v0.2.2

v0.2.2 - Add query and check hmem info

• Affinity: Added host memory (hmem) info for better GPU/system topology view.
• Docs: Updated documentation for experiments to improve reproducibility and clarity.

v0.2.1

v0.2.1 - Proxy hang Bug Fixes

Bug Fixes

• Fixed hang where idle ranks waited indefinitely during pairwise proxy tests
• Fixed race condition in SymmetricMemory::Complete() - now uses cuda::std::atomic with relaxed memory order for proper CPU-GPU synchronization on managed memory

v0.2.0

v0.2.0 - GPU-Initiated RDMA Proxy

New Features

• Proxy Experiment (experiments/proxy): Benchmarks GPU-initiated RDMA writes where:

  • GPU kernel pushes requests to a device-accessible queue
  • CPU proxy loop processes requests and issues RDMA operations
  • Supports both single-channel and multi-channel modes for performance comparison

• Device Request Queue: Lock-free MPSC queue for GPU-to-CPU communication with proper memory fencing (__threadfence_system)

• Completion Tracking: Quiet() and Fence() device functions for GPU-side synchronization with CPU-initiated RDMA completions