Investigate bunching of multiple events

[fwyzard]

During the ACAT presentations, we have been reminded of the tried approach of improving GPU throughput by bunching the computation over multiple events.

A production-level approach will likely require dedicated support in the framework, but I think we can already try something ad-hoc on the level of O(10) events at a time.

[makortel]

IIRC Simulations of Chris indicate a negative effect on throughput via the overall scheduling of modules. Maybe a good first step would be to simulate with our workflow parameters?

[fwyzard]

I think to make sense it would need to happen for all modules (cpu and gpu ones) to work on bunches of events, instead of individual events.

[makortel]

Was the concern more about the transfers or about the kernels (or both)?

[fwyzard]

Nothing specific, just general questions and comments.