Hello,
I’m working on a Floyd BB02 board (Xavier NX 16GB) and added a 10G ethernet M2 daughtercard from Innovision (The World’s First M.2 10GbE LAN Module | InnoNews - Innodisk).
With iperf3, I can get close to the maximum throughput (>=9 Gbps in either directions), but when using some application or test code (python or C++), throughput seems to be much lower:
- Python seems to be limited to ~200MB/s (1.6Gbps)
- using C++ and transferring V2L image buffers (allocated with V4L2_MEMORY_MMAP), I too get low throughput ~200MB/s
- using C++ and CUDA-allocated memory (pinned), I get around 800MB/s (~6.5Gbps)
Has anybody been able to write application code running on Jetson that reaches close-to-max throughput on a 10G port?
Maybe using libraries like UCX (GitHub - openucx/ucx: Unified Communication X (mailing list - https://elist.ornl.gov/mailman/listinfo/ucx-group)) or UCX-py?
or Rivermax (GitHub - NVIDIA/Rivermax: Extended and advanced applications to the Rivermax SDK (Networking SDK for Media and Data Streaming).) if supported by Jetson?
Any feedback or suggestion is welcome!
Thanks,
Alain