RNTuple/RDataFrame prefetching of pages should handle common analysis use cases

[pcanal]

Feature description

This is inherited from similar request for TTree: https://its.cern.ch/jira/browse/ROOT-5098

Efficiently prefetch RTNuple pages in for common analysis cases including:

• reading branches for the first time outside the learning period (note: this description is in term of TTree behavior, the RNTuple behavior is different, with no 'learning' but instead prefetching all the columns that are referred to within any of the active views, whether they are actively retrieved or not).
• select and skim pattern (read a few columns for all entries and all columns for a few entries).

Investigate alternative algorithm to try to leverage the os prefetcher; i.e. be able to (always) go faster than the case with read ordered baskets (and/or disable the os prefetcher when it works against the direct prefetching).

Related: we have had request to allow customization of the prefetching algorithm https://its.cern.ch/jira/browse/ROOT-5097

[sanchayanghosh]

I have a fair idea of what to do after going through your feature description.

Could I take up this issue?

[sanchayanghosh]

I am going through the prefetch in TTree.cxx, and the allied functions.

I will make a draft PR with what I intend to add to RNTuple

[jblomer]

The current RNTuple approach is this:

• Prefetch all the requested columns (union of views and model [model only if LoadEntry is used])
• Assume monotonic access (increasing event number) per data stream
• Advise the OS against block-level prefetching (patch to be submitted) because we know the exact byte ranges

This should cover the relevant cases. I'd argue that the "select and skim" pattern is better served with two readers than by attempting to manually interfere with the prefetcher.