[Rust] [Experiment] [WIP]: Use SmallVec in ArrayData to optimize the common usecase of single-buffer arrays

[jhorstmann]

Another experiement for speeding up ArrayData. Most array types only contain a single buffer (except for the validity buffer) and also at most one child data object. I think the only types this does not apply to are struct and union arrays. Using a SmallVec trades one comparison against an allocation and the resulting indirection in those cases.

The ArrayData::new method is left in place unchanged for compatibility.

TODO:

• Use the optimized new_smallvec method in kernels whenever applicable
• Run all benchmarks on a non-throttling machine

@jorgecarleitao @Dandandan this is also related to the discussion on #9271. I tried this initially some month ago and the benchmarks were not conclusive, but might be worth trying again or in combination with removing the Arc indirection.

[github-actions]

Thanks for opening a pull request!

Could you open an issue for this pull request on JIRA?
https://issues.apache.org/jira/browse/ARROW

Then could you also rename pull request title in the following format?

ARROW-${JIRA_ID}: [${COMPONENT}] ${SUMMARY}

See also:

• Other pull requests
• Contribution Guidelines - How to contribute patches

[alamb]

I wonder if using something like a specific enum might also work (and be more memory efficient) as well:

pub enum  ArrayBuffer {
  One(ArrayDataRef),
  Two(ArrayDataRef, ArrayDataRef),
  Many(Vec<ArrayDataRef>)
}

I honestly don't know what the maximum number of buffers any specific ArrayData can have,

[jorgecarleitao]

fyi, max is 3 including the null buffer, so max is 2 in this context. See MutableDataArray, where we use two buffers on the stack to avoid the overhead. When unused, a buffer takes something like a usize + a pointer.

[jhorstmann]

Good idea! I haven't checked the memory layout of both, but in the common case of 1 buffer this could inline better (in the absence of LTO). Since we already have the validity bitmap in a separate Option, the variants are Zero, One and Many.

[alamb]

@jhorstmann I am closing this PR for the time being to clean up the Rust/Arrow PR backlog. Please let me know if this is a mistake