在前文中,我们介绍了PipeDream的总体架构和Profile阶段,本文我们继续介绍计算分区阶段。其功能是:依据profile结果确定所有层的运行时间,然后使用动态规划对模型进行划分,将模型划分为不同的stage,以及得到每个stage的replication数。计算结果具体如下图所示:
流水线并行其他文章链接如下:
[源码解析] 深度学习流水线并行Gpipe(1)—流水线基本实现
[源码解析] 深度学习流水线并行GPipe (2) ----- 梯度累积
[源码解析] 深度学习流水线并行之PipeDream(1)— Profile阶段
我们首先看看profile文件 profiler/translation/profiles/gnmt/graph.txt 内容,这里只是做摘录。
node1 -- Input0 -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=0.0, parameter_size=0.000
node4 -- Embedding(32320, 1024, padding_idx=0) -- forward_compute_time=0.073, backward_compute_time=6.949, activation_size=6291456.0, parameter_size=132382720.000
node5 -- EmuBidirLSTM( (bidir): LSTM(1024, 1024, bidirectional=True) (layer1): LSTM(1024, 1024) (layer2): LSTM(1024, 1024)) -- forward_compute_time=5.247, backward_compute_time=0.016, activation_size=12582912.0, parameter_size=67174400.000
node2 -- Input1 -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=0.0, parameter_size=0.000
node6 -- Dropout(p=0.2) -- forward_compute_time=0.077, backward_compute_time=0.196, activation_size=12582912.0, parameter_size=0.000
node7 -- LSTM(2048, 1024) -- forward_compute_time=3.190, backward_compute_time=5.348, activation_size=[6291456.0; 131072.0; 131072.0], parameter_size=50364416.000
node8 -- __getitem__(0) -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=6291456.0, parameter_size=0.000
node9 -- __getitem__(1) -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=131072.0, parameter_size=0.000
node10 -- Dropout(p=0.2) -- forward_compute_time=0.064, backward_compute_time=0.128, activation_size=6291456.0, parameter_size=0.000
node11 -- LSTM(1024, 1024) -- forward_compute_time=2.491, backward_compute_time=4.203, activation_size=[6291456.0; 131072.0; 131072.0], parameter_size=33587200.000
node12 -- __getitem__(0) -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=6291456.0, parameter_size=0.000
node13 -- __getitem__(1) -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=131072.0, parameter_size=0.000
node14 -- Add -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=6291456.0, parameter_size=0.000
node15 -- Dropout(p=0.2) -- forward_compute_time=0.059, backward_compute_time=0.121, activation_size=6291456.0, parameter_size=0.000
node16 -- LSTM(1024, 1024) -- forward_compute_time=2.492, backward_compute_time=4.201, activation_size=[6291456.0; 131072.0; 131072.0], parameter_size=33587200.000
node17 -- __getitem__(0) -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=6291456.0, parameter_size=0.000
......
node1 -- node4
node4 -- node5
node2 -- node5
node5 -- node6
node6 -- node7
node7 -- node8
node7 -- node9
node8 -- node10
node10 -- node11
node11 -- node12
node11 -- node13
node12 -- node14
node8 -- node14
node14 -- node15
node15 -- node16
node16 -- node17
node16 -- node18
node17 -- node19
node14 -- node19
......
在前文我们也提到了几个挑战,其中有:
因此当跨机器将层划分为不同的阶段时,PipeDream的自动划分算法必须确保每个阶段大致执行相同的总工作量。同时还必须确保各阶段之间通信的数据量尽可能小,以避免通信中断。
PipeDream的自动划分算法总体目标是输出一个平衡的管道,算法如下:
这个划分问题等价于最小化流水线的最慢阶段所花费的时间,并且具有最优子问题属性:在给定worker工作量前提下,吞吐量最大化的流水线由一系列子流水线构成,其中每一个子流水线针对较小worker工作量来最大化自己的输出。因此PipeDream使用动态规划来寻找最优解。
这里给出对应的架构图如下:
我们下面先看看计算分区之前的准备工作:图相关工作和构建反链。
图的定义位于 graph/graph.py 文件之中,主要数据结构有两个:Graph 和 Node。
Graph就是图的数据结构,其主要成员包括:
class Graph(object):
def __init__(self, node=None):
self.nodes = {} # 节点
if node is not None:
self.nodes[node.node_id] = node
self.edges = {} # 出边
self.in_edges = {} # 入边
self._predecessors = {} #每个节点的前序节点
self._successors = {} # 每个节点的后序节点
self._augmented_antichains = {}
self._deaugmented_augmented_antichains = {}
self._next_antichains = {}
self._antichain_dag = None # 反链DAG
if node is not None:
self.in_edges[node.node_id] = list()
节点定义如下,里面就是从profile获取到的结构,比如:
class Node(object):
def __init__(self, node_id, node_desc="", forward_compute_time=0.0,
backward_compute_time=0.0, activation_size=0.0, parameter_size=0.0,
stage_id=None):
self.node_id = node_id
self.node_desc = node_desc
self.forward_compute_time = forward_compute_time
self.backward_compute_time = backward_compute_time
self.activation_size = activation_size
self.parameter_size = parameter_size
self.stage_id = stage_id
self.depth = None
self.height = None
我们打印出运行时看看,可以发现 Graph 的具体情况。
gr = {Graph}
# 边
edges = {dict: 39}
'node1' = {list: 1}
0 = {Node} node4 -- Embedding(32320, 1024, padding_idx=0) -- forward_compute_time=0.073, backward_compute_time=6.949, activation_size=6291456.0, parameter_size=132382720.000
'node4' = {list: 1}
0 = {Node} node5 -- EmuBidirLSTM( (bidir): LSTM(1024, 1024, bidirectional=True) (layer1): LSTM(1024, 1024) (layer2): LSTM(1024, 1024)) -- forward_compute_time=5.247, backward_compute_time=0.016, activation_size=12582912.0, parameter_size=67174400.000
......
