Experimental: Overlapping communication and computation

Compiled Graph currently provides experimental support for GPU communication and computation overlap. When you turn this feature on, it automatically overlaps the GPU communication with computation operations, thereby hiding the communication overhead and improving performance.

To enable this feature, specify _overlap_gpu_communication=True when calling dag.experimental_compile().

The following code has GPU communication and computation operations that benefit from overlapping.

import ray
import time
import torch
from ray.dag import InputNode, MultiOutputNode


@ray.remote(num_cpus=0, num_gpus=1)
class TorchTensorWorker:
    def send(self, shape, dtype, value: int, send_tensor=True):
        if not send_tensor:
            return 1
        return torch.ones(shape, dtype=dtype, device="cuda") * value

    def recv_and_matmul(self, two_d_tensor):
        """
        Receive the tensor and do some expensive computation (matmul).

        Args:
            two_d_tensor: a 2D tensor that has the same size for its dimensions
        """
        # Check that tensor got loaded to the correct device.
        assert two_d_tensor.dim() == 2
        assert two_d_tensor.size(0) == two_d_tensor.size(1)
        torch.matmul(two_d_tensor, two_d_tensor)
        return (two_d_tensor[0][0].item(), two_d_tensor.shape, two_d_tensor.dtype)


def test(overlap_gpu_communication):
    num_senders = 3
    senders = [TorchTensorWorker.remote() for _ in range(num_senders)]
    receiver = TorchTensorWorker.remote()

    shape = (10000, 10000)
    dtype = torch.float16

    with InputNode() as inp:
        branches = [sender.send.bind(shape, dtype, inp) for sender in senders]
        branches = [
            branch.with_tensor_transport(
                transport="nccl", _static_shape=True, _direct_return=True
            )
            # For a ray version before 2.42, use `with_type_hint()` instead.
            # branch.with_type_hint(
            #     TorchTensorType(
            #         transport="nccl", _static_shape=True, _direct_return=True
            #     )
            # )
            for branch in branches
        ]
        branches = [receiver.recv_and_matmul.bind(branch) for branch in branches]
        dag = MultiOutputNode(branches)

    compiled_dag = dag.experimental_compile(
        _overlap_gpu_communication=overlap_gpu_communication
    )

    start = time.monotonic()
    for i in range(5):
        ref = compiled_dag.execute(i)
        result = ray.get(ref)
        assert result == [(i, shape, dtype)] * num_senders
    duration = time.monotonic() - start
    print(f"{overlap_gpu_communication=}, {duration=}")
    compiled_dag.teardown(kill_actors=True)


for overlap_gpu_communication in [False, True]:
    test(overlap_gpu_communication)

The output of the preceding code includes the following two lines:

overlap_gpu_communication=False, duration=1.0670117866247892
overlap_gpu_communication=True, duration=0.9211348341777921

The actual performance numbers may vary on different hardware, but enabling _overlap_gpu_communication improves latency by about 14% for this example.