Pytorch - 分布式通信原语（附源码）

前言

由于工作需要，最近在补充分布式训练方面的知识。经过一番理论学习后仍觉得意犹未尽，很多知识点无法准确get到（例如：分布式原语scatter、all reduce等代码层面应该是什么样的，ring all reduce 算法在梯度同步时是怎么使用的，parameter server参数是如何部分更新的）。

著名物理学家，诺贝尔奖得主Richard Feynman办公室的黑板上写了："What I cannot create, I do not understand."。在程序员界也经常有"show me the code"的口号。因此，我打算写一系列的分布式训练的文章，将以往抽象的分布式训练的概念以代码的形式展现出来，并保证每个代码可执行、可验证、可复现，并贡献出来源码让大家相互交流。

经过调研发现pytorch对于分布式训练做好很好的抽象且接口完善，因此本系列文章将以pytorch为主要框架进行，文章中的例子很多都来自pytorch的文档，并在此基础上进行了调试和扩充。

最后，由于分布式训练的理论介绍网络上已经很多了，理论部分的介绍不会是本系列文章的重点，我会将重点放在代码层面的介绍上面。

1 基本介绍

近些年随着深度学习的火爆，模型的参数规模也飞速增长，OpenAI数据显示：

2012年以前，模型计算耗时每2年增长一倍，和摩尔定律保持一致；
2012年后，模型计算耗时每3.4个月翻一倍，远超硬件发展速度；

近一年来，百亿、千亿级的参数模型陆续面世，谷歌、英伟达、阿里、智源研究院更是发布了万亿参数模型。因此，大模型已经成为了未来深度学习的趋势。提到大模型，就不得不提分布式训练，由于模型参数和训练数据的不断增多，只有通过分布式训练才能完成大模型的训练任务。

分布式训练可以分为数据并行、模型并行，流水线并行和混合并行。分布式算法又有典型的parameter server和ring all-reduce。无论是哪一种分布式技术一个核心的关键就是如何进行communication，这是实现分布式训练的基础，因此要想掌握分布式训练或当前流行的大模型训练务必对worker间的通信方式有所了解。

互联网上已经有很多关于分布式训练的通信方面的文章，但是均没有代码层面的例子。我是属于比较愚钝类型的，只有通过自己手动实现一下方能对一些抽象的概念有较深的理解。

Pytorch的分布式训练的通信是依赖torch.distributed模块来实现的，torch.distributed提供了point-2-point communication 和collective communication两种通信方式。

point-2-point communication提供了send和recv语义，用于任务间的通信
collective communication主要提供了scatter/broadcast/gather/reduce/all_reduce/all_gather 语义，不同的backend在提供的通信语义上具有一定的差异性。

Device	CPU	GPU	CPU	GPU	CPU	GPU
send	✓	✘	✓	?	✘	✘
recv	✓	✘	✓	?	✘	✘
broadcast	✓	✓	✓	?	✘	✓
all_reduce	✓	✓	✓	?	✘	✓
reduce	✓	✘	✓	?	✘	✓
all_gather	✓	✘	✓	?	✘	✓
gather	✓	✘	✓	?	✘	✘
scatter	✓	✘	✓	?	✘	✘
reduce_scatter	✘	✘	✘	✘	✘	✓
all_to_all	✘	✘	✓	?	✘	✓
barrier	✓	✘	✓	?	✘	✓

2 P2P communication

下面通过torch.distributed的send/recv接口实现一个简易的ping-pong 程序。程序功能如下：

tensor 初始值为0
process 0 （或叫rank 0)：对tensor加1，然后发送给process 1(或叫rank1）；
process 1：接收到tensor后，对tensor 加2，然后在发送给process 0;
process 0：接收process1发送的tensor；

2.1 初始化

pytorch中在分布式通信原语使用之前，需要对分布式模块进行初始化。pytorch的分布式模块通过torch.distributed.init_process_group来完成

