Efficient Data Structures (Overview)

Array of Struct (AoS) vs Struct of Array (SoA).
Data structures should adapt to both CPU and GPU.
Collaboration with Jolly (ROOT) and Daniele (with Stefan Roiser) to ensure portability accross experiments.

Updates from last week

We have investigated the following existing approaches for SoA data structures at Cern.

Description	Link	Advantages	Disadvantages
Jolly's code	reflmempp	nice interface, easy to read and maintain, flexible	reflections are still experimental, no member functions yet
Attila's SoA code	vecmem_emd, silicon_cell	used for ATLAS	Attila wants to replace it; offers not quite the same features we need
O2 analysis SoA code	O2, documentation	used for O2 analysis	depends on Apache Arrow; offers many features (e.g. table operations) we don't need
Stephen Swatman's SoA code	array_wrapper, godbolt	interface is very nice and similar to Jolly's work	pointers and heap allocation, a lot of boilerplate code, no member functions yet

Summary

Jolly's code is the most elegant solution, but reflections are only experimental. (long term solution)
Stephen Swatman's code exposes the same interface but uses only classical techniques. (temporary solution)

General ongoing work

Try to understand O2 code (in particular TPC reconstruction).
Provide member function support in Jolly's SoA code.