hi， I want to know the performance benchmark about gloo with nccl

[mullerhai]

I want to know how ‌Gloo‌ compares to ‌NCCL‌ in terms of performance for distributed training on ‌NVIDIA GPUs (CUDA)‌, as we currently plan to use ‌Gloo‌ for distributed inference in ‌Java‌ with ‌javacpp-pytorch‌.

Key Points:
‌Gloo vs. NCCL‌:

‌NCCL‌ is optimized for NVIDIA GPUs (CUDA), offering ‌higher throughput and lower latency‌ in GPU clusters.
‌Gloo‌ is a ‌cross-platform‌ alternative but may underperform NCCL in CUDA environments unless specific optimizations are applied.
‌Java Integration with PyTorch (javacpp-pytorch)‌:

‌Gloo‌ is supported in PyTorch, but ‌NCCL‌ is the default for GPU training.
If using ‌Gloo in Java‌, ensure compatibility with CUDA and distributed inference workflows.
Would you like a deeper comparison (e.g., benchmarks, setup guidance) for your specific use case?
thanks

[d4l3k]

Performance is generally significantly lower than NCCL. The new IB+cuda backend is significantly faster than the TCP backend but I expect both will be lower than NCCL

It depends a lot on what your network is. Are you using TCP, IB, etc?

I'm not aware of any public benchmarks comparing Gloo vs NCCL but it's fairly simple in PyTorch to benchmark them yourself

[mullerhai]

Performance is generally significantly lower than NCCL. The new IB+cuda backend is significantly faster than the TCP backend but I expect both will be lower than NCCL

It depends a lot on what your network is. Are you using TCP, IB, etc?

I'm not aware of any public benchmarks comparing Gloo vs NCCL but it's fairly simple in PyTorch to benchmark them yourself

something I WANT to.comfirm that if gloo support send and recv for cuda gpu？the sourcecode couldfind SendWork and RecvWork