通过环境变量MASTER_ADDR和MASTER_PORT设置rank0的IP和PORT信息，rank0的作用相当于是协调节点，需要其他所有节点知道其访问地址;
本例中后端选择的是gloo，通过设置NCCL_DEBUG环境变量为INFO，输出NCCL的调试信息；
init_process_group：执行网络通信模块的初始化工作
- backend：设置后端网络通信的实现库，可选的为gloo、nccl和mpi；本例选择gloo作为backend(注：nccl不支持p2p通信，mpi需要重新编译pytorch源码才能使用）；
- rank：为当前rank的index，用于标记当前是第几个rank，取值为0到work_size - 1之间的值；
- world_size: 有多少个进程参与到分布式训练中;

def init_process(rank_id, size, fn, backend='gloo'):""" Initialize the distributed environment. """os.environ['MASTER_ADDR'] = '127.0.0.1'os.environ['MASTER_PORT'] = '29500'dist.init_process_group(backend, rank=rank_id, world_size=size)fn(rank_id, size)

2.2 通信逻辑

下面的代码展示了rank0和rank1进行ping-pong通信的实现：

通过rank_id来区分当前应该执行哪一个rank的业务逻辑；
pytorch 中通过torch.distributed.send(tensor, dst, group=None, tag=0) 和torch.distributed.isend(tensor, dst, group=None, tag=0) 来实现tensor的发送，其中send是同步函数，isend是异步函数；
- tensor：要发送的数据
- dst：目标rank，填写目标rank id即可
pytorch中通过torch.distributed.recv(tensor, src=None, group=None, tag=0)和torch.distributed.irecv(tensor, src=None, group=None, tag=0)来实现tensor的接收，其中recv是同步函数，irecv是异步函数；
- tensor：接收的数据
- src：接收数据来源的rank id

def run(rank_id, size):tensor = torch.zeros(1)if rank_id == 0:tensor += 1# Send the tensor to process 1dist.send(tensor=tensor, dst=1)print('after send, Rank ', rank_id, ' has data ', tensor[0])dist.recv(tensor=tensor, src=1)print('after recv, Rank ', rank_id, ' has data ', tensor[0])else:# Receive tensor from process 0dist.recv(tensor=tensor, src=0)print('after recv, Rank ', rank_id, ' has data ', tensor[0])tensor += 1dist.send(tensor=tensor, dst=0)print('after send, Rank ', rank_id, ' has data ', tensor[0])

2.3 任务启动

通过下面的代码来启动两个process进行ping-pong通信：

这里使用torch.multiprocessing来启动多进程，torch.multiprocessing是python库中multiprocessing的封装，并且兼容了所有的接口
multiprocessing.set_start_method : 用于指定创建child process的方式，可选的值为fork、spawn和forkserver。使用spawn，child process仅会继承parent process的必要resource，file descriptor和handle均不会继承。
multiprocessing.Process(group=None, target=None, name=None, args=(), kwargs={}, *, daemon=None) ：用来启动child process

if __name__ == "__main__":size = 2processes = []mp.set_start_method("spawn")for rank in range(size):p = mp.Process(target=init_process, args=(rank, size, run))p.start()processes.append(p)for p in processes:p.join()

2.4 测试

完整代码如下：

import os
import torch
import torch.distributed as dist
import torch.multiprocessing as mpdef run(rank_id, size):tensor = torch.zeros(1)if rank_id == 0:tensor += 1# Send the tensor to process 1dist.send(tensor=tensor, dst=1)print('after send, Rank ', rank_id, ' has data ', tensor[0])dist.recv(tensor=tensor, src=1)print('after recv, Rank ', rank_id, ' has data ', tensor[0])else:# Receive tensor from process 0dist.recv(tensor=tensor, src=0)print('after recv, Rank ', rank_id, ' has data ', tensor[0])tensor += 1dist.send(tensor=tensor, dst=0)print('after send, Rank ', rank_id, ' has data ', tensor[0])def init_process(rank_id, size, fn, backend='gloo'):""" Initialize the distributed environment. """os.environ['MASTER_ADDR'] = '127.0.0.1'os.environ['MASTER_PORT'] = '29500'dist.init_process_group(backend, rank=rank_id, world_size=size)fn(rank_id, size)

