Need to address significant design flaws leading to poor performance.

[russfellows]

Describe the feature

This library needs to implement at a minimum three techniques to improve performance:

1. Use of HTTP/2
2. io_uring
3. RDMA and zero-copy buffers

Use Case

Any use case that requires high performance. The current implementation achieves at most 50% of what other S3 libraries can achieve using the identical hardware. This indicates significant performance issues with this software stack.

Proposed Solution

AWS C S3 Library Enhancement Analysis: HTTP/2, io_uring, and RDMA Support

Current HTTP Implementation

The library currently uses HTTP/1.1 through the aws-c-http dependency. The code shows:

• HTTP connections are managed through aws_http_connection_manager
• The library uses traditional socket-based networking with TCP keepalive options
• Connection pooling and multiplexing is handled at the application layer
• The mock server tests confirm HTTP/1.1 usage (using Python's h11 library)

Implementation Strategy for Your Requirements

1. HTTP/2 Support

To enable HTTP/2, you'll need to:

Modify the HTTP layer dependency:

• Update aws-c-http to support HTTP/2 or integrate a different HTTP/2 library
• The connection management code in source/s3_client.c would need updates to handle HTTP/2 multiplexing
• Modify the request creation logic in the request message builders to use HTTP/2 headers

Key files to modify:

• source/s3_client.c (connection management)
• include/aws/s3/private/s3_client_impl.h (connection configuration)
• Request message creation functions

2. io_uring Integration

For io_uring support to improve performance:

I/O layer modifications:

• Replace the current async I/O implementation with io_uring-based operations
• The library currently uses aws-c-io for I/O operations - this would need io_uring backend support
• Modify the buffer management and memory pooling system shown in docs/memory_aware_request_execution.md

Implementation areas:

• File I/O operations for uploads/downloads
• Network I/O for HTTP requests
• Buffer management for zero-copy operations

3. RDMA and Zero-Copy Support

For RDMA integration to eliminate memory copies:

Memory management overhaul:

• The current buffer pooling system in the library would need modification to support RDMA-registered memory
• Implement RDMA-aware buffer allocation that can be directly accessed by the network hardware
• Modify the part-based upload/download mechanism to use RDMA transfers

Key implementation points:

• Replace the current memory pool (source/s3_client.c shows 8MB default part sizes)
• Implement RDMA connection management alongside HTTP connections
• Create zero-copy paths for data transfer between user space and S3

Implementation Approach

1. Start with HTTP/2: This is the most straightforward enhancement and will provide immediate performance benefits through connection multiplexing.

2. Add io_uring support: This can be done incrementally, starting with file I/O operations and then extending to network operations.

3. Implement RDMA: This is the most complex change and would require significant architectural modifications to support zero-copy operations.

Current Library Architecture Analysis

The library's current architecture with its connection pooling, part-based transfers, and async I/O foundation provides a good starting point for these enhancements. Key characteristics:

• Default part size: 8MB (configurable)
• Connection management: Uses connection pooling with automatic scaling
• Memory management: Custom buffer pooling system for efficient memory usage
• Async operations: Thread pool-based async I/O system
• Request splitting: Automatic parallel chunk processing

Technical Implementation Details

HTTP/2 Migration Path

// Current HTTP/1.1 connection setup (simplified)
struct aws_http_connection_manager_options http_options = {
    .initial_window_size = initial_window_size,
    .socket_options = &socket_options,
    .tls_options = tls_options,
    // Add HTTP/2 specific options
    .http2_settings = &http2_settings,
    .protocol_version = AWS_HTTP_VERSION_2
};

io_uring Integration Points

1. File I/O operations in upload/download paths
2. Network socket operations for HTTP requests
3. Buffer management for zero-copy file-to-network transfers

RDMA Implementation Strategy

1. RDMA connection management parallel to HTTP connections
2. Memory registration for S3 transfer buffers
3. Protocol negotiation to determine RDMA capability
4. Fallback mechanisms for non-RDMA endpoints

Benefits Expected

• HTTP/2: Reduced connection overhead, better multiplexing, header compression
• io_uring: Lower CPU usage, reduced system call overhead, better I/O efficiency
• RDMA: Zero-copy transfers, reduced memory bandwidth usage, lower latency

Dependencies and Prerequisites

• HTTP/2: Update or replace aws-c-http dependency
• io_uring: Linux kernel 5.1+ support, liburing integration
• RDMA: InfiniBand or RoCE hardware, libibverbs integration

The modular design with separate HTTP, I/O, and client layers makes it possible to implement these features incrementally while maintaining backward compatibility.

Other Information

No response

Acknowledgements

• I may be able to implement this feature request
• This feature might incur a breaking change

[DmitriyMusatkin]

AWS S3 does not support http/2 on server side, so there is no current plans to add support to this library.
We will consider consider your other 2 points as feature requests.

[russfellows]

Dmitry,

Two important items:

1. Either S3 is a standard, widely used protocol, or it isn't. If it IS, then http/2 is important, regardless if AWS' particular implementation of S3 happens to support it currently, or not. Other S3 implementations do, and THEREFORE, having http/2 support in the client stack is VERY important.

2. There is MUCH more information available about the issues with the AWS http stack documented in the issue id 1349 in the AWS Smithy HTTP client stack repository here: Issues with AWS HTTP stack