def run_p2p()

in bench_cluster/communication/p2p.py [0:0]

• 37 lines of code
• 8 McCabe index (conditional complexity)

def run_p2p(local_rank, trials, warmups, maxsize, async_op, bw_unit, scan, raw, dtype, mem_factor, debug=False):
    # Prepare benchmark header
    print_header(bw_unit, raw, 'p2p')
    global_rank = dist.get_rank()
    world_size = dist.get_world_size()

    start_event = torch.cuda.Event(enable_timing=True)
    end_event = torch.cuda.Event(enable_timing=True)

    if scan:
        # Create list of message sizes
        M_LIST = [2**p for p in range(1, maxsize)]

        sync_all()
        # loop over various tensor sizes
        for M in M_LIST:
            try:
                mat = torch.ones(M, dtype=getattr(torch, dtype)).cuda(local_rank)
                sync_all()
                input = mat.mul_(float(global_rank))
                del mat
                torch.cuda.empty_cache()
            except RuntimeError as e:
                if 'out of memory' in str(e):
                    print_rank_0('WARNING: Ran out of GPU memory. Exiting comm op.')
                    sync_all()
                    break
                else:
                    raise e
            sync_all()
            timed_p2p(input, start_event, end_event, warmups, trials, async_op, bw_unit, raw)
    else:
        # Send the biggest message size our GPUs can fit. If you're facing OOM errors, reduce the mem_factor
        # Don't need output tensor, so double mem_factor
        elements_per_gpu = max_numel('p2p', getattr(torch, dtype), mem_factor * 2, local_rank)
        try:
            mat = torch.ones(elements_per_gpu, dtype=getattr(torch, dtype)).cuda(local_rank)
            input = mat.mul_(float(global_rank))
        except RuntimeError as e:
            if 'out of memory' in str(e):
                print_rank_0('WARNING: Ran out of GPU memory. Try to reduce the --mem-factor argument!')
                sync_all()
                return
        sync_all()
        timed_p2p(input, start_event, end_event, warmups, trials, async_op, bw_unit, raw)