执行效果如下：

root@g48r13:/workspace/communication# python sync_p2p.py
after send, Rank  0  has data  tensor(1.)
after recv Rank  1  has data  tensor(1.)
after send Rank  1  has data  tensor(2.)
after recv, Rank  0  has data  tensor(2.)

3 collective communication

3.1 broadcast

broadcast的计算方式如上图所示。

在pytorch中通过torch.distributed.broadcast(tensor, src, group=None, async_op=False) 来broadcast通信。

参数tensor在src rank是input tensor，在其他rank是output tensor；
参数src设置哪个rank进行broadcast，默认为rank 0；

使用方式如下面代码所示：

import os
import torch
import torch.distributed as dist
import torch.multiprocessing as mpdef run(rank_id, size):tensor = torch.arange(2, dtype=torch.int64) + 1 + 2 * rank_idprint('before broadcast',' Rank ', rank_id, ' has data ', tensor)dist.broadcast(tensor, src = 0)print('after broadcast',' Rank ', rank_id, ' has data ', tensor)def init_process(rank_id, size, fn, backend='gloo'):""" Initialize the distributed environment. """os.environ['MASTER_ADDR'] = '127.0.0.1'os.environ['MASTER_PORT'] = '29500'dist.init_process_group(backend, rank=rank_id, world_size=size)fn(rank_id, size)if __name__ == "__main__":size = 4processes = []mp.set_start_method("spawn")for rank in range(size):p = mp.Process(target=init_process, args=(rank, size, run))p.start()processes.append(p)for p in processes:p.join()

输出内容为：

一共有4个rank参与了broadcast计算，计算之前：rank0 为[1, 2]，rank1 为[3, 4]， rank2为[5, 6]， rank3为[7, 8]
broadcast计算之后，所有rank的结果均rank0的tensor即[1, 2]（因为在调用torch.distributed.broadcast时src设置为0，表示rank0进行broadcast）

before broadcast  Rank  1  has data  tensor([3, 4])
before broadcast  Rank  0  has data  tensor([1, 2])
before broadcast  Rank  2  has data  tensor([5, 6])
before broadcast  Rank  3  has data  tensor([7, 8])
after broadcast  Rank  1  has data  tensor([1, 2])
after broadcast  Rank  0  has data  tensor([1, 2])
after broadcast  Rank  2  has data  tensor([1, 2])
after broadcast  Rank  3  has data  tensor([1, 2])

3.2 scatter

scatter的计算方式如上图所示。

在pytorch中通过torch.distributed.scatter(tensor, scatter_list=None, src=0, group=None, async_op=False) 来实现scatter通信。

参数tensor为除 src rank外，其他rank获取output tensor的参数
scatter_list为进行scatter计算tensor list
参数src设置哪个rank进行scatter，默认为rank 0；

