[ntuple] set default initial page size to 16B (was 256B)

[jblomer]

This change has a negligible hit on performance (<1% for serializing 100k entries to 10k double columns uncompressed into the Null sink). It does help considerably keeping the page count low when wide and sparse EDMs are written under memory pressure (real world CMS SIM example: 3M --> 1.3M pages).

[pcanal]

I am at a tad bit confused on the unit of fInitialUnzippedPageSize, the title and comments are consistent in their use of byte but the RNTupleModel.cxx line:

    minPageBufferSize += r.size() * options.GetInitialUnzippedPageSize();

seems to use it as 'number' of elements. What am I misunderstanding?

On the other hand we also have:

// tree/ntuple/src/RColumn.cxx
42:   fInitialNElements = pageSink.GetWriteOptions().GetInitialUnzippedPageSize() / fElement->GetSize();

So on the surface it appears the usage might be inconsistent.

[pcanal]

See the comment at #23638 (comment) which may or may not hint at other possible updates.

[jblomer]

@pcanal In your referenced code snippet, r.size() is not the element size. It is the length of a vector with column representatives. We need one page for every column, so that calculation is correct. The unit is B.

[pcanal]

In your referenced code snippet, r.size() is not the element size. It is the length of a vector with column representatives. We need one page for every column, so that calculation is correct. The unit is B.

Doesn't that need to be also timed by the sizeof then ?

[jblomer]

In your referenced code snippet, r.size() is not the element size. It is the length of a vector with column representatives. We need one page for every column, so that calculation is correct. The unit is B.

Doesn't that need to be also timed by the sizeof then ?

Sorry, I don't understand that question. Can you please clarify?

[silverweed]

Doesn't that need to be also timed by the sizeof then ?

The sizeof is taken into account by options.GetInitialUnzippedPageSize(): it's in the end minBufSize = nPages * minPageSize (where r.size() == nColumnReprs == nPages)

[github-actions]

Test Results

    24 files      24 suites   3d 23h 9m 9s ⏱️
 3 884 tests  3 880 ✅ 0 💤 4 ❌
83 135 runs  83 131 ✅ 0 💤 4 ❌

For more details on these failures, see this check.

Results for commit fa7024a.

[pcanal]

Doesn't that need to be also timed by the sizeof then ?
Sorry, I don't understand that question. Can you please clarify?

Re-reading the code clarified that the question was moot/misplaced.

Questions were clarified.

[pcanal]

LGTM.