Throughput improvements to the AWS SDK S3 client were released in v3.649.0

[kuhe]

Describe the feature

Investigate and improve throughput performance of the S3Client in this SDK.

Use Case

Improve throughput of parallel or high-volume of operations, including S3 HeadObject.

Proposed Solution

improve request handler creation time

• handler cache feat(smithy-client): add handler cache smithy-lang/smithy-typescript#1383

avoid FS / loadNodeConfig calls in the request path

• fix(endpoints): resolve service-specific endpoint once per client smithy-lang/smithy-typescript#1382

improve endpoint resolution time

• endpoint cache feat(util-endpoint): add endpoint ruleset cache smithy-lang/smithy-typescript#1385

defer calls to socket.setKeepAlive and other socket event listener attachments

• feat(node-http-handler): defer socket event listener attachment smithy-lang/smithy-typescript#1384
• fix(node-http-handler): call socket methods if socket is defined smithy-lang/smithy-typescript#1392

Acknowledgements

• I may be able to implement this feature request

SDK version used

3.600.0+

Environment details (OS name and version, etc.)

Node.js

[kuhe]

The set of changes were released in https://github.com/aws/aws-sdk-js-v3/releases/tag/v3.649.0.

https://www.npmjs.com/package/@aws-sdk/client-s3/v/3.649.0

There is an additional client configuration field related to this.

  const s3 = new S3Client({
    cacheMiddleware: true
  });

cacheMiddleware caches the middleware resolver function stack per Client+Command combination.

When, for example, making 400x HeadObject requests with the same S3Client, this can be enabled for a performance boost. The caveat is that modification of the middlewareStack on client or Command after the first request will be ignored.

[kuhe]

The expected performance profile of S3 HeadObject with e.g. parallel batch size 10 and total size 400 objects is that with v3.649.0 and cacheMiddleware=true, the v3 SDK runs at a similar throughput to the v2 SDK.

sdk v2 x 55.51 ops/sec ±0.57% (87 runs sampled)
sdk v3 x 55.62 ops/sec ±0.75% (87 runs sampled)
Fastest is sdk v3
- sdk v2 is 1.00x slower

Using https://github.com/bolt-juri-gavshin/aws-sdk-v3-vs-v2-perf-comparison

Note, the node-http2 requestHandler should not be used with S3, I've excluded it from the sample.

Disclaimer: due to the greater number of features in the JSv3 SDK and the fact that the v2 SDK sacrifices a lot of correctness (e.g. endpoint resolution) for simplicity, the v2 SDK will still be faster than the v3 SDK in many cases. However, JSv2 has entered maintenance mode and development efforts are focused on this iteration of the AWS SDK for JavaScript.