# 输入边
in_edges = {dict: 44}
'node4' = {list: 1}
0 = {Node} node1 -- Input0 -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=0.0, parameter_size=0.000
'node5' = {list: 2}
0 = {Node} node4 -- Embedding(32320, 1024, padding_idx=0) -- forward_compute_time=0.073, backward_compute_time=6.949, activation_size=6291456.0, parameter_size=132382720.000
1 = {Node} node2 -- Input1 -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=0.0, parameter_size=0.000
......
# 节点
nodes = {dict: 48}
'node1' = {Node} node1 -- Input0 -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=0.0, parameter_size=0.000
'node4' = {Node} node4 -- Embedding(32320, 1024, padding_idx=0) -- forward_compute_time=0.073, backward_compute_time=6.949, activation_size=6291456.0, parameter_size=132382720.000
'node5' = {Node} node5 -- EmuBidirLSTM( (bidir): LSTM(1024, 1024, bidirectional=True) (layer1): LSTM(1024, 1024) (layer2): LSTM(1024, 1024)) -- forward_compute_time=5.247, backward_compute_time=0.016, activation_size=12582912.0, parameter_size=67174400.000
......
# 前置节点
_predecessors = {dict: 36}
'node4' = {set: 0} set()
__len__ = {int} 0
'node5' = {set: 1} {<graph.graph.Node object at 0x7fb055e4bf28>}
{Node} node4 -- Embedding(32320, 1024, padding_idx=0) -- forward_compute_time=0.073, backward_compute_time=6.949, activation_size=6291456.0, parameter_size=132382720.000
__len__ = {int} 1
'node6' = {set: 2} {<graph.graph.Node object at 0x7fb055e4bf98>, <graph.graph.Node object at 0x7fb055e4bf28>}
{Node} node5 -- EmuBidirLSTM( (bidir): LSTM(1024, 1024, bidirectional=True) (layer1): LSTM(1024, 1024) (layer2): LSTM(1024, 1024)) -- forward_compute_time=5.247, backward_compute_time=0.016, activation_size=12582912.0, parameter_size=67174400.000
{Node} node4 -- Embedding(32320, 1024, padding_idx=0) -- forward_compute_time=0.073, backward_compute_time=6.949, activation_size=6291456.0, parameter_size=132382720.000
__len__ = {int} 2
'node7' = {set: 3} {<graph.graph.Node object at 0x7fb055e4bf98>, <graph.graph.Node object at 0x7fb055e4bf28>, <graph.graph.Node object at 0x7fb055e670f0>}
{Node} node5 -- EmuBidirLSTM( (bidir): LSTM(1024, 1024, bidirectional=True) (layer1): LSTM(1024, 1024) (layer2): LSTM(1024, 1024)) -- forward_compute_time=5.247, backward_compute_time=0.016, activation_size=12582912.0, parameter_size=67174400.000
{Node} node4 -- Embedding(32320, 1024, padding_idx=0) -- forward_compute_time=0.073, backward_compute_time=6.949, activation_size=6291456.0, parameter_size=132382720.000
{Node} node6 -- Dropout(p=0.2) -- forward_compute_time=0.077, backward_compute_time=0.196, activation_size=12582912.0, parameter_size=0.000
__len__ = {int} 3
# 其他变量
_antichain_dag = {NoneType} None
_augmented_antichains = {dict: 0} {}
_deaugmented_augmented_antichains = {dict: 0} {}
_next_antichains = {dict: 0} {}
_successors = {dict: 0} {}
图是由profile文件的字符串构建出来。找出来profile文件内容我们就可以知道,具体是针对每行进行不同处理。
node1 -- Input0 -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=0.0, parameter_size=0.000
node4 -- Embedding(32320, 1024, padding_idx=0) -- forward_compute_time=0.073, backward_compute_time=6.949, activation_size=6291456.0, parameter_size=132382720.000
node5 -- EmuBidirLSTM( (bidir): LSTM(1024, 1024, bidirectional=True) (layer1): LSTM(1024, 1024) (layer2): LSTM(1024, 1024)) -- forward_compute_time=5.247, backward_compute_time=0.016, activation_size=12582912.0, parameter_size=67174400.000
node1 -- node4
node4 -- node5
node2 -- node5
构建图具体代码如下:
@staticmethod
def from_str(graph_str):
gr = Graph()
graph_str_lines = graph_str.strip().split('\n')
for graph_str_line in graph_str_lines: # 逐行处理
if not graph_str_line.startswith('\t'):
node = Node.from_str(graph_str_line.strip()) # 构建节点
gr.nodes[node.node_id] = node
else:
# 构建边
[in_node_id, node_id] = graph_str_line.strip().split(" -- ")
if node_id not in gr.in_edges: # 每个节点的输入边
gr.in_edges[node_id] = [gr.nodes[in_node_id]]
else:
gr.in_edges[node_id].append(gr.nodes[in_node_id])
if in_node_id not in gr.edges: # 每个节点的输出边
gr.edges[in_node_id] = [gr.nodes[node_id]]
else:
gr.edges[in_node_id].append(gr.nodes[node_id])
return gr
构建节点具体代码如下:
@staticmethod
def from_str(node_str):
node_str_tokens = node_str.strip().split(" -- ")
node_id = node_str_tokens[0] # 节点名字
node_desc = node_str_tokens[1] # 节点描述
node_metadata = node_str_tokens[2] # 元数据
stage_id = None
if len(node_str_tokens) > 3:
stage_id = int(node_str_tokens[3].split("=")[1]) # 阶段信息
[forward_compute_time, backward_compute_time, activation_size, parameter_size] = node_metadata.split(", ")
forward_compute_time = float(forward_compute_time.split("=")[1]) # 前向传播计算时间
backward_compute_time = float(backward_compute_time.split("=")[1]) # 后向传播计算时间
if "[" in activation_size:
activation_size = activation_size.split("=")[1] # 激活值大小
activation_size = sum([float(x) for x in activation_size.lstrip("[").rstrip("]").split("; ")])
else:
activation_size = float(activation_size.split("=")[1])
parameter_size = float(parameter_size.split("=")[1]) # 参数大小
# 构建节点
return Node(node_id, node_desc, forward_compute_time=forward_compute_time,
backward_compute_time=backward_compute_time, activation_size=activation_size,
parameter_size=parameter_size, stage_id=stage_id)
在有向无环图中,有如下的一些概念:
链 :一条链是一些点的集合,在此链上的任意两个点x, y,满足以下条件:或者 x 能到达 y ,或者 y 能到达 x 。也可以认为是某一个偏序集S的全序子集(所谓全序是指其中任意两个元素可以比较)
反链 :一条反链也是一些点的集合,在此链上任意两个点x, y,满足如下条件: x 不能到达 y,且 y 也不能到达 x。也可以认为是某一个偏序集S的子集,其中任意两个元素不可比较。
在PipeDream的图数据结构之中,也有反链的概念。反链节点定义如下:
class AntichainNode(Node):
def __init__(self, node_id, antichain, node_desc=""):
self.antichain = antichain
self.output_activation_size = 0.0
super(AntichainNode, self).__init__(node_id, node_desc)
因为此处过于复杂,所以我们会在下面用一节专门分析。
因为本节概念比较绕,所以我们先提前剧透。
寻找某节点后续反链的目的就是找到下一个图分割点 A(可能是若干node的组合),为了确定 A 的运行时间(或者其他信息),我们需要找到 A 的增强反链。
此处具体代码位于optimizer_graph_hierarchical.py
文件。
我们利用如下逻辑来演示:
+-------+ +-------+
| node1 | | node2 |
+---+---+ +---+---+
| |
| |
| |
v v
+---+---+ +---+---+ +-------+ +-------+
| node4 +-----> | node5 +------> | node6 +------->+ node7 |
+-------+ +-------+ +-------+ +-+-+---+
| |
| |
+-------------+ |
| |
v v
+----+--+ +---+---+
| node9 | | node8 +-----+
+-------+ +---+---+ |
| |
+---------------------------------+ |
| |
v |
+----+---+ +--------+ +--------+ |
| node10 +-----> | node11 +------> | node12 | |
+--------+ +---+----+ +----+---+ |
| | |
| | |
v v |
+---+----+ +----+---+ |
| node13 | | node14 +<---+
+--------+ +-+----+-+
| |
+------+ +---+
| |
v v
+----+---+ +--+-----+
| node15 | | node19 |
+--------+ +--------+
我们首先从 main 函数看起。main函数第一部分是构建反链和拓扑排序,具体如下:
具体代码如下:
def main(all_num_machines, profile_filename, network_bandwidths, memory_size,
straight_pipeline, use_memory_constraint, use_fewer_machines,
activation_compression_ratio, output_directory,
print_configuration=True, verbose=False):
gr = graph.Graph.from_str(open(profile_filename, 'r').read())
# Zero out all metadata associated with inputs in graph, since the optimizer
# shouldn't really get a choice with where to place the input (should always
# be in the first stage).
# 排除干扰,因为input必然在第一层,没必要让优化器再来选择把输入放在哪里,所以先去除,后续会再加上。
sources = gr.sources() # 对图的输入进行处理
nodes_to_remove = OrderedDict()
for source in sources:
if source.node_desc.startswith("Input"): # 只处理input
source.forward_compute_time = 0.0
source.backward_compute_time = 0.0
source.activation_size = 0.0
source.parameter_size = 0.0
nodes_to_remove[source] = []
for out_node in gr.edges[source.node_id]:
nodes_to_remove[source].append(out_node) # 记录这些删除source对应了哪些out节点,因为后续还要处理
gr.remove_node(source) # 在图中移除这些input source
# Remove all unneeded sinks that are not used, makes code generation and
# optimization easier.
sinks = gr.sinks() # 对图的输出进行处理,移除没有用到的输出
for sink in sinks:
if sink.node_desc.startswith("__getitem__"):
gr.remove_node(sink)
antichain_gr = gr.antichain_dag() # 得到反链DAG
states = antichain_gr.topological_sort() # 拓扑排序,得到一个排序好的节点列表
# 后续代码暂时省略
这里再取出反链节点定义如下,可以看出来和代码对应关系。
class AntichainNode(Node):
def __init__(self, node_id, antichain, node_desc=""):
self.antichain = antichain
self.output_activation_size = 0.0
super(AntichainNode, self).__init__(node_id, node_desc)
首先要介绍先增强反链概念。每个节点的增强反链包括:本身节点 + 部分前序节点。
这个前序节点的选取算法是:
从下面图例中可以看出来,如果某一个节点 A,其前置节点中有一个分叉节点 Z,且这个分叉之中,有一个分叉绕过了节点 A,则对于节点 A,他的增强反链就是 [A, Z]。
对于增强反链概念,可以理解为:对于节点 A,他只有把节点 Z 一起考虑,才能唯一确定自己节点的运行时间。因为如果思考节点 A 的运行时间,我理解的大致思路是:
所以,需要把 [ A,Z ] 放在一起作为一个状态考虑,事实上 PipeDream 就是这么处理的,用 [ A,Z ] 这个状态来统一计算。
因为作为一个状态考虑,所以给节点 A 计算输出激活值大小,具体是通过遍历其反链(增强反链)来计算,就是把其增强反链的前序节点给自己的输出都叠加起来。
+-----+ +-----+
| X | | Z |
+--+--+ +--+-++
| | |
| | |
+------+ +-------+ |
| | |
v v |
++---++ |
| A | |
++-+--+ |
| | |
+---------+ | |
| | |
v v v
+---+-+ +--+--+ +-+---+
| B | | C | | D |
+-----+ +-----+ +-----+
在代码之中,_augmented_antichains
是增强反链,也是一个字典类,key是节点名字,value是 key 节点的增强反链,比如:
augment_antichain函数作用就是对每个节点,找到其增强反链。
def augment_antichain(self, antichain):
# 参数 antichain 是一个节点列表
antichain_key = tuple(sorted(antichain))
# 如果key已经在扩大反链之中,就直接返回对应key的增强反链
if antichain_key in self._augmented_antichains:
return self._augmented_antichains[antichain_key]
extra_nodes = set()
all_predecessors = set()
# 遍历参数list之中的反链节点,获取每个节点的前置节点,归并在all_predecessors之中。
for antichain_node in antichain:
predecessors = self.predecessors(antichain_node)
all_predecessors = all_predecessors.union(predecessors)
# 遍历参数list之中的反链节点
for antichain_node in antichain:
# 获取每个反链节点的前置节点列表
predecessors = self.predecessors(antichain_node)
# 遍历每个前置节点
for predecessor in predecessors:
# 看每个前置节点的出边,如果出边不在前置节点列表之中,且 出边节点不等于本反链节点
for out_node in self.edges[predecessor.node_id]:
if out_node not in predecessors and out_node.node_id != antichain_node:
# 把这个前置节点插入到附加节点列表中
extra_nodes.add(predecessor.node_id)
# 最终把个附加节点列表插入到增强节点之中
self._augmented_antichains[antichain_key] = list(extra_nodes) + antichain
return self._augmented_antichains[antichain_key]
比如对应下图中的逻辑,初始化之后,_augmented_antichains 就是
_augmented_antichains = {dict: 1}
('node4',) = {list: 1} ['node4']
后续迭代node 5之后,_augmented_antichains 就是
_augmented_antichains = {dict: 2}
('node4',) = {list: 1} ['node4']
('node5',) = {list: 1} ['node5']
__len__ = {int} 2
继续迭代,增强反链为:
_augmented_antichains = {dict: 7}
('node4',) = {list: 1} ['node4'] # node4的增强反链只有自己
('node5',) = {list: 1} ['node5'] # node5的增强反链只有自己
('node6',) = {list: 1} ['node6']
('node7',) = {list: 1} ['node7']
('node8',) = {list: 1} ['node8']
('node10',) = {list: 2} ['node8', 'node10'] # node10的增强反链是'node8', 'node10'
('node14',) = {list: 1} ['node14']
('node11',) = {list: 2} ['node8', 'node11'] # node11的增强反链是'node8', 'node11'
('node15',) = {list: 2} ['node14', 'node15']
('node19',) = {list: 1} ['node19']
('node12',) = {list: 2} ['node8', 'node12']
('node16',) = {list: 2} ['node14', 'node16']
('node23',) = {list: 2} ['node20', 'node23']
('node17',) = {list: 2} ['node14', 'node17']
图例中可以看出来,因为有 node 8的出边 [node 8,node 14] 存在,对于 node 10, node 11, node 12 来说,他们必须把 node 8 加入自己的增强反链之中。
对于 node 10,我们可以认为,必须结合 node 8之后,node 10 才能确定 node 10 的运行时间。下面图上标记出来了 node 10 的 augmented 反链(本身节点 + 部分前序节点)。
+-------+ +-------+
| node1 | | node2 |
+---+---+ +---+---+
| |
| |
| |
v v
+---+---+ +---+---+ +-------+ +-------+
| node4 +-----> | node5 +------> | node6 +------->+ node7 |
+-------+ +-------+ +-------+ +-+-+---+
| |
| |
+-------------+ |
| |
v v augmented
+----+--+ +---+---+
| node9 | | node8 +-----+
+-------+ +---+---+ |
| |
+---------------------------------+ |
| |
v |
+----+---+ +--------+ +--------+ |
antichain | node10 +-----> | node11 +------> | node12 | |
+--------+ +---+----+ +----+---+ |
augmented | | |
| | |
v v |
+---+----+ +----+---+ |
| node13 | | node14 +<---+
+--------+ +-+----+-+
| |
+------+ +---+
| |
v v
+----+---+ +--+-----+
| node15 | | node19 |
+--------+ +--------+
在代码之中,_next_antichains 是一个字典类,key是节点名字,value是 key 节点的后续反链。
比如,对于 node A 来说,下一个反链是 [ node B, node C ],其中 node B 和 node C 彼此之间无法排序。寻找反链的目的就是找到下一个图分割点。
+-----+ +-----+
| X | | Z |
+--+--+ +--+-++
| | |
| | |
+------+ +-------+ |
| | |
v v |
++---++ |
| A | |
++-+--+ |
| | |
+---------+ | |
| | |
v v v
+---+-+ +--+--+ +-+---+
| B | | C | | D |
+-----+ +-----+ +-----+
对于每个节点 antichain ,next_antichains 函数获取其后续反链。
def next_antichains(self, antichain):
# 构建antichain的反链key,其实就是 antichain 自己作为key
antichain_key = tuple(sorted(antichain))
# 如果key已经在后续反链之中,则返回这个后续反链
if antichain_key in self._next_antichains:
return self._next_antichains[antichain_key]
next_antichains = []
antichain_set = set(antichain)
# 获取 antichain 的增强反链
augmented_antichain = self.augment_antichain(antichain)
# 遍历增强反链
for augmented_antichain_node in augmented_antichain:
# 遍历增强反链某节点的出边
next_nodes = self.edges[augmented_antichain_node] if augmented_antichain_node in self.edges else []
# 遍历增强反链某节点的出边
for next_node in next_nodes:
# 如果出边节点已经在反链集合之中,跳过,进入下一循环
if next_node.node_id in antichain_set:
continue
# 如果出边节点是后续反链,则假如到反链列表
if self.is_next_antichain(augmented_antichain, next_node.node_id):
next_antichain = self.construct_antichain(augmented_antichain,
augmented_antichain_node,
next_node.node_id)
next_antichains.append(next_antichain)
# 最终把反链列表设置为key对应的反链
self._next_antichains[antichain_key] = next_antichains
return self._next_antichains[antichain_key]
is_next_antichain 方法用来判断某新节点是否为后续反链。
def is_next_antichain(self, augmented_antichain, new_node):
successors = self.successors(new_node)
augmented_antichain_set = set(augmented_antichain)
# 遍历新节点的后续节点
for successor in successors:
# 如果后续节点有一个在增强节点之中,就返回false,说明不是后续反链
if successor.node_id in augmented_antichain_set:
return False
# 否则就是后续反链
return True
_next_antichains举例如下,大家可以结合之前的增强反链对比看看。
所以 node 10 的后续反链是 [ [‘node14’] ,[ ‘node11’] ]。
对比 看看,node 10 的增强反链是 [‘node8’, ‘node10’],
_next_antichains = {dict: 99}
('node4',) = {list: 1} [['node5']]
('node5',) = {list: 1} [['node6']]
('node6',) = {list: 1} [['node7']]
('node7',) = {list: 1} [['node8']]
('node8',) = {list: 2} [['node10'], ['node14']]
('node10',) = {list: 2} [['node14'], ['node11']] # 这里
('node14',) = {list: 2} [['node15'], ['node19']]
('node11',) = {list: 2} [['node14'], ['node12']]
('node15',) = {list: 2} [['node19'], ['node16']]
('node19',) = {list: 1} [['node23']]
('node12',) = {list: 2} [['node14'], ['node14']]
('node16',) = {list: 2} [['node19'], ['node17']]
具体如下图,可以看出来,node 11和 node 14确实是 node 10的后续反链,就是在这两个节点上可以对于图进行分割。