使用方式如下面代码所示：

这里需要注意的是，仅有src rank才能设置scatter_list( 本例中为rank 0），否则会报错

import os
import torch
import torch.distributed as dist
import torch.multiprocessing as mpdef run(rank_id, size):tensor = torch.arange(2, dtype=torch.int64) + 1 + 2 * rank_idprint('before scatter',' Rank ', rank_id, ' has data ', tensor)if rank_id == 0:scatter_list = [torch.tensor([0,0]), torch.tensor([1,1]), torch.tensor([2,2]), torch.tensor([3,3])]print('scater list:', scatter_list)dist.scatter(tensor, src = 0, scatter_list=scatter_list)else:dist.scatter(tensor, src = 0)print('after scatter',' Rank ', rank_id, ' has data ', tensor)def init_process(rank_id, size, fn, backend='gloo'):""" Initialize the distributed environment. """os.environ['MASTER_ADDR'] = '127.0.0.1'os.environ['MASTER_PORT'] = '29500'dist.init_process_group(backend, rank=rank_id, world_size=size)fn(rank_id, size)if __name__ == "__main__":size = 4processes = []mp.set_start_method("spawn")for rank in range(size):p = mp.Process(target=init_process, args=(rank, size, run))p.start()processes.append(p)for p in processes:p.join()

输出内容为：

一共有4个rank参与了scatter计算，计算之前：rank0 为[1, 2]，rank1 为[3, 4]， rank2为[5, 6]， rank3为[7, 8]，scatter list为[0,0], [1,1], [2,2], [3,3];
scatter计算之后，rank按顺序被分配scatter list的每一个tensor, rank0为[0,0], rank1为 [1, 1] , rank2为 [2, 2], rank3[3, 3];

root@g48r13:/workspace/communication# python scatter.py
before scatter  Rank  1  has data  tensor([3, 4])
before scatter  Rank  0  has data  tensor([1, 2])
before scatter  Rank  2  has data  tensor([5, 6])
scater list: [tensor([0, 0]), tensor([1, 1]), tensor([2, 2]), tensor([3, 3])]
before scatter  Rank  3  has data  tensor([7, 8])
after scatter  Rank  1  has data  tensor([1, 1])
after scatter  Rank  0  has data  tensor([0, 0])
after scatter  Rank  3  has data  tensor([3, 3])
after scatter  Rank  2  has data  tensor([2, 2])

3.3 gather

gather计算方式如上图所示。在pytorch中通过torch.distributed.gather(tensor, gather_list=None, dst=0, group=None, async_op=False) 来实现gather的通信；

参数tensor是所有rank的input tensor
gather_list是dst rank的output 结果
dst为目标dst

使用方式如下：

这里需要注意的是在rank 0（也就是dst rank）中要指定gather_list，并且要在gather_list构建好的tensor，否是会报错

import os
import torch
import torch.distributed as dist
import torch.multiprocessing as mpdef run(rank_id, size):tensor = torch.arange(2, dtype=torch.int64) + 1 + 2 * rank_idprint('before gather',' Rank ', rank_id, ' has data ', tensor)if rank_id == 0:gather_list = [torch.zeros(2, dtype=torch.int64) for _ in range(4)]dist.gather(tensor, dst = 0, gather_list=gather_list)print('after gather',' Rank ', rank_id, ' has data ', tensor)print('gather_list:', gather_list)else:dist.gather(tensor, dst = 0)print('after gather',' Rank ', rank_id, ' has data ', tensor)def init_process(rank_id, size, fn, backend='gloo'):""" Initialize the distributed environment. """os.environ['MASTER_ADDR'] = '127.0.0.1'os.environ['MASTER_PORT'] = '29500'dist.init_process_group(backend, rank=rank_id, world_size=size)fn(rank_id, size)if __name__ == "__main__":size = 4processes = []mp.set_start_method("spawn")for rank in range(size):p = mp.Process(target=init_process, args=(rank, size, run))p.start()processes.append(p)for p in processes:p.join()

输出内容如下：

一共有4个rank参与了gather计算，计算之前：rank0 为[1, 2]，rank1 为[3, 4]， rank2为[5, 6]， rank3为[7, 8]
gather计算之后，gather_list的值为[tensor([1, 2]), tensor([3, 4]), tensor([5, 6]), tensor([7, 8])]

root@g48r13:/workspace/communication# python gather.py
before gather  Rank  0  has data  tensor([1, 2])
before gather  Rank  3  has data  tensor([7, 8])
after gather  Rank  3  has data  tensor([7, 8])
before gather  Rank  1  has data  tensor([3, 4])
before gather  Rank  2  has data  tensor([5, 6])
after gather  Rank  1  has data  tensor([3, 4])
after gather  Rank  2  has data  tensor([5, 6])
after gather  Rank  0  has data  tensor([1, 2])
gather_list: [tensor([1, 2]), tensor([3, 4]), tensor([5, 6]), tensor([7, 8])]

