Improve the performance of low cardinality groupby #16619
Conversation
This file contains 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
PointKernel
added
2 - In Progress
3 tasks
rapids-bot bot
pushed a commit
that referenced
this pull request
Aug 27, 2024
This PR introduces a new `num_multiprocessors` utility and moves the existing `elements_per_thread` host utility to the new `cuda.hpp` header. Needed by #16619. Authors: - Yunsong Wang (https://github.com/PointKernel) Approvers: - David Wendt (https://github.com/davidwendt) - Mark Harris (https://github.com/harrism) URL: #16628
This was referenced Oct 23, 2024
rapids-bot bot
pushed a commit
that referenced
this pull request
Oct 28, 2024
This work is part of splitting the original bulk shared memory groupby PR #16619. This PR introduces the `compute_mapping_indices` API, which is used by the shared memory groupby. libcudf will opt for the shared memory code path when the aggregation request is compatible with shared memory, i.e. there is enough shared memory space and no dictionary aggregation requests. Aggregating with shared memory involves two steps. The first step, introduced in this PR, calculates the offset for each input key within the shared memory aggregation storage, as well as the offset when merging the shared memory results into global memory. Authors: - Yunsong Wang (https://github.com/PointKernel) Approvers: - Mark Harris (https://github.com/harrism) - Nghia Truong (https://github.com/ttnghia) URL: #17147
rapids-bot bot
pushed a commit
that referenced
this pull request
Oct 31, 2024
This work is part of splitting the original bulk shared memory groupby PR #16619. This PR introduces the `compute_shared_memory_aggs` API, which is utilized by the shared memory groupby. The shared memory groupby process consists of two main steps. The first step was introduced in #17147, and this PR implements the second step, where the actual aggregations are performed based on the offsets from the first step. Each thread block is designed to handle up to 128 unique keys. If this limit is exceeded, there won't be enough space to store temporary aggregation results in shared memory, so a flag is set to indicate that follow-up global memory aggregations are needed to complete the remaining aggregation requests. Authors: - Yunsong Wang (https://github.com/PointKernel) Approvers: - David Wendt (https://github.com/davidwendt) - Nghia Truong (https://github.com/ttnghia) - Bradley Dice (https://github.com/bdice) URL: #17162
PointKernel
added
3 - Ready for Review
|
This PR is ready for review. |
davidwendt
approved these changes
Nov 6, 2024
Co-authored-by: David Wendt <45795991+davidwendt@users.noreply.github.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Description
This PR enhances groupby performance for low-cardinality input cases. When applicable, it leverages shared memory for initial aggregation, followed by global memory aggregation to reduce atomic contention and improve performance.
Checklist