可以这么理解:对于 node 10 来说,下一个反链是 [ node 11, node 14],其中 node 11 和 node 14 彼此之间无法排序。寻找后续反链的目的就是找到下一个图分割点。
+-------+ +-------+
| node1 | | node2 |
+---+---+ +---+---+
| |
| |
| |
v v
+---+---+ +---+---+ +-------+ +-------+
| node4 +-----> | node5 +------> | node6 +------->+ node7 |
+-------+ +-------+ +-------+ +-+-+---+
| |
| |
+-------------+ |
| |
v v augmented
+----+--+ +---+---+
| node9 | | node8 +-----+
+-------+ +---+---+ |
| |
+---------------------------------+ |
| |
v next |
+----+---+ +--------+ +--------+ |
antichain | node10 +-----> | node11 +------> | node12 | |
+--------+ +---+----+ +----+---+ |
augmented | | |
| | |
v next v |
+---+----+ +----+---+ |
| node13 | | node14 +<---+
+--------+ +-+----+-+
| |
+------+ +---+
| |
v v
+----+---+ +--+-----+
| node15 | | node19 |
+--------+ +--------+
antichain_dag 的目的是依据 增强反链列表 和 后续反链列表来构建一个反链 DAG。
我们以上面的图例进行讲解,以 node 8 为例。
def antichain_dag(self):
if self._antichain_dag is not None:
return self._antichain_dag
antichain_dag = Graph()
antichain_id = 0
antichain = [self.sources()[0].node_id] # 获取source第一个节点。
# 构建首节点,同时利用 augment_antichain 来往_augmented_antichains 之中添加首节点。
source_node = AntichainNode("antichain_%d" % antichain_id, self.augment_antichain(antichain))
antichain_dag.source = source_node
antichain_queue = [antichain] # 把第一个节点插入queue
antichain_mapping = {tuple(sorted(antichain)): source_node}
# 如果queue之中还有节点
while len(antichain_queue) > 0:
antichain = antichain_queue.pop(0) # 弹出第一个节点,赋值为 antichain,这里为 node 8
# key就是由 antichain 节点名字构建,比如 antichain_key = {tuple: 1} node8
antichain_key = tuple(sorted(antichain))
# 如果 antichain_key 已经位于self._next_antichains之中,即 antichain_key 的后续反链已经被记录,就跳过去
if antichain_key in self._next_antichains:
continue
# 获取 antichain 的后续反链,对于8,这里是[[10],[14]]
next_antichains = self.next_antichains(antichain)
# 遍历后续反链[10,14]
for next_antichain in next_antichains:
# 下一个反链节点的key 10
next_antichain_key = tuple(sorted(next_antichain))
if next_antichain_key not in antichain_mapping: # 如果存在,就跳过
antichain_id += 1
# 下一反链节点 10 被设置为其增强节点 [ 8, 10 ]
next_antichain_node = AntichainNode("antichain_%d" % antichain_id, self.augment_antichain(next_antichain))
# 设置 antichain_mapping
antichain_mapping[next_antichain_key] = next_antichain_node
# 向 反链DAG 插入边:
antichain_dag.add_edge(antichain_mapping[antichain_key],
antichain_mapping[next_antichain_key])
# 把最新反链节点插入queue,下次迭代使用
antichain_queue.append(next_antichain)
self._antichain_dag = antichain_dag
return antichain_dag
这里其实目的是设置 antichain_mapping。
流程是:
可以看到,寻找某节点后续反链的目的就是找到下一个图分割点 A,然后为了确定 A 的运行时间(或者其他信息),需要找到 A 的增强反链(一些增强反链就是一些状态),A 的 antichain_mapping 就是其增强反链。
antichain_mapping 示例如下:
antichain_mapping = {dict: 99}
('node4',) = {AntichainNode} antichain_0 -- ['node4']
('node5',) = {AntichainNode} antichain_1 -- ['node5']
('node6',) = {AntichainNode} antichain_2 -- ['node6']
('node7',) = {AntichainNode} antichain_3 -- ['node7']
('node8',) = {AntichainNode} antichain_4 -- ['node8']
('node10',) = {AntichainNode} antichain_5 -- ['node8', 'node10'] # 最新设置
('node14',) = {AntichainNode} antichain_6 -- ['node14']
('node11',) = {AntichainNode} antichain_7 -- ['node8', 'node11']
('node15',) = {AntichainNode} antichain_8 -- ['node14', 'node15']
('node19',) = {AntichainNode} antichain_9 -- ['node19']
('node12',) = {AntichainNode} antichain_10 -- ['node8', 'node12']
('node16',) = {AntichainNode} antichain_11 -- ['node14', 'node16']
('node23',) = {AntichainNode} antichain_12 -- ['node20', 'node23']
('node17',) = {AntichainNode} antichain_13 -- ['node14', 'node17']
antichain_dag 示例如下,可以认为就是增强反链DAG:
antichain_dag = {Graph}
nodes = {dict: 99}
'antichain_0' = {AntichainNode} antichain_0 -- ['node4']
'antichain_1' = {AntichainNode} antichain_1 -- ['node5']
