Add column projection resolution benchmarks for the Parquet readers by qbacpey · Pull Request #24193 · NVIDIA/cudf

qbacpey · 2026-09-16T10:02:56Z

Description

Split out of #24162

Adds a new axis for an existed benchmark, both of them are about column-name resolution

Checklist

I am familiar with the Contributing Guidelines.
New or existing tests cover these changes (benchmark-only change; no production code modified).
The documentation is up to date with these changes.

copy-pr-bot · 2026-09-16T10:03:00Z

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

qbacpey · 2026-09-16T14:51:36Z

Benchmark results

Machine: NVIDIA GH200 480GB (132-SM Hopper, Neoverse-V2 CPU), driver 580.95.05, CUDA 13.3. All runs with CUDF_BENCHMARK_DROP_CACHE=1.

Base = main;
PR = this branch after merging upstream/main.

Raw nvbench output — PR branch, parquet_column_selection

| io_type  | selection_method | num_cols | Samples |  CPU Time  | Noise |  GPU Time  | Noise | cols_per_sec | peak_memory_usage |
|----------|------------------|----------|---------|------------|-------|------------|-------|--------------|-------------------|
| FILEPATH |             none |       64 |   5744x | 106.029 us | 2.67% |  87.200 us | 3.01% |       733941 |        44.561 KiB |
| FILEPATH |            names |       64 |   4304x | 136.041 us | 3.21% | 116.289 us | 3.48% |       550352 |        44.561 KiB |
| FILEPATH |             none |      512 |    880x | 673.656 us | 1.91% | 651.361 us | 1.98% |       786045 |       345.907 KiB |
| FILEPATH |            names |      512 |    384x |   1.377 ms | 1.18% |   1.354 ms | 1.19% |       378118 |       345.907 KiB |
| FILEPATH |             none |     2048 |    864x |   3.283 ms | 1.88% |   3.259 ms | 1.89% |       628432 |         1.385 MiB |
| FILEPATH |            names |     2048 |     50x |  10.090 ms | 0.39% |  10.064 ms | 0.39% |       203501 |         1.385 MiB |

… produced (#24191) Split out of #24162. This PR is the shared groundwork that #24192 and #24193 build on. It also sets `max_page_fragment_size` and raises `max_page_size_bytes` to ensure benchmark `parquet_read_file_shape` can actually generate requested Parquet shape. Authors: - Qi Chen (https://github.com/qbacpey) Approvers: - Vukasin Milovanovic (https://github.com/vuule) - Kyle Edwards (https://github.com/KyleFromNVIDIA) URL: #24191

copy-pr-bot · 2026-09-29T11:31:24Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

Moves the writer inlined in BM_parquet_filter_name_resolution into write_named_resolution_parquet_file so the projection benchmarks can resolve against the same deterministically named schema.

hybrid_scan_projection times the filter and payload column chunk byte range calls across a side axis; parquet_read_column_projection is the naive reader baseline, timing chunked reader construction with a full explicit projection.

Introduces the `named_resolution_column_names` function to generate deterministic column names for Parquet files. This function is utilized in `write_named_resolution_parquet_file` and various benchmarks to ensure consistent naming across different Parquet operations, enhancing code clarity and maintainability.

coderabbitai · 2026-09-30T15:50:04Z

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

@coderabbitai resume to resume automatic reviews.
@coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

▶️ Resume reviews
🔍 Trigger review

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository: NVIDIA/cudf/.coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: a2016fe2-6161-4c66-b86a-9ac9e77d2d99

📥 Commits

Reviewing files that changed from the base of the PR and between df8428a and 0178f4b.

📒 Files selected for processing (2)

cpp/benchmarks/io/parquet/experimental/hybrid_scan/hybrid_scan_projection.cpp
cpp/benchmarks/io/parquet/parquet_reader_metadata.cpp

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.

📝 Summary

Summary by CodeRabbit

Chores
- Expanded Parquet benchmark coverage to include hybrid-scan payload-column selection across schemas with 64 to 4,096 columns.
- Added comparisons of positional and name-based column selection in Parquet reader benchmarks.
- Benchmarks report schema column counts and peak memory usage. Parquet filter-name-resolution results now report schema column counts instead of columns processed per second.

