GUSLI Backend Integration for NIXL

[aranadive]

What?

Added GUSLI (G3+ User Space Access Library) backend support to NIXL and NIXLBench. This integration leverages GUSLI's user-space block device access to provide storage performance measurements through the NIXL framework.

Why?

• User-Space Block Access: GUSLI provides direct, user-space access to block storage devices, bypassing kernel overhead for low-latency storage operations
• G3+ Storage Integration: Enables benchmarking of next-generation storage architectures that benefit from user-space I/O control
• Storage Performance Measurement: Extends NIXLBench from network/memory benchmarks to include precise storage I/O performance analysis

How?

GUSLI Architecture

• Block Device Management: GUSLI manages devices through configuration:
  • Local Files: File-backed storage (./store0.bin)
  • Block Devices: Direct device access (/dev/nvme0)
• Asynchronous I/O: Implements pollable async operations with configurable concurrency
• LBA Addressing: Uses Logical Block Address offsets for precise block-level positioning

GUSLI I/O Model

• Buffer Registration: Explicit memory buffer registration before I/O operations
• Request Lifecycle: Prepare → Submit → Poll → Complete with comprehensive error handling
• Multi-Range Operations: Supports scatter-gather I/O across multiple memory regions

NIXLBench Integration

• Configuration: Mirrors reference test setup with client context and device mapping
• Transfer Patterns: Supports DRAM↔Block bidirectional transfers using GUSLI's optimized paths
• Performance Features: Direct I/O, batch processing, and priority scheduling for accurate storage benchmarking

# Usage example
./nixlbench --backend=GUSLI --initiator_seg_type=DRAM --target_seg_type=BLK

The integration maintains GUSLI's user-space performance advantages while providing NIXLBench's standardized benchmarking interface for G3+ storage subsystem analysis.

Original from PR #494. Thanks to @danielhe-nvidia.

[github-actions]

👋 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]

/build

[ovidiusm]

/build

[ovidiusm]

/build

[ovidiusm]

LGTM