3.4 reduce

reduce的计算方式如上图所示。在pytorch中通过torch.distributed.reduce(tensor, dst, op=<ReduceOp.SUM: 0>, group=None, async_op=False)来实现reduce通信；

参数tensor是需要进行reduce计算的数据，对于dst rank来说，tensor为最终reduce的结果
参数dist设置目标rank的ID
参数op为reduce的计算方式，pytorch中支持的计算方式有SUM, PRODUCT, MIN, MAX, BAND, BOR, and BXOR

使用方式如下：

import os
import torch
import torch.distributed as dist
import torch.multiprocessing as mpdef run(rank_id, size):tensor = torch.arange(2, dtype=torch.int64) + 1 + 2 * rank_idprint('before reudce',' Rank ', rank_id, ' has data ', tensor)dist.reduce(tensor, dst = 3, op=dist.ReduceOp.SUM,)print('after reudce',' Rank ', rank_id, ' has data ', tensor)def init_process(rank_id, size, fn, backend='gloo'):""" Initialize the distributed environment. """os.environ['MASTER_ADDR'] = '127.0.0.1'os.environ['MASTER_PORT'] = '29500'dist.init_process_group(backend, rank=rank_id, world_size=size)fn(rank_id, size)if __name__ == "__main__":size = 4processes = []mp.set_start_method("spawn")for rank in range(size):p = mp.Process(target=init_process, args=(rank, size, run))p.start()processes.append(p)for p in processes:p.join()

执行结果如下：

一共有4个rank参与了gather计算，计算之前：rank0 为[1, 2]，rank1 为[3, 4]， rank2为[5, 6]， rank3为[7, 8]；dst rank设置为3
可见rank 3为reduce sum计算的最终结果；
需要注意这里有个副作用，就是rank 0、rank 1和rank 2的tensor也会被修改

root@g48r13:/workspace/communication# python reduce.py
before reudce  Rank  3  has data  tensor([7, 8])
before reudce  Rank  0  has data  tensor([1, 2])
before reudce  Rank  2  has data  tensor([5, 6])
before reudce  Rank  1  has data  tensor([3, 4])
after reudce  Rank  1  has data  tensor([15, 18])
after reudce  Rank  0  has data  tensor([16, 20])
after reudce  Rank  3  has data  tensor([16, 20]) # reduce 的最终结果
after reudce  Rank  2  has data  tensor([12, 14])

3.5 all-gather

all-gather计算方式如上图所示。在pytorch中通过torch.distributed.all_gather(tensor_list, tensor, group=None, async_op=False)来实现。

参数tensor_list，rank从该参数中获取all-gather的结果
参数tensor，每个rank参与all-gather计算输入数据

使用方式如下：

同gather的使用方式基本一样，区别是all_gather中每个rank都要指定gather_list，并且要在gather_list构建好的tensor，否是会报错；

import os
import torch
import torch.distributed as dist
import torch.multiprocessing as mpdef run(rank_id, size):tensor = torch.arange(2, dtype=torch.int64) + 1 + 2 * rank_idprint('before gather',' Rank ', rank_id, ' has data ', tensor)gather_list = [torch.zeros(2, dtype=torch.int64) for _ in range(4)]dist.all_gather(gather_list, tensor)print('after gather',' Rank ', rank_id, ' has data ', tensor)print('after gather',' Rank ', rank_id, ' has gather list ', gather_list)def init_process(rank_id, size, fn, backend='gloo'):""" Initialize the distributed environment. """os.environ['MASTER_ADDR'] = '127.0.0.1'os.environ['MASTER_PORT'] = '29500'dist.init_process_group(backend, rank=rank_id, world_size=size)fn(rank_id, size)if __name__ == "__main__":size = 4processes = []mp.set_start_method("spawn")for rank in range(size):p = mp.Process(target=init_process, args=(rank, size, run))p.start()processes.append(p)for p in processes:p.join()