Walkthrough

Adds Parquet reader-selection and hybrid-scan projection benchmarks. The reader-selection benchmark covers positional and name-based selection. The filter-name-resolution benchmark now reports schema column count instead of throughput.

Changes

Parquet Projection Benchmarks

Layer / File(s)	Summary
Reader selection and metadata benchmarks `cpp/benchmarks/io/parquet/parquet_reader_metadata.cpp`	Adds positional and name-based selection cases to the column-selection benchmark. The filter-name-resolution benchmark reports `schema_columns` instead of `cols_per_sec`; peak-memory reporting remains.
Hybrid-scan projection benchmark `cpp/benchmarks/io/parquet/experimental/hybrid_scan/hybrid_scan_projection.cpp`, `cpp/benchmarks/CMakeLists.txt`	Adds and registers a benchmark that writes a configurable-width Parquet table, prepares filter column chunks, and times payload column-chunk selection. It reports schema column count and peak memory usage.

Priority: ⬇️ Low

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Other

Suggested reviewers: vyasr, pointkernel

Merge Risk: ⚪ Minimal · up to 0178f

This change adds a registered projection benchmark and changes the filter benchmark’s reported metric. No concrete repository-internal breakage or other actionable merge risk is established.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 42.86% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 7 functions across 4 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly summarizes the main changes: it adds Parquet reader benchmarks for column projection resolution, including the new hybrid scan benchmark and the expanded existing benchmark.
Description check	✅ Passed	The description is directly related to the changeset. It identifies the new benchmark axis and the focus on column-name resolution.

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create a new PR

Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Autopilot is currently an internal CodeRabbit preview.

_{Comment @coderabbitai help to get the list of available commands.}

mhaseeb123 · 2026-09-30T19:11:57Z

+using projection_sides = nvbench::enum_type_list<projection_side::FILTER,
+                                                 projection_side::PAYLOAD,
+                                                 projection_side::PAYLOAD_EXPLICIT>;
+
+NVBENCH_BENCH_TYPES(BM_hybrid_scan_projection, NVBENCH_TYPE_AXES(projection_sides))
+  .set_name("hybrid_scan_projection")
+  .set_type_axes_names({"side"})
+  .set_min_samples(4)
+  .add_int64_axis("num_cols", {64, 512, 2048, 4096});


Do we really need enums here. We could just use a string axis.

mhaseeb123 · 2026-09-30T19:13:44Z

+{
+  auto const num_cols = static_cast<cudf::size_type>(state.get_int64("num_cols"));
+
+  auto source_sink = write_named_resolution_parquet_file(num_cols, io_type::FILEPATH);


By default, the parquet writer writes column names as: _col0, _col1, .., ... Can we just use that instead of explicitly writing col0, col1... names? We can also select columns by index as well if needed. We don't need the two new helpers in parquet_common.xpp that way

mhaseeb123 · 2026-09-30T19:18:51Z

+  // Caller-supplied payload list
+  if constexpr (Side == projection_side::PAYLOAD_EXPLICIT) {
+    read_opts_builder.column_names(named_resolution_column_names(num_cols));
+  }
+  auto const read_opts = read_opts_builder.build();


Is PAYLOAD_EXPLICIT even needed? Isn't the column selection cost just proportional to the number of columns being selected? Also please check if any of the existing parquet benchmarks cover column selection. If so, then PAYLOAD_EXPLICIT as well as PAYLOAD modes are just duplicates and need to be removed

Removed PAYLOAD_EXPLICIT.

None of the existing benchmarks cover the payload case (a filter and no projection):

No other benchmark calls payload_column_chunks_byte_ranges; the hybrid scan benchmarks time the filters or read in a single step through all_column_chunks_byte_ranges

parquet_column_selection passes no names, so it times the linear path

parquet_read_column_selection times a full read of a few dozen columns, where selection is negligible

Selection cost is only proportional to the number of columns when no names are passed. With names, select_columns looks up each one with a linear find_if over every schema path, so the cost is roughly selected columns × schema width. With a filter and no projection, select_payload_columns turns every non-filter column into a name and passes all n−1 of them to that lookup, while the regular reader in the same situation passes no names and selects all columns in one pass.

Same build of this branch, GH200:

num_cols hybrid payload selection regular reader, filter, no projection regular reader, all names projected

64 0.130 ms 0.079 ms 0.099 ms

512 1.45 ms 0.72 ms 1.21 ms

2048 10.58 ms 2.46 ms 9.08 ms

4096 38.52 ms 5.20 ms 32.64 ms

The columns come from hybrid_scan_projection, parquet_filter_name_resolution (case_sensitive=1, heavy_filter=0) and parquet_read_column_projection.

Should we further drop the the payload case (a filter and no projection)?

mhaseeb123 · 2026-09-30T19:20:34Z

+  cudf::ast::tree filter_tree;
+  auto const& col_ref = filter_tree.push(cudf::ast::column_name_reference("col0"));
+  auto const& lit     = filter_tree.push(cudf::ast::literal(filter_literal));
+  auto const& filter_expr =
+    filter_tree.push(cudf::ast::operation(cudf::ast::ast_operator::GREATER_EQUAL, col_ref, lit));


The filter AST tree seems trivial (single predicate) so the filter column selection cost would be trivial. Can we use something like here 4a59ca5#diff-c9f9907616a2121d2c0cd6dd2d50126a42964461acf05132bfe59db0b67b2a21 to build a varying depth/columns AST tree or skip this altogether as filter/payload columns ratio is usually small in real world.

Agreed, dropped it since parquet_filter_name_resolution sweeps heavy filters for the regular reader.

…ing metadata handling - Removed the `named_resolution_column_names` and `write_named_resolution_parquet_file` functions from `parquet_common.cpp` and `parquet_common.hpp`. - Updated benchmarks in `parquet_reader_metadata.cpp` and `hybrid_scan_projection.cpp` to directly handle column names and metadata without the removed functions. - Adjusted CMake configuration to reflect the changes in benchmark source files. This cleanup enhances code maintainability and reduces unnecessary complexity in the benchmark implementations.

mhaseeb123 · 2026-10-01T23:19:42Z

+               // Select the filter columns first, as a real read does, so the timed call covers
+               // only the payload selection
+               std::ignore = reader->filter_column_chunks_byte_ranges(row_groups, read_opts);


I think this might not be needed anymore. The hybrid_scan_impl::select_columns() now computes the filter column names (if not already done and filter is present) before deducing payload columns. Please check and update 🙂

mhaseeb123 · 2026-10-01T23:20:03Z

+
+  state.exec(nvbench::exec_tag::sync | nvbench::exec_tag::timer,
+             [&](nvbench::launch& launch, auto& timer) {
+               drop_page_cache_if_enabled(source_info.filepaths());


Don't think this is needed either

mhaseeb123 · 2026-10-01T23:23:30Z

    mem_stats_logger.peak_memory_usage(), "peak_memory_usage", "peak_memory_usage");
 }

+// Benchmark full-projection column-name resolution during naive reader construction.


Can we just add an axis (selection method?) to the other benchmark instead where we either provide column selection or not instead of this separate benchmark?

Done, parquet_column_selection now has a selection_method axis

…mn name selection. Removed unused benchmark for full-projection column-name resolution. Updated benchmark parameters to include selection method options.

…akeLists.txt. This cleanup eliminates unused code and simplifies the benchmark setup.

qbacpey · 2026-10-02T19:42:31Z

/ok to test 5f4e0f0

qbacpey · 2026-10-02T19:42:37Z

/merge

github-actions Bot assigned qbacpey Sep 16, 2026

github-actions Bot added libcudf Affects libcudf (C++/CUDA) code. CMake CMake build issue labels Sep 16, 2026

qbacpey mentioned this pull request Sep 16, 2026

Tighten Parquet page config in benchmarks so requested page layout is produced #24191

