Skip to content

[GLUTEN-12303][VL] Support async multipart upload for S3 writes#12305

Open
ReemaAlzaid wants to merge 13 commits into
apache:mainfrom
ReemaAlzaid:multi-threaded-s3
Open

[GLUTEN-12303][VL] Support async multipart upload for S3 writes#12305
ReemaAlzaid wants to merge 13 commits into
apache:mainfrom
ReemaAlzaid:multi-threaded-s3

Conversation

@ReemaAlzaid

@ReemaAlzaid ReemaAlzaid commented Jun 16, 2026

Copy link
Copy Markdown
Contributor

The PR is ported from facebookincubator/velox#14472. The original author is @weixiuli

What changes were proposed in this pull request?

This PR adds Gluten-side support for multi-threaded asynchronous multipart upload for S3 compatible object storage in the Velox backend.

When enabled, Gluten uses an async S3 write path that uploads multipart upload parts through a shared upload thread pool. The number of in-flight part uploads per file and the upload thread pool size are configurable.

The existing synchronous S3 write path remains the default behavior.

This PR also adds a benchmark for comparing synchronous and asynchronous S3 multipart upload performance.
#12303

How was this patch tested?

Built the Velox backend with S3 enabled and ran the S3 async upload benchmark against MinIO.

Example benchmark command:

export AWS_ACCESS_KEY_ID=minioadmin
export AWS_SECRET_ACCESS_KEY=minioadmin
export GLUTEN_S3_BENCH_BUCKET=writedata
export GLUTEN_S3_BENCH_ENDPOINT=http://127.0.0.1:9000
export GLUTEN_S3_BENCH_REGION=us-east-1
export GLUTEN_S3_BENCH_PATH_STYLE_ACCESS=true
export GLUTEN_S3_BENCH_SSL_ENABLED=false
export GLUTEN_S3_BENCH_MAX_CONCURRENCY=8
export GLUTEN_S3_BENCH_UPLOAD_THREADS=8
export GLUTEN_S3_BENCH_MIN_PART_SIZE=32MB

./cpp/build/velox/benchmarks/s3_async_upload_benchmark \
  --bm_min_iters=1 \
  --bm_regex='(sync|async)_upload_(16|32|64|128|256|512|1024)M'
============================================================================
[...]/benchmarks/S3AsyncUploadBenchmark.cc     relative  time/iter   iters/s
============================================================================  
sync_upload_16M                                            80.58ms     12.41
async_upload_16M                                128.28%    62.81ms     15.92
sync_upload_32M                                           117.05ms      8.54
async_upload_32M                                93.949%   124.59ms      8.03
sync_upload_64M                                           204.63ms      4.89
async_upload_64M                                132.00%   155.03ms      6.45
sync_upload_128M                                          360.72ms      2.77
async_upload_128M                               135.32%   266.56ms      3.75
sync_upload_256M                                          709.47ms      1.41
async_upload_256M                               203.33%   348.93ms      2.87
sync_upload_512M                                             1.50s   667.72m
async_upload_512M                               237.77%   629.87ms      1.59
sync_upload_1024M                                            2.90s   345.41m
async_upload_1024M                              313.82%   922.55ms      1.08

@FelixYBW

Copy link
Copy Markdown
Contributor

FYI, @weixiuli. Since the PR isn't imported into Velox. We extend Velox's S3 file system and move your enhancement into Gluten.

@zhouyuan

Copy link
Copy Markdown
Member

@ReemaAlzaid the configuration issue should be re-generated by https://github.com/apache/gluten/blob/main/dev/gen-all-config-docs.sh

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Adds an optional Velox-backend S3 write path that uploads multipart parts asynchronously using a shared upload thread pool, along with new Spark/Gluten configs and documentation to control concurrency.

Changes:

  • Introduces spark.gluten.velox.s3UploadPartAsync, spark.gluten.velox.s3MaxConcurrentUploadNum, and spark.gluten.velox.s3UploadThreads (plus docs) to enable/tune async multipart uploads.
  • Extends native config extraction to populate corresponding hive.s3.* keys and adds a focused C++ unit test.
  • Adds a new s3_async_upload_benchmark to compare sync vs async upload performance.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 3 comments.

Show a summary per file
File Description
docs/velox-configuration.md Documents new S3 async upload knobs (and reorders/updates a couple of config rows).
docs/get-started/VeloxS3.md Adds a section describing how to enable and tune async multipart uploads.
cpp/velox/utils/ConfigExtractor.cc Maps the new Spark/Gluten configs into Velox hive.s3.* configuration.
cpp/velox/tests/GlutenS3FileSystemTest.cc Adds a unit test validating extraction of the new async upload configs.
cpp/velox/filesystem/GlutenS3FileSystem.h Extends the S3 FS wrapper with async-upload-related state.
cpp/velox/filesystem/GlutenS3FileSystem.cc Implements async multipart upload write path and config handling.
cpp/velox/benchmarks/S3AsyncUploadBenchmark.cc New benchmark for sync vs async upload throughput/latency comparisons.
cpp/velox/benchmarks/CMakeLists.txt Registers the new benchmark behind ENABLE_S3.
backends-velox/src/main/scala/org/apache/gluten/config/VeloxConfig.scala Adds the three new Spark configs to the Velox config registry.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Comment thread cpp/velox/filesystem/GlutenS3FileSystem.cc
Comment on lines +355 to +359
static std::shared_ptr<folly::CPUThreadPoolExecutor> uploadThreadPool(uint32_t uploadThreads) {
std::lock_guard<std::mutex> l(uploadThreadPoolMutex_);
if (!sharedUploadThreadPool_) {
sharedUploadThreadPool_ = std::make_shared<folly::CPUThreadPoolExecutor>(
uploadThreads, std::make_shared<folly::NamedThreadFactory>("s3-upload-thread"));
Comment on lines +336 to +340
RECORD_METRIC_VALUE(filesystems::kMetricS3StartedUploads);
uploadPart({currentPart_->data(), currentPart_->size()}, true);
waitForAsyncUploads();
VELOX_CHECK_EQ(uploadState_.partNumber, uploadState_.completedParts.size());
completeMultipartUpload();
@zhouyuan zhouyuan requested a review from JkSelf July 2, 2026 11:08
ReemaAlzaid and others added 3 commits July 2, 2026 18:05
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants