README.md

usage:

1. modify run_comm_bench.sh and modify parameters as you need, especially bind_core_list to fit your machine configuration
2. run with the following args: run_comm_bench.sh [args]

• --elements: number of elements, default 16384
• --dtype: data type (bf16 or fp32), default bf16
• --count: number of iterations, default 10000
• --ccl: if present use oneccl allreduce, otherwise call deepspeed inference_all_reduce which intend to optimize for inference

3. Alternative way of benchmarking is run_mod_sweep.sh which generates report that is more oneCCL flavor

validate correctness:

• ./validate.sh
• ./validate.sh --ccl -- validate ccl implementation
• ./validate.sh --ipex -- validate ipex implementation

performance test:

• ./run_mod_sweep.sh -- test inference_all_reduce performance, which is deepspeed built-in implmentation
• ./run_mod_sweep.sh --ipex -- test ipex performance For the following two, you need to have oneCCL binding for PyTorch and oneCCL installed in your environment
• ./run_mod_sweep.sh --ccl -- test ccl performance
• ./run_mod_sweep.sh --torch -- test torch.distributed performance, when oneCCL binding for PyTorch is installed, this will test allreduce performance through oneCCL binding for PyTorch.