Skip to content

Conversation

@aranadive
Copy link
Contributor

What?

This PR introduces the S3 CRT client as an additional S3 client type for the object plugin to improve performance for large object transfers by automatically using multipart uploads and ranged GET requests, which increase throughput for large KV-to-object transfers. Both S3Client and S3CRTClient are supported and selected via a new backend parameter, crtMinLimit, so that when the object size exceeds this threshold the S3CRTClient is used. Shared logic between S3 and S3 CRT flows is refactored into common files, and the Dockerfiles are updated to compile and include the S3 CRT client in the build.

Why?

The S3 CRT client is designed for high-throughput transfers and leverages multipart upload and byte-range fetches automatically, which yields better performance and reliability for large objects compared with the standard S3 client. Using it only above a configurable size threshold via crtMinLimit allows the system to gain these benefits for large transfers while keeping behavior unchanged for smaller objects and minimizing unnecessary resource usage.

How?

  • Add crtMinLimit as a backend configuration parameter to determine when to route operations to S3CRTClient instead of S3Client based on object size.
  • Instantiate and use S3CRTClient for objects larger than crtMinLimit, taking advantage of automatic multipart uploads and ranged GETs for improved throughput.
  • Refactor common S3 logic into shared modules used by both clients and update Dockerfiles so the build process compiles and packages the S3 CRT client and its runtime dependencies.

@github-actions
Copy link

👋 Hi aranadive! Thank you for contributing to ai-dynamo/nixl.

Your PR reviewers will review your contribution then trigger the CI to test your changes.

🚀

@aranadive
Copy link
Contributor Author

/build

@aranadive aranadive marked this pull request as ready for review December 15, 2025 08:48
Lets use S3 CRT Client as an additional client type when creating the
object plugin. The CRT Client generally performs better for large
objects by supporting multi-parting automatically and ranged GETs.
This increases throughput and will help large KV transfers to objects.

Both the S3Client and S3CRTClient are used based on a new
backend parameter called crtMinLimit. This determines the client type
used of the object size to be transfered is larger than this value, then
the S3CRTClient is used.

Since S3 and S3CRT are very similar API wise, most of the common code is
refactored into separate files. Also, Dockerfiles are modified to
compile the s3crt client as well.

Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
By default the tests use small object sizes, I modified the transfer
handler to create larger objects based on parameters passed. The CRT
tests will use these when testing.

Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
Move the common utilities for S3 and S3CRT to the utils folder. Other
S3-based object plugins can link to the utils lib as required.

Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
Some more reorg of the object plugin to let each object storage api have
its own directory. Should make it easier to onboard other S3-type
plugins.

Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
@aranadive
Copy link
Contributor Author

/build

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.

2 participants