Asynchronously fetch byte ranges from multiple sources by mhaseeb123 · Pull Request #22586 · NVIDIA/cudf

mhaseeb123 · 2026-05-19T19:04:12Z

Description

Contributes to #22583. Follow up #22550

This PR adds a new overload of fetch_byte_ranges_to_device_async Parquet IO utility that enables asynchronously fetching byte ranges across multiple sources.

This new API is now being used to read column chunk data by both Parquet and Hybrid scan readers enabling common (pipelining) behavior for multithreaded reads.

Checklist

Run and compare PARQUET_READER_NVBENCH against main. No perf improvement or regression.
Let Reduce mutex scoped region in parquet io utilities #22550 merge and pull in changes
No performance hit to the experiment from this comment
I am familiar with the Contributing Guidelines.
New or existing tests cover these changes.
The documentation is up to date with these changes.

…i-source-parquet-io-utils

copy-pr-bot · 2026-05-19T19:04:16Z

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.

mhaseeb123 · 2026-05-25T23:19:12Z


+}  // namespace detail
+
+std::tuple<std::vector<rmm::device_buffer>,


Moved impl in detail vectorized API. Public APIs just call that wrapped with a CUDF_FUNC_RANGE() like every other libcudf API

Co-authored-by: Bradley Dice <bdice@bradleydice.com>

mhaseeb123 · 2026-05-26T22:50:05Z

@vuule added you as reviewer as you just reviewed its predecessor PR (same thing here except vectorized and using another mutex to serialize host reads too across threads)

vuule · 2026-05-28T21:26:17Z

+  // Schedule host reads holding the `host_read_mutex` so that all reads for a caller thread
+  // are scheduled without interleaving with reads from other threads yielding better pipelining
+  {
+    std::scoped_lock<std::mutex> lock(host_read_mutex);


if I'm reading this right, we now have a mutex around read_parquet IO as well? Have you checked the impact on the multithreaded benchmark?

Your understanding is correct. I measured performance of Parquet multithreaded benchmark and saw no difference (See #22586 (comment)). I was expecting a speed improvement though.

I also ran this experiment #22529 (comment) again with this PR and compared the profiles and saw no difference

mhaseeb123 · 2026-05-28T23:05:09Z

Performance Impact to PARQUET_MULTITHREAD_READER_NVBENCH

Summary: No impact. Everything within noise threshold.

`parquet_multithreaded_read_decode_fixed_width`

cardinality	io_type	num_cols	num_iterations	num_threads	run_length	total_data_size	main (ms)	PR (ms)	Δ	noise(main/PR)
1000	PINNED_BUFFER	4	1	1	8	1073741824	14.334	14.603	+1.88%	0.41%/1.03%
1000	PINNED_BUFFER	4	1	2	8	1073741824	14.002	14.246	+1.74%	1.68%/1.79%
1000	PINNED_BUFFER	4	1	4	8	1073741824	14.734	15.033	+2.03%	5.11%/4.68%
1000	PINNED_BUFFER	4	1	8	8	1073741824	16.532	16.676	+0.87%	7.18%/6.57%

`parquet_multithreaded_read_decode_list`

cardinality	io_type	num_cols	num_iterations	num_threads	run_length	total_data_size	main (ms)	PR (ms)	Δ	noise(main/PR)
1000	PINNED_BUFFER	4	1	1	8	1073741824	29.630	29.837	+0.70%	0.87%/0.81%
1000	PINNED_BUFFER	4	1	2	8	1073741824	28.109	28.047	-0.22%	2.70%/4.85%
1000	PINNED_BUFFER	4	1	4	8	1073741824	27.215	27.480	+0.97%	7.30%/9.90%
1000	PINNED_BUFFER	4	1	8	8	1073741824	35.343	35.172	-0.49%	14.16%/13.49%

`parquet_multithreaded_read_decode_mixed`

cardinality	io_type	num_cols	num_iterations	num_threads	run_length	total_data_size	main (ms)	PR (ms)	Δ	noise(main/PR)
1000	PINNED_BUFFER	4	1	1	8	1073741824	12.903	13.042	+1.07%	0.67%/1.61%
1000	PINNED_BUFFER	4	1	2	8	1073741824	13.045	13.376	+2.53%	6.03%/6.13%
1000	PINNED_BUFFER	4	1	4	8	1073741824	14.900	15.040	+0.94%	6.66%/8.13%
1000	PINNED_BUFFER	4	1	8	8	1073741824	18.373	18.664	+1.58%	10.06%/9.21%

