Tobias Zagorni created ARROW-16562:
--------------------------------------
Summary: [C++] Avoid slicing array inputs in ExecBatchIterator
that would result in one slice
Key: ARROW-16562
URL: https://issues.apache.org/jira/browse/ARROW-16562
Project: Apache Arrow
Issue Type: Sub-task
Components: C++
Reporter: Tobias Zagorni
Attachments: avoid-slicing-performance.txt
For scalar functions, {{ExecBatchIterator}} is used to iterate over batches in
smaller units. It is implemented by calling {{{}Array::slice(){}}}. For small
batches, this is unecessary, since only one slice is created. The slice
operation still causes some overhead by copying the shrared_ptrs of the
ArrayData object, inclung the type pointer, which can lead to contention
(ARROW-16161).
This Patch checks if the batch size is smaller than the slice size first, and
uses std::move in this case.
I have attached a comparision of the ExecuteScalarExpressionOverhead benchmark
here: [^avoid-slicing-performance.txt]
(created with --benchmark_min_time=20, the standard low runtime tends to be
noisy with this, but also shows a positive tendency)
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
