Reduce overhead when computing compound hash-based groupby aggregations #20736
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: Nghia Truong <[email protected]>
Signed-off-by: Nghia Truong <[email protected]>
Signed-off-by: Nghia Truong <[email protected]>
ttnghia
added
3 - Ready for Review
ttnghia
commented
Nov 26, 2025
Comment on lines
-96
to
-105
| std::vector<std::unique_ptr<aggregation>> visit( | ||
| data_type, cudf::detail::correlation_aggregation const&) override | ||
| { | ||
| std::vector<std::unique_ptr<aggregation>> aggs; | ||
| aggs.push_back(make_sum_aggregation()); | ||
| // COUNT_VALID | ||
| aggs.push_back(make_count_aggregation()); | ||
|
|
||
| return aggs; | ||
| } |
Contributor
Author
There was a problem hiding this comment.
Unused code: correlation aggregation is not yet supported in hash-based pipeline.
davidwendt
reviewed
Nov 26, 2025
| // If it is not an input aggregation, we can force its output to be non-nullable | ||
| // (by storing `1` value in the `force_non_nullable` vector). | ||
| auto const is_input_agg = input_agg_kinds_set.contains(agg->kind); | ||
| force_non_nullable.push_back(is_input_agg ? 0 : 1); |
Contributor
There was a problem hiding this comment.
Suggested change
| force_non_nullable.push_back(is_input_agg ? 0 : 1); | |
| force_non_nullable.push_back(not is_input_agg); |
Using int (4-bytes) for this vector seems like a waste. Perhaps using int8 instead. Using bool will not work with spans unfortunately.
I also recommend using std::span instead of cudf::host_span if possible.
We can revisit the other host_span usages here in a later PR. Reference: #20539
Contributor
Author
There was a problem hiding this comment.
I've updated the code to use int8_t.
I also tried using cuda::std::span but couldn't compile due to:
error: no suitable user-defined conversion from "std::tuple_element<1UL, const std::tuple<cudf::table_view, cudf::detail::host_vector<cudf::aggregation::Kind>, std::vector<std::unique_ptr<cudf::aggregation, std::default_delete<cudf::aggregation>>, std::allocator<std::unique_ptr<cudf::aggregation, std::default_delete<cudf::aggregation>>>>, std::vector<int8_t, std::allocator<int8_t>>, bool>>::type"
(aka "const std::__tuple_element_t<1UL, std::tuple<cudf::table_view, cudf::detail::host_vector<cudf::aggregation::Kind>, std::vector<std::unique_ptr<cudf::aggregation, std::default_delete<cudf::aggregation>>, std::allocator<std::unique_ptr<cudf::aggregation, std::default_delete<cudf::aggregation>>>>, std::vector<signed char, std::allocator<signed char>>, bool>>")
to "cuda::std::__4::span<const cudf::aggregation::Kind, 18446744073709551615UL>" exists
Contributor
There was a problem hiding this comment.
Since this is host-only, the std::span should be enough. No need for cuda::std::span.
I'm not sure if that will help this though. I can look at this more closely next week.
Contributor
Author
There was a problem hiding this comment.
Ah, I found that the issue is due to implicit converting cudf::detail::host_vector to (cuda::)std::span. The implicit conversion is not implemented yet, so we should do that soon.
When trying to use std::span in place of host_span, the modification chains down to nearly dozen of unrelated files thus I would rather have std::span adoption implemented in other separate PR.
Contributor
There was a problem hiding this comment.
Ok, I tried std::span locally and only had to change 6 files in this PR.
Contributor
There was a problem hiding this comment.
The host_vector should not come into play. Are you trying to change all of the host_span usages?
I would recommend to only change the host_span usage for this new int8 variable and we'll change any of the other host_span usages in a follow up PR.
Contributor
Author
There was a problem hiding this comment.
Yes I tried to change host_span for this new variable and another one, since they are passing to the same function:
std::unique_ptr<table> create_results_table(size_type output_size,
table_view const& values,
host_span<aggregation::Kind const> agg_kinds,
host_span<int8_t const> force_non_nullable,
so it makes more sense to change both together.
Contributor
There was a problem hiding this comment.
I understand. But it would require too many changes.
Whereas changing just the one for this PR will be enough for it to pass.
Please just change the one variable. We can change the other one(s) in a later PR.
Contributor
Author
There was a problem hiding this comment.
Done. std::span looks great now. Excited waiting for the full adoption 👍
Signed-off-by: Nghia Truong <[email protected]>
Signed-off-by: Nghia Truong <[email protected]>
Contributor
Author
|
Benchmark for For |
Signed-off-by: Nghia Truong <[email protected]>
davidwendt
reviewed
Nov 26, 2025
Signed-off-by: Nghia Truong <[email protected]>
Signed-off-by: Nghia Truong <[email protected]>
Signed-off-by: Nghia Truong <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
During computing the compound aggregations in hash-based groupby, we may perform unnecessary work or repeat the same computation multiple times. For example, we may output nulls for the aggregation results and launch a kernel to count nulls for them, although they are computed only to serve as intermediate data for other compound aggregation and their null masks/null counts are never accessed.
In the situations when we have many aggregations (hundred+ to thousand+ and up), such repeated or unnecessary work accumulates to a significant runtime overhead. As such, we want to reduce them as little as possible. This PR intends to do so, modifying the existing hash-based groupby aggregation framework to allow ignoring nulls for the aggregation outputs. By doing so, the results of aggregations requested by the users can have nulls as before (no change) while the aggregations generated only to serve as intermediate data for computing compound aggregations are always non-nullable.
Closes #20734.