`parquet_multithreaded_read_decode_string`

cardinality	io_type	num_cols	num_iterations	num_threads	run_length	total_data_size	main (ms)	PR (ms)	Δ	noise(main/PR)
1000	PINNED_BUFFER	4	1	1	8	1073741824	13.568	13.789	+1.64%	0.50%/1.22%
1000	PINNED_BUFFER	4	1	2	8	1073741824	14.126	14.149	+0.17%	6.52%/5.48%
1000	PINNED_BUFFER	4	1	4	8	1073741824	15.821	15.937	+0.73%	14.11%/14.23%
1000	PINNED_BUFFER	4	1	8	8	1073741824	20.153	20.702	+2.72%	8.61%/8.47%

Co-authored-by: Yunsong Wang <12716979+PointKernel@users.noreply.github.com>

PointKernel

One last comment but non blocking.

…e sources (#22613) Contributes to #22583. Follow up #22586 This PR adds new overloads of remaining Parquet IO utility to fetch parquet footers and page index bytes from multiple sources. Authors: - Muhammad Haseeb (https://github.com/mhaseeb123) Approvers: - Qi Chen (https://github.com/qbacpey) - Yunsong Wang (https://github.com/PointKernel) URL: #22613

mhaseeb123 · 2026-06-02T01:23:09Z

/merge

vyasr

Wow I'm sorry I started reviewing here and then completely forgot to submit. Looks like almost all of my requests have already been addressed so please proceed with merging. I'll just leave the question.

vyasr · 2026-05-30T01:58:35Z

 {
-  static std::mutex mutex;
+  static std::mutex host_read_mutex;
+  static std::mutex device_read_mutex;


This is implementing what we discussed in #22550, right?

mhaseeb123 · 2026-06-03T02:38:27Z

This is implementing what we discussed in #22550, right?

Yes :)

mhaseeb123 and others added 10 commits May 15, 2026 22:43

Fix the parquet io utils

41d27df

Remove unnecessary comments

dbbdea2

Remove unnecessary sync

5a95c7e

Minor updates

410d495

Minor update

96e1778

Minor improvement

7ae60be

Merge branch 'release/26.06' into fix/parquet-io-utils-stream

0008932

Optimize parquet io utils

f552803

Add and use multisource parquet io utils

da4c404

Merge branch 'main' of https://github.com/rapidsai/cudf into fea/mult…

7b000c5

…i-source-parquet-io-utils

github-actions Bot assigned mhaseeb123 May 19, 2026

github-actions Bot added the libcudf Affects libcudf (C++/CUDA) code. label May 19, 2026

mhaseeb123 added 2 - In Progress Currently a work in progress cuIO cuIO issue improvement Improvement / enhancement to an existing function non-breaking Non-breaking change labels May 19, 2026

Minor variable name improvement

683acc9

This was referenced May 20, 2026

Parquet utilities to fetch footer and page index buffers from multiple sources #22613

Merged

Add hybrid scan multifile reader basics #22616

Merged

Reduce mutex scoped region in parquet io utilities #22550

Merged

mhaseeb123 added 2 commits May 22, 2026 15:43

Merge branch 'main' into fea/multi-source-parquet-io-utils

407ef0d

Update

a7928f1

mhaseeb123 marked this pull request as ready for review May 25, 2026 23:08

mhaseeb123 requested a review from a team as a code owner May 25, 2026 23:08

mhaseeb123 requested a review from PointKernel May 25, 2026 23:08

Merge branch 'main' into fea/multi-source-parquet-io-utils

bfbd7e1

mhaseeb123 requested a review from davidwendt May 25, 2026 23:08

mhaseeb123 added 3 - Ready for Review Ready for review by team and removed 2 - In Progress Currently a work in progress labels May 25, 2026

