doesn't look like a tracking fault or a problem on BDSIM's side so far. It would seem like the job progresses as intended and it's down to the physics list, although more testing needs to be done to confirm this.

If this is the case, we could think about putting in a mechanism for a maximum event time / # of secondaries etc and kill the event (and mark as not be included in analysis - or not even written out).

[Laurie Nevay]

From recent studies this is mostly likely due to very low energy particles spiralling in a relatively strong magnetic field - ie a dipole that has a several Tesla field designed for a high energy primary beam and occasionally, very low energy particles < 1 MeV are produced that spiral without loosing energy endlessly in the field. Our steppers / integrators do not deal with this well.

Stewart is working on termination conditions and modifications to the integrators that will cut on an absolute momentum calculated form a certain gyroradius and length scale supplied - ie ~ beam pipe width / 2 or 5.

[Laurie Nevay]

I've modified the dipole integrators to deal with low energy spiralling particles. They now artificially advance the particle faster along the axis of the helix if their bending radius is less than a specified value (default 10cm).

I've recently run the LHC collimation system with halo for a few million events (1000k events per job) and all have terminated in good time. An initial analysis shows the mean event time has a definite end point.