执行结果如下：

一共有4个rank参与了gather计算，计算之前：rank0 为[1, 2]，rank1 为[3, 4]， rank2为[5, 6]， rank3为[7, 8]；
执行完gather_list后，每个rank均可以拿到最终gather_list的结果

root@g48r13:/workspace/communication# python all_gather.py
before gather  Rank  0  has data  tensor([1, 2])
before gather  Rank  2  has data  tensor([5, 6])
before gather  Rank  3  has data  tensor([7, 8])
before gather  Rank  1  has data  tensor([3, 4])
after gather  Rank  1  has data  tensor([3, 4])
after gather  Rank  0  has data  tensor([1, 2])
after gather  Rank  3  has data  tensor([7, 8])
after gather  Rank  2  has data  tensor([5, 6])
after gather  Rank  1  has gather list  [tensor([1, 2]), tensor([3, 4]), tensor([5, 6]), tensor([7, 8])]
after gather  Rank  0  has gather list  [tensor([1, 2]), tensor([3, 4]), tensor([5, 6]), tensor([7, 8])]
after gather  Rank  3  has gather list  [tensor([1, 2]), tensor([3, 4]), tensor([5, 6]), tensor([7, 8])]
after gather  Rank  2  has gather list  [tensor([1, 2]), tensor([3, 4]), tensor([5, 6]), tensor([7, 8])]

3.6 all-reduce

all-reduce计算方式如上图所示。在pytorch中通过torch.distributed.all_reduce(tensor, op=<ReduceOp.SUM: 0>, group=None, async_op=False) 来实现all-reduce的调用；

使用方式如下面代码所示

import os
import torch
import torch.distributed as dist
import torch.multiprocessing as mpdef run(rank_id, size):tensor = torch.arange(2, dtype=torch.int64) + 1 + 2 * rank_idprint('before reudce',' Rank ', rank_id, ' has data ', tensor)dist.all_reduce(tensor, op=dist.ReduceOp.SUM)print('after reudce',' Rank ', rank_id, ' has data ', tensor)def init_process(rank_id, size, fn, backend='gloo'):""" Initialize the distributed environment. """os.environ['MASTER_ADDR'] = '127.0.0.1'os.environ['MASTER_PORT'] = '29500'dist.init_process_group(backend, rank=rank_id, world_size=size)fn(rank_id, size)if __name__ == "__main__":size = 4processes = []mp.set_start_method("spawn")for rank in range(size):p = mp.Process(target=init_process, args=(rank, size, run))p.start()processes.append(p)for p in processes:p.join()

输出内内容为：

一共有4个rank参与了all-reduce计算，计算之前：rank0 为[1, 2]，rank1 为[3, 4]， rank2为[5, 6]， rank3为[7, 8]
all-reduce计算之后，所有rank的结果均相同，为rank0-rank3的tensor计算sum的结果[1+3 + 5 + 7, 2 + 4 + 6 + 8]=[16, 20]

root@g48r13:/workspace/communication# python all_reduce.py
before reudce  Rank  3  has data  tensor([7, 8])
before reudce  Rank  2  has data  tensor([5, 6])
before reudce  Rank  0  has data  tensor([1, 2])
before reudce  Rank  1  has data  tensor([3, 4])
after reudce  Rank  0  has data  tensor([16, 20])
after reudce  Rank  3  has data  tensor([16, 20])
after reudce  Rank  2  has data  tensor([16, 20])
after reudce  Rank  1  has data  tensor([16, 20])