'antichain_2' = {AntichainNode} antichain_2 -- ['node6']
'antichain_3' = {AntichainNode} antichain_3 -- ['node7']
'antichain_4' = {AntichainNode} antichain_4 -- ['node8']
'antichain_5' = {AntichainNode} antichain_5 -- ['node8', 'node10']
'antichain_6' = {AntichainNode} antichain_6 -- ['node14']
'antichain_7' = {AntichainNode} antichain_7 -- ['node8', 'node11']
'antichain_8' = {AntichainNode} antichain_8 -- ['node14', 'node15']
'antichain_9' = {AntichainNode} antichain_9 -- ['node19']
'antichain_10' = {AntichainNode} antichain_10 -- ['node8', 'node12']
'antichain_11' = {AntichainNode} antichain_11 -- ['node14', 'node16']
'antichain_12' = {AntichainNode} antichain_12 -- ['node20', 'node23']
'antichain_13' = {AntichainNode} antichain_13 -- ['node14', 'node17']
'antichain_14' = {AntichainNode} antichain_14 -- ['node20', 'node30', 'node23']
'antichain_15' = {AntichainNode} antichain_15 -- ['node20', 'node36', 'node23']
'antichain_16' = {AntichainNode} antichain_16 -- ['node20', 'node43', 'node23']
'antichain_17' = {AntichainNode} antichain_17 -- ['node20', 'node23', 'node24']
得到了增强反链之后,需要进行拓扑排序之后才能使用。
antichain_gr = gr.antichain_dag()
states = antichain_gr.topological_sort()
得出拓扑排序的目的是:如果按照拓扑序列的顶点次序,在到达某节点之前,可以保证它的所有前序活动都已经完成,从而整个工程顺序执行,不会冲突。
在图论中,**拓扑排序(Topological Sorting)是一个有向无环图(DAG, Directed Acyclic Graph)**的所有顶点的线性序列。且该序列必须满足下面两个条件:
有向无环图(DAG)才有拓扑排序,非DAG图没有拓扑排序一说。一个有向无环图可以有一个或多个拓扑排序序列。
例如,下面这个图:
+--------+ +--------+
| +----------------> | |
| 1 | | 4 +------------+
| | +-----------> | | |
+-----+--+ | +---+----+ |
| | | v
| | | +--+--+
| | | +---> | 5 |
| | | | +-----+
v | | |
| v |
+--------+ | +---+-----+ |
| +----+ | | |
| 2 +----------------->+ 3 +--+
| | | |
+--------+ +---------+
得到拓扑排序后的结果是 { 1, 2, 4, 3, 5 }。
这里的拓扑排序算法使用的是深度优先排序。
def topological_sort(self):
# Algorithm from https://en.wikipedia.org/wiki/Topological_sorting
self.sorted_nodes = []
self.marked_nodes = set()
self.temporarily_marked_nodes = set()
nodes = list(self.nodes.values())
nodes.sort(key=lambda x: x.node_desc)
for node in nodes:
if node.node_id in self.marked_nodes:
continue
self.topological_sort_helper(node.node_id)
return [self.nodes[node_id] for node_id in self.sorted_nodes]
def topological_sort_helper(self, node_id):
if node_id in self.marked_nodes:
return
if node_id in self.temporarily_marked_nodes:
raise Exception("Graph has a cycle")
self.temporarily_marked_nodes.add(node_id)
if node_id in self.edges:
out_nodes = list(self.edges[node_id])
out_nodes.sort(key=lambda x: (x.node_desc, x.height))
for out_node in out_nodes:
self.topological_sort_helper(out_node.node_id)
self.marked_nodes.add(node_id)
self.temporarily_marked_nodes.remove(node_id)
self.sorted_nodes.insert(0, node_id)
最终结果举例如下,可以和上面的反链DAG antichain_dag 比对,看看异同:
states = {list: 99}
00 = {AntichainNode} antichain_0 -- ['node4']
01 = {AntichainNode} antichain_1 -- ['node5']
02 = {AntichainNode} antichain_2 -- ['node6']
03 = {AntichainNode} antichain_3 -- ['node7']
04 = {AntichainNode} antichain_4 -- ['node8']
05 = {AntichainNode} antichain_5 -- ['node8', 'node10']
06 = {AntichainNode} antichain_7 -- ['node8', 'node11']
07 = {AntichainNode} antichain_10 -- ['node8', 'node12']
08 = {AntichainNode} antichain_6 -- ['node14']
09 = {AntichainNode} antichain_8 -- ['node14', 'node15']
10 = {AntichainNode} antichain_11 -- ['node14', 'node16']
11 = {AntichainNode} antichain_13 -- ['node14', 'node17']
12 = {AntichainNode} antichain_9 -- ['node19']
13 = {AntichainNode} antichain_12 -- ['node20', 'node23']
14 = {AntichainNode} antichain_18 -- ['node23', 'node20', 'node26']
15 = {AntichainNode} antichain_17 -- ['node23', 'node20', 'node24']
