Dynamically connected (DC) queue pairs

In RDMA, the scalability issue has been puzzled people for a long time. The resource contention in both host and NIC sides forms the bottleneck when the number of QPs exceeds some threshold. Therefore, some SW-based multiplex approaches, including DC queue pairs [^1] have been proposed to cope with these issues. DC transport can use one single QP to manage multiple outstanding connections, and the WQEs corresponding to different connections are marked with stream_id to identify their affiliation and handling order (somewhat similar with CUDA stream).

BlueFlame (MLX5)

BlueFlame is the codename by Mellanox standing for programmed I/O inlining WQE with the doorbell to the NIC. With this, we can sidestep the DMA-read from NIC for the WQE.