Merged

3 tasks done

qbacpey changed the title ~~Bench/hybrid scan projection~~ Add column projection resolution benchmarks for the Parquet readers Sep 16, 2026

qbacpey added improvement Improvement / enhancement to an existing function non-breaking Non-breaking change labels Sep 16, 2026

qbacpey force-pushed the bench/hybrid-scan-projection branch 3 times, most recently from 94ee62d to eb6a7b2 Compare September 16, 2026 12:30

qbacpey added the 2 - In Progress Currently a work in progress label Sep 16, 2026

qbacpey mentioned this pull request Sep 16, 2026

Add metadata and setup benchmarks for the hybrid scan Parquet reader #24162

Closed

3 tasks done

qbacpey force-pushed the bench/hybrid-scan-projection branch from eb6a7b2 to f0d6376 Compare September 16, 2026 15:11

qbacpey force-pushed the bench/hybrid-scan-projection branch from f0d6376 to a23494a Compare September 29, 2026 11:31

qbacpey and others added 3 commits September 30, 2026 14:41

Share the named column Parquet file writer via parquet_common

b931c7a

Moves the writer inlined in BM_parquet_filter_name_resolution into write_named_resolution_parquet_file so the projection benchmarks can resolve against the same deterministically named schema.

qbacpey force-pushed the bench/hybrid-scan-projection branch from a23494a to bbe33c6 Compare September 30, 2026 15:28

qbacpey requested a review from mhaseeb123 September 30, 2026 15:45

Merge branch 'main' into bench/hybrid-scan-projection

6735d0d

qbacpey marked this pull request as ready for review September 30, 2026 15:45

qbacpey requested review from a team as code owners September 30, 2026 15:45

qbacpey requested a review from ttnghia September 30, 2026 15:45

qbacpey added 3 - Ready for Review Ready for review by team and removed 2 - In Progress Currently a work in progress labels Sep 30, 2026

mhaseeb123 requested changes Sep 30, 2026

View reviewed changes

qbacpey and others added 3 commits October 1, 2026 18:39

Merge branch 'main' into bench/hybrid-scan-projection

c12098c

Merge branch 'main' into bench/hybrid-scan-projection

ac9c740

qbacpey requested a review from mhaseeb123 October 1, 2026 18:12

mhaseeb123 requested changes Oct 1, 2026

View reviewed changes

qbacpey and others added 2 commits October 2, 2026 10:40

Merge branch 'main' into bench/hybrid-scan-projection

df8428a

Enhance parquet column selection benchmark by adding support for colu…

0178f4b

…mn name selection. Removed unused benchmark for full-projection column-name resolution. Updated benchmark parameters to include selection method options.

qbacpey requested a review from mhaseeb123 October 2, 2026 11:35

mhaseeb123 approved these changes Oct 2, 2026

View reviewed changes

mhaseeb123 reviewed Oct 2, 2026

View reviewed changes

Comment thread cpp/benchmarks/io/parquet/experimental/hybrid_scan/hybrid_scan_projection.cpp Outdated

ttnghia approved these changes Oct 2, 2026

View reviewed changes

anon and others added 2 commits October 2, 2026 19:39

Remove hybrid scan projection benchmark and its configuration from CM…

24fc6ea

…akeLists.txt. This cleanup eliminates unused code and simplifies the benchmark setup.

Merge branch 'main' into bench/hybrid-scan-projection

5f4e0f0

rapids-bot Bot merged commit baaaea9 into NVIDIA:main Oct 2, 2026
72 of 74 checks passed

num_cols	hybrid payload selection	regular reader, filter, no projection	regular reader, all names projected
64	0.130 ms	0.079 ms	0.099 ms
512	1.45 ms	0.72 ms	1.21 ms
2048	10.58 ms	2.46 ms	9.08 ms
4096	38.52 ms	5.20 ms	32.64 ms

Conversation

qbacpey commented Sep 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Checklist

Uh oh!

copy-pr-bot Bot commented Sep 16, 2026

Uh oh!

qbacpey commented Sep 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Benchmark results

Uh oh!

copy-pr-bot Bot commented Sep 29, 2026

Uh oh!

coderabbitai Bot commented Sep 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Reviews paused

Summary by CodeRabbit

Walkthrough

Changes

❌ Failed checks (1 warning)

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

qbacpey commented Oct 2, 2026

Uh oh!

qbacpey commented Oct 2, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

qbacpey commented Sep 16, 2026 •

edited

Loading

qbacpey commented Sep 16, 2026 •

edited

Loading

coderabbitai Bot commented Sep 30, 2026 •

edited

Loading