16 = {AntichainNode} antichain_32 -- ['node23', 'node20', 'node28']
17 = {AntichainNode} antichain_31 -- ['node23', 'node20', 'node26', 'node24']
18 = {AntichainNode} antichain_63 -- ['node23', 'node20', 'node26', 'node28']
19 = {AntichainNode} antichain_33 -- ['node20', 'node26', 'node29']
20 = {AntichainNode} antichain_16 -- ['node20', 'node43', 'node23']
21 = {AntichainNode} antichain_30 -- ['node23', 'node20', 'node43', 'node26']
22 = {AntichainNode} antichain_29 -- ['node23', 'node20', 'node43', 'node24']
23 = {AntichainNode} antichain_59 -- ['node23', 'node20', 'node43', 'node28']
我们 也可以和如下增强反链比对,看到 states 就是对增强反链DAG进行拓扑排序之后的结果,按照这个顺序进行训练是符合逻辑的。
_augmented_antichains = {dict: 99}
('node4',) = {list: 1} ['node4']
('node5',) = {list: 1} ['node5']
('node6',) = {list: 1} ['node6']
('node7',) = {list: 1} ['node7']
('node8',) = {list: 1} ['node8']
('node10',) = {list: 2} ['node8', 'node10']
('node14',) = {list: 1} ['node14']
('node11',) = {list: 2} ['node8', 'node11']
('node15',) = {list: 2} ['node14', 'node15']
('node19',) = {list: 1} ['node19']
('node12',) = {list: 2} ['node8', 'node12']
('node16',) = {list: 2} ['node14', 'node16']
('node23',) = {list: 2} ['node20', 'node23']
('node17',) = {list: 2} ['node14', 'node17']
('node23', 'node30') = {list: 3} ['node20', 'node30', 'node23']
('node23', 'node36') = {list: 3} ['node20', 'node36', 'node23']
('node23', 'node43') = {list: 3} ['node20', 'node43', 'node23']
('node24',) = {list: 3} ['node23', 'node20', 'node24']
('node26',) = {list: 3} ['node23', 'node20', 'node26']
('node23', 'node30', 'node36') = {list: 4} ['node20', 'node36', 'node30', 'node23']
('node23', 'node30', 'node43') = {list: 4} ['node20', 'node43', 'node30', 'node23']
('node31',) = {list: 3} ['node20', 'node26', 'node31']
('node24', 'node30') = {list: 4} ['node23', 'node20', 'node30', 'node24']
('node26', 'node30') = {list: 4} ['node23', 'node20', 'node30', 'node26']
('node23', 'node36', 'node43') = {list: 4} ['node20', 'node43', 'node36', 'node23']
('node37',) = {list: 4} ['node32', 'node20', 'node26', 'node37']
('node24', 'node36') = {list: 4} ['node23', 'node20', 'node36', 'node24']
('node26', 'node36') = {list: 4} ['node23', 'node20', 'node36', 'node26']
('node44',) = {list: 2} ['node40', 'node44']
('node24', 'node43') = {list: 4} ['node23', 'node20', 'node43', 'node24']
('node26', 'node43') = {list: 4} ['node23', 'node20', 'node43', 'node26']
('node24', 'node26') = {list: 4} ['node23', 'node20', 'node26', 'node24']
因为目前的算法比较复杂,所以我们暂时总结一下目前为止的工作:
_augmented_antichains
。_next_antichains
和 _augmented_antichains
进行处理,构建一个反链 DAG,就是变量 antichain_dag。至此,图已经依据后续反链被分割成若干状态(states),每个状态很重要的一个属性是其增强反链。states 就是对增强反链进行拓扑排序之后的结果,按照这个顺序进行训练是符合逻辑的。
自动分区算法具体分为两部分。
下面我们逐一分析。
main函数接下来与计算分区相关的逻辑如下:
具体代码如下:
def main(all_num_machines, profile_filename, network_bandwidths, memory_size,
straight_pipeline, use_memory_constraint, use_fewer_machines,
activation_compression_ratio, output_directory,
print_configuration=True, verbose=False):
gr = graph.Graph.from_str(open(profile_filename, 'r').read())
# Zero out all metadata associated with inputs in graph, since the optimizer
# shouldn't really get a choice with where to place the input (should always
# be in the first stage).
# 排除干扰,因为input必然在第一层,没必要让优化器再来选择把输入放在哪里,所以先去除,后续会再加上。
sources = gr.sources() # 对图的输入进行处理
nodes_to_remove = OrderedDict()
for source in sources:
if source.node_desc.startswith("Input"): # 只处理input
source.forward_compute_time = 0.0
source.backward_compute_time = 0.0
source.activation_size = 0.0
source.parameter_size = 0.0
nodes_to_remove[source] = []
for out_node in gr.edges[source.node_id]:
nodes_to_remove[source].append(out_node) # 记录这些删除source对应了哪些out节点,因为后续还要