mhaseeb123 commented May 25, 2026

View reviewed changes

bdice approved these changes May 26, 2026

View reviewed changes

Comment thread cpp/src/io/parquet/io_utils/parquet_io_utils.cpp

Comment thread cpp/src/io/parquet/io_utils/parquet_io_utils.cpp Outdated

Comment thread cpp/src/io/parquet/reader_impl_preprocess_utils.cu

Update cpp/src/io/parquet/io_utils/parquet_io_utils.cpp

ca2e3f8

Co-authored-by: Bradley Dice <bdice@bradleydice.com>

mhaseeb123 requested a review from vuule May 26, 2026 22:49

vuule reviewed May 28, 2026

View reviewed changes

mhaseeb123 requested review from qbacpey and vuule May 28, 2026 23:09

Merge branch 'main' into fea/multi-source-parquet-io-utils

fc055b1

PointKernel reviewed Jun 1, 2026

View reviewed changes

Comment thread cpp/src/io/parquet/io_utils/parquet_io_utils.cpp Outdated

Comment thread cpp/src/io/parquet/io_utils/parquet_io_utils.cpp Outdated

Comment thread cpp/src/io/parquet/reader_impl_preprocess_utils.cu

Update cpp/src/io/parquet/io_utils/parquet_io_utils.cpp

867341e

Co-authored-by: Yunsong Wang <12716979+PointKernel@users.noreply.github.com>

mhaseeb123 requested a review from PointKernel June 1, 2026 23:22

Merge branch 'main' into fea/multi-source-parquet-io-utils

a2c4f0c

PointKernel approved these changes Jun 2, 2026

View reviewed changes

Comment thread cpp/src/io/parquet/io_utils/parquet_io_utils.cpp Outdated

mhaseeb123 added 2 commits June 1, 2026 18:04

Merge branch 'main' into fea/multi-source-parquet-io-utils

72b579f

Address comments

3a56e67

mhaseeb123 added 5 - Ready to Merge Testing and reviews complete, ready to merge and removed 3 - Ready for Review Ready for review by team labels Jun 2, 2026

vuule approved these changes Jun 2, 2026

View reviewed changes

mhaseeb123 added 2 commits June 2, 2026 12:55

Merge branch 'main' into fea/multi-source-parquet-io-utils

72d5da9

Merge branch 'main' into fea/multi-source-parquet-io-utils

78d2bb4

vyasr approved these changes Jun 3, 2026

View reviewed changes

rapids-bot Bot merged commit d612f5c into NVIDIA:main Jun 3, 2026
215 of 219 checks passed

mhaseeb123 deleted the fea/multi-source-parquet-io-utils branch June 3, 2026 02:38

coderabbitai Bot mentioned this pull request Jun 13, 2026

Add hybrid scan multifile two-step materializers #22833

Merged

4 tasks

coderabbitai Bot mentioned this pull request Jun 26, 2026

Fix hybrid scan parquet reader incorrectly parse bloom filter #22901

Merged

3 tasks


		} // namespace detail

		std::tuple<std::vector<rmm::device_buffer>,

Uh oh!

Conversation

mhaseeb123 commented May 19, 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 May 19, 2026

Uh oh!

mhaseeb123 May 25, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

mhaseeb123 commented May 26, 2026

Uh oh!

vuule May 28, 2026

Choose a reason for hiding this comment

Uh oh!

mhaseeb123 May 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mhaseeb123 May 28, 2026

Choose a reason for hiding this comment

Uh oh!

mhaseeb123 commented May 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Performance Impact to PARQUET_MULTITHREAD_READER_NVBENCH

parquet_multithreaded_read_decode_fixed_width

parquet_multithreaded_read_decode_list

parquet_multithreaded_read_decode_mixed

parquet_multithreaded_read_decode_string

Uh oh!

Uh oh!

Uh oh!

Uh oh!

PointKernel left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

mhaseeb123 commented Jun 2, 2026

Uh oh!

vyasr left a comment

Choose a reason for hiding this comment

Uh oh!

vyasr May 30, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

mhaseeb123 commented Jun 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

mhaseeb123 commented May 19, 2026 •

edited

Loading

mhaseeb123 May 28, 2026 •

edited

Loading

mhaseeb123 commented May 28, 2026 •

edited

Loading

`parquet_multithreaded_read_decode_fixed_width`

`parquet_multithreaded_read_decode_list`

`parquet_multithreaded_read_decode_mixed`

`parquet_multithreaded_read_decode_string`