Contents

• 1. Overview
  • 1.1. How GPUDirect RDMA Works
  • 1.2. Standard DMA Transfer
  • 1.3. GPUDirect RDMA Transfers
  • 1.4. Changes in CUDA 6.0
  • 1.5. Changes in CUDA 7.0
  • 1.6. Changes in CUDA 8.0
  • 1.7. Changes in CUDA 10.1
  • 1.8. Changes in CUDA 11.2
  • 1.9. Changes in CUDA 11.4
• 2. Design Considerations
  • 2.1. Lazy Unpinning Optimization
  • 2.2. Registration Cache
  • 2.3. Unpin Callback
  • 2.4. Supported Systems
  • 2.5. PCI BAR sizes
  • 2.6. Tokens Usage
  • 2.7. Synchronization and Memory Ordering
• 3. How to Perform Specific Tasks
  • 3.1. Displaying GPU BAR space
  • 3.2. Pinning GPU memory
  • 3.3. Unpinning GPU memory
  • 3.4. Handling the free callback
  • 3.5. Buffer ID Tag Check for A Registration Cache
  • 3.6. Linking a Kernel Module against nvidia.ko
  • 3.7. Using nvidia-peermem
• 4. References
  • 4.1. Basics of UVA CUDA Memory Management
  • 4.2. Userspace API
  • 4.3. Kernel API
  • 4.4. Porting to Tegra
    • 4.4.1. Changing the allocator
    • 4.4.2. Modification to Kernel API
    • 4.4.3. Other highlights
• 5. Notices
  • 5.1. Notice
  • 5.2. OpenCL
  • 5.3. Trademarks