Skip to main content
Version: v2509

observability.nccl_benchmark_torch.run

byte_value​

def byte_value(byte: str) -> int

converts data to bytes if in MB, or GB format

factor_range​

def factor_range(minbytes: str, maxbytes: str, stepfactor: float) -> List[int]

creates the range for the values for which nccl bw needs to be tested

run_allgather​

def run_allgather(group: torch.distributed.ProcessGroup, bytes: int,
                  buffers: torch.Tensor, world_size: int,
                  dtype: torch.dtype) -> Tuple[float, float, float]

group: distributed group bytes: total data that needs to be transferred buffers: torch tensor that can fit the maximum num of elements specified in -e argument

returns: (inplace_time, inplace_algbw, inplace_busbw)

run_allredeuce​

def run_allredeuce(group: torch.distributed.ProcessGroup, bytes: int,
                   buffers: torch.Tensor, world_size: int,
                   dtype: torch.dtype) -> Tuple[float, float, float]

group: distributed group bytes: total data that needs to be transferred buffers: torch tensor that can fit the maximum num of elements specified in -e argument

returns: (inplace_time, inplace_algbw, inplace_busbw)

calculate_bw​

def calculate_bw(time: float, bytes: int, world_size: int,
                 routine: str) -> Tuple[float, float, float]

takes teh execution time and calculates algo bw and busbe returns (algbw, busbw)

get_routine​

def get_routine(routine: str) -> callable

all routines will have same interface for calling returns which routine to use for the test