Evaluate direct GPU I/O with cuda GDS

[stloufra]

Description

Investigate whether direct GPU I/O can improve I/O performance.

The use cases are:

• Read: load events directly into GPU memory e.g. for reweighting.
• Write: write unweighted events directly from GPU memory.

The goal is to determine whether avoiding CPU < - > GPU copies provides a measurable performance improvement.

References

• GPUDirect Storage Overview — cuFile API and GPU I/O model.
• cuFile API Reference — cuFileRead / cuFileWrite.
• GPUDirect Storage Best Practices — performance and asynchronous I/O.

(to not be lost in time)

[theoheimel]

madspace reaches unweighted event throughputs of around 10M events/second for simple processes like g g > t t~. For more complicated processes, I/O becomes less of an issue as ME evaluation times go up and unweighting efficiencies go down. So I'd say implementing separate I/O for GPUs is probably not worth the effort and added code complexity.