PyTorch 分布式操作之 Barrier

关于 barrier 的概念

关于 barrier 这个概念可以参考 Wiki 中的介绍：同步屏障(Barrier)是并行计算中的一种同步方法。对于一群进程或线程，程序中的一个同步屏障意味着任何线程/进程执行到此后必须等待，直到所有线程/进程都到达此点才可继续执行下文。

这里要注意，barrier 这一方法并不是 pytorch 独有的，这是并行计算中的一个基本概念，其他的并行计算的场景下也可能会涉及这一概念和操作。本文主要讨论 pytorch 中的情况。

  1. torch.distributed.barrier(group=None, async_op=False, device_ids=None)
  3. Synchronizes all processes.
  5. This collective blocks processes until the whole group enters this function, if async_op is False, or if async work handle is called on wait().
  7. Parameters
  8. group (ProcessGroup, optional) – The process group to work on. If None, the default process group will be used.
  9. async_op (bool, optional) – Whether this op should be an async op
  10. device_ids ([int], optional) – List of device/GPU ids. Valid only for NCCL backend.
  12. Returns
  13. Async work handle, if async_op is set to True. None, if not async_op or if not part of the group

在多卡训练的时候，由于不同的 GPU 往往被设定在不同的进程中，有时候为了在单独的进程中执行一些任务，但是又同时希望限制其他进程的执行进度，就有了使用barrier的需求。
一个实际的场景是准备数据集：我们只需要在 0 号进程处理，其他进程没必要也执行这一任务，但是其他进程的后续工作却依赖准备好的数据。于是就需要在 0 号进程执行过程中阻塞其他的进程，使其进入等待状态。等到处理好之后，再一起放行。

这种需求下，一个典型的基于上下文管理器形式的构造如下：

  1. # https://github.com/ultralytics/yolov5/blob/7d56d451241e94cd9dbe4fcb9bfba0e92c6e0e23/utils/torch_utils.py#L29-L38
  3. @contextmanager
  4. def torch_distributed_zero_first(local_rank: int):
  5.     """
  6.     Decorator to make all processes in distributed training
  7.     wait for each local_master to do something.
  8.     """
  9.     if local_rank not in [-1, 0]:
  10.         dist.barrier(device_ids=[local_rank])
  11.     yield
  12.     if local_rank == 0:
  13.         dist.barrier(device_ids=[0])

关于 barrier 的细节

  1. # -*- coding: utf-8 -*-
  2. # @Time    : 2021/1/6
  3. # @Author  : Lart Pang
  4. # @GitHub  : https://github.com/lartpang
  6. import os
  7. import time
  9. import torch.distributed as dist
  10. import torch.multiprocessing as mp
  13. def ddp_test_v0(local_rank, word_size):
  14.     # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
  15.     dist.init_process_group(backend="nccl", world_size=word_size, rank=local_rank)
  17.     print("first before barrier{}\n".format(local_rank))
  18.     if local_rank != 0:
  19.         dist.barrier()
  20.     print("first after barrier{}\n".format(local_rank))
  22.     print("inter {}".format(local_rank))
  24.     print("second before barrier{}\n".format(local_rank))
  25.     if local_rank == 0:
  26.         dist.barrier()
  27.     print("second after barrier{}\n".format(local_rank))
  29.     print("{} exit".format(local_rank))
  32. def ddp_test_v1(local_rank, word_size):
  33.     # Initializes the distributed backend which will take care of synchronizing nodes/GPUs
  34.     dist.init_process_group(backend="nccl", world_size=word_size, rank=local_rank)
  36.     if local_rank != 0:
  37.         print("1 before barrier{}\n".format(local_rank))
  38.         start = time.time()
  39.         time.sleep(5)
  40.         dist.barrier()
  41.         print(time.time() - start)
  42.         print("1 after barrier{}\n".format(local_rank))
  43.         dist.barrier()
  44.         print("1 after barrier{}\n".format(local_rank))
  45.     else:
  46.         print("0 before barrier{}\n".format(local_rank))
  47.         start = time.time()
  48.         dist.barrier()
  49.         print(time.time() - start)
  50.         print("0 after barrier{}\n".format(local_rank))
  51.         print("0 after barrier{}\n".format(local_rank))
  52.         dist.barrier()
  53.         print("0 after barrier{}\n".format(local_rank))
  55.     print("{} exit".format(local_rank))
  58. def main():
  59.     world_size = 2
  60.     os.environ["MASTER_ADDR"] = "127.0.0.1"
  61.     os.environ["MASTER_PORT"] = "29500"
  62.     mp.spawn(ddp_test_v0, args=(world_size,), nprocs=world_size, join=True)
  65. if __name__ == "__main__":
  66.     main()

这里展示了两个例子，实际上在官方展示的 dist.barrier 之外显示了该方法的一个重要特性，就是其操作实际上是每一个进程内部都需要对应的执行同样的次数，才会对应的由阻塞变为正常运行。
先看第一个例子：

def ddp_test(local_rank, word_size):
    # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
    dist.init_process_group(backend="nccl", world_size=word_size, rank=local_rank)

    print("first before barrier{}\n".format(local_rank))
    if local_rank != 0:
        dist.barrier()
    print("first after barrier{}\n".format(local_rank))

    print("inter {}".format(local_rank))

    print("second before barrier{}\n".format(local_rank))
    if local_rank == 0:
        dist.barrier()
    print("second after barrier{}\n".format(local_rank))

    print("{} exit".format(local_rank))

