Direct Device-to-Device Communication with TensorPipe CUDA RPC


Direct device-to-device RPC (CUDA RPC) is introduced in PyTorch 1.8 as a prototype feature. This API is subject to change.

In this recipe, you will learn:

  • The high-level idea of CUDA RPC.

  • How to use CUDA RPC.

What is CUDA RPC?

CUDA RPC supports directly sending Tensors from local CUDA memory to remote CUDA memory. Prior to v1.8 release, PyTorch RPC only accepts CPU Tensors. As a result, when an application needs to send a CUDA Tensor through RPC, it has to first move the Tensor to CPU on the caller, send it via RPC, and then move it to the destination device on the callee, which incurs both unnecessary synchronizations and D2H and H2D copies. Since v1.8, RPC allows users to configure a per-process global device map using the set_device_map API, specifying how to map local devices to remote devices. More specifically, if worker0’s device map has an entry "worker1" : {"cuda:0" : "cuda:1"}, all RPC arguments on "cuda:0" from worker0 will be directly sent to "cuda:1" on worker1. The response of an RPC will use the inverse of the caller device map, i.e., if worker1 returns a Tensor on "cuda:1", it will be directly sent to "cuda:0" on worker0. All intended device-to-device direct communication must be specified in the per-process device map. Otherwise, only CPU tensors are allowed.

Under the hood, PyTorch RPC relies on TensorPipe as the communication backend. PyTorch RPC extracts all Tensors from each request or response into a list and packs everything else into a binary payload. Then, TensorPipe will automatically choose a communication channel for each Tensor based on Tensor device type and channel availability on both the caller and the callee. Existing TensorPipe channels cover NVLink, InfiniBand, SHM, CMA, TCP, etc.

How to use CUDA RPC?

The code below shows how to use CUDA RPC. The model contains two linear layers and is split into two shards. The two shards are placed on worker0 and worker1 respectively, and worker0 serves as the master that drives the forward and backward passes. Note that we intentionally skipped DistributedOptimizer to highlight the performance improvements when using CUDA RPC. The experiment repeats the forward and backward passes 10 times and measures the total execution time. It compares using CUDA RPC against manually staging to CPU memory and using CPU RPC.

import torch
import torch.distributed.autograd as autograd
import torch.distributed.rpc as rpc
import torch.multiprocessing as mp
import torch.nn as nn

import os
import time

class MyModule(nn.Module):
    def __init__(self, device, comm_mode):
        self.device = device
        self.linear = nn.Linear(1000, 1000).to(device)
        self.comm_mode = comm_mode

    def forward(self, x):
        # is a no-op if x is already on self.device
        y = self.linear(
        return y.cpu() if self.comm_mode == "cpu" else y

    def parameter_rrefs(self):
        return [rpc.RRef(p) for p in self.parameters()]

def measure(comm_mode):
    # local module on "worker0/cuda:0"
    lm = MyModule("cuda:0", comm_mode)
    # remote module on "worker1/cuda:1"
    rm = rpc.remote("worker1", MyModule, args=("cuda:1", comm_mode))
    # prepare random inputs
    x = torch.randn(1000, 1000).cuda(0)

    tik = time.time()
    for _ in range(10):
        with autograd.context() as ctx:
            y = rm.rpc_sync().forward(lm(x))
            autograd.backward(ctx, [y.sum()])
    # synchronize on "cuda:0" to make sure that all pending CUDA ops are
    # included in the measurements
    tok = time.time()
    print(f"{comm_mode} RPC total execution time: {tok - tik}")

def run_worker(rank):
    os.environ['MASTER_ADDR'] = 'localhost'
    os.environ['MASTER_PORT'] = '29500'
    options = rpc.TensorPipeRpcBackendOptions(num_worker_threads=128)

    if rank == 0:
        options.set_device_map("worker1", {0: 1})

    # block until all rpcs finish

if __name__=="__main__":
    world_size = 2
    mp.spawn(run_worker, nprocs=world_size, join=True)

Outputs are displayed below, which shows that CUDA RPC can help to achieve 34X speed up compared to CPU RPC in this experiment.

cpu RPC total execution time: 2.3145179748535156 Seconds
cuda RPC total execution time: 0.06867480278015137 Seconds


Access comprehensive developer documentation for PyTorch

View Docs


Get in-depth tutorials for beginners and advanced developers

View Tutorials


Find development resources and get your questions answered

View Resources