Skip to content

CHANGELOG.md

Latest commit

 

History

History
106 lines (84 loc) · 3.34 KB

CHANGELOG.md

File metadata and controls

106 lines (84 loc) · 3.34 KB

CHANGELOG

All notable changes to MultiGpuHelper will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[1.0.0] – 2024-01-07 – Initial Release

Added

  • Core GPU Management

    • GpuManager for device registry and selection
    • GpuDevice model with VRAM tracking
    • Support for multiple GPU devices

  • GPU Auto-Detection

    • NvidiaSmiProbeProvider for automatic NVIDIA GPU detection
    • Graceful fallback if nvidia-smi is unavailable
    • IGpuProbeProvider abstraction for extensibility

  • Device Selection Policies

    • GpuPolicy.RoundRobin – distribute work evenly
    • GpuPolicy.MostFreeVram – select GPU with most available memory
    • GpuPolicy.SpecificDevice – route to explicit GPU ID

  • VRAM Soft-Budgeting

    • VramBudget class with thread-safe reservations
    • Per-device VRAM limits
    • TryReserve() / Release() API
    • Automatic budget enforcement

  • Concurrency Control

    • Per-GPU semaphores via SemaphoreSlim
    • MaxConcurrentJobs configuration per device
    • Thread-safe work dispatching

  • Work Dispatching

    • GpuDispatcher for async work scheduling
    • Support for async/sync lambdas with/without return values
    • Timeout and cancellation token support
    • GpuWorkItem for work metadata (tags, VRAM requests, priorities)

  • Logging

    • IGpuLogger abstraction for custom logging
    • NoOpLogger default implementation

  • Utilities

    • Size helper for human-readable byte formatting (MiB, GiB)
    • XML documentation for all public APIs

  • Error Handling

    • GpuSelectionException – no suitable GPU found
    • GpuProbeException – GPU detection failed
    • GpuBudgetExceededException – VRAM budget exceeded
    • Rich context in exception messages

  • NuGet Packaging

    • Auto-generated .nupkg and .snupkg (symbol package)
    • Strong-name signing with 2048-bit RSA key
    • MIT license metadata
    • Package tags: gpu;cuda;ai;inference;multi-gpu;scheduler

  • CI/CD

    • GitHub Actions workflow (Windows, .NET 8.x)
    • Automated build, test, and pack
    • Artifact upload to GitHub

  • Documentation

    • Comprehensive README with quick-start examples
    • VERSIONING.md with SemVer policy and release workflow
    • MIT LICENSE

  • Sample Applications

    • .NET 8 console sample with async dispatching
    • .NET Framework 4.7.2 sample (device selection)
    • Hardware verification test for real GPU devices

  • Testing

    • 13 unit tests (xUnit)
    • VramBudget functionality tests
    • GpuManager selection policy tests
    • Hardware test for P4000 GPUs

Tested On

  • Hardware: NVIDIA Quadro P4000 (2 units)
  • Frameworks: .NET 8.0, .NET Standard 2.0, .NET Framework 4.7.2
  • OS: Windows 10/11

Known Limitations

  • NVIDIA GPU detection only (extensible to AMD ROCm, Intel oneAPI)
  • Requires nvidia-smi for auto-detection (gracefully handles missing driver)
  • Net472 sample requires .NET Framework SDK (code compiles, may need runtime)

[Unreleased]

Planned (Future)

  • AMD ROCm probe provider
  • Intel oneAPI probe provider
  • OpenCL support
  • GPU memory profiling hooks
  • Work queue persistence
  • Multi-machine GPU clustering
  • Performance optimizations
  • Linux/macOS support (after AMD/Intel providers)