其输出是：

first before barrier1
first before barrier0


first after barrier0

inter 0
second before barrier0

second after barrier0

0 exit
first after barrier1

inter 1
second before barrier1

second after barrier1

1 exit

Process finished with exit code 0

可以看到，有几个细节：

  • barrier 之前，所有的操作都是各 GPU 进程自己输出自己的。
    • 由于 local_rank=0 执行到自己可见的 barrier 中间会输出多个，而 local_rank=1 则只有一条 first before barrier1 。
  • second before barrier0 之后，0 号执行到了属于自己的 barrier ，这回让使得其他进程不再阻塞，开始正常运行。由于中间操作的时间，所以先是 0 号输出自己的 second after barrier0 并随之退出，之后 1 号也接着开始输出自己的结果。

这里有一点值得注意，不同进程的 barrier 实际上是互相对应的，必须所有进程都执行一次barrier，才会重新放行正常前进。
对于第二段代码：

def ddp_test_v1(local_rank, word_size):
    # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
    dist.init_process_group(backend="nccl", world_size=word_size, rank=local_rank)

    if local_rank != 0:
        print("1 before barrier{}\n".format(local_rank))
        start = time.time()
        time.sleep(5)
        dist.barrier()
        print(time.time() - start)
        print("1 after barrier{}\n".format(local_rank))
        dist.barrier()
        print("1 after barrier{}\n".format(local_rank))
    else:
        print("0 before barrier{}\n".format(local_rank))
        start = time.time()
        dist.barrier()
        print(time.time() - start)
        print("0 after barrier{}\n".format(local_rank))
        print("0 after barrier{}\n".format(local_rank))
        dist.barrier()
        print("0 after barrier{}\n".format(local_rank))

    print("{} exit".format(local_rank))

则是有输出：

1 before barrier1
0 before barrier0


5.002117395401001
5.0021262168884281 after barrier1


0 after barrier0

0 after barrier0

0 after barrier0

0 exit
1 after barrier1

1 exit

Process finished with exit code 0

可以看到一个重要的点，就是这两处 print(time.time() - start) 的输出是基本一样的，不管前面延时多少， barrier 后面的时间都是按照最长到达并执行 barrier 的间隔时间来的。这个更体现了不同进程 barrier 之间的互相限制关系。而 0 到达自己的第二个 barrier 之后，会使得 1 号再次运行。但是此时 0 是先结束的。
另外，可以验证，如果某个编号对应的代码中的两个 barrier 之中的一个，那么另一个就会陷入无限等待之中。
例如：


def ddp_test_v1(local_rank, word_size):
    # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
    dist.init_process_group(backend="nccl", world_size=word_size, rank=local_rank)

    if local_rank != 0:
        print("1 before barrier{}\n".format(local_rank))
        start = time.time()
        time.sleep(5)
        dist.barrier()
        print(time.time() - start)
        print("1 after barrier{}\n".format(local_rank))
        # dist.barrier()
        print("1 after barrier{}\n".format(local_rank))
    else:
        print("0 before barrier{}\n".format(local_rank))
        start = time.time()
        time.sleep(3)
        dist.barrier()
        print(time.time() - start)
        print("0 after barrier{}\n".format(local_rank))
        print("0 after barrier{}\n".format(local_rank))
        dist.barrier()
        print("0 after barrier{}\n".format(local_rank))

    print("{} exit".format(local_rank))

输出：

0 before barrier0
1 before barrier1


5.002458572387695
1 after barrier1

1 after barrier1

1 exit
5.002473831176758
0 after barrier0

0 after barrier0

Traceback (most recent call last):
  File "/home/lart/Coding/SODBetterProj/tools/dist_experiment_test.py", line 67, in <module>
    main()
  File "/home/lart/Coding/SODBetterProj/tools/dist_experiment_test.py", line 63, in main
    mp.spawn(ddp_test_v1, args=(world_size,), nprocs=world_size, join=True)
  File "/home/lart/miniconda3/envs/pt17/lib/python3.8/site-packages/torch/multiprocessing/spawn.py", line 199, in spawn
    return start_processes(fn, args, nprocs, join, daemon, start_method='spawn')
  File "/home/lart/miniconda3/envs/pt17/lib/python3.8/site-packages/torch/multiprocessing/spawn.py", line 157, in start_processes
    while not context.join():
  File "/home/lart/miniconda3/envs/pt17/lib/python3.8/site-packages/torch/multiprocessing/spawn.py", line 75, in join
    ready = multiprocessing.connection.wait(
  File "/home/lart/miniconda3/envs/pt17/lib/python3.8/multiprocessing/connection.py", line 931, in wait
    ready = selector.select(timeout)
  File "/home/lart/miniconda3/envs/pt17/lib/python3.8/selectors.py", line 415, in select
    fd_event_list = self._selector.poll(timeout)
KeyboardInterrupt

Process finished with exit code 137 (interrupted by signal 9: SIGKILL)

会在第二个 barrier 处无限等待下去。
这一特点在这个回答中也被提到了：

when a process encounters a barrier it will block the position of the barrier is not important (not all processes have to enter the same if-statement, for instance) a process is blocked by a barrier until all processes have encountered a barrier, upon which the barrier is lifted for all processes

https://stackoverflow.com/a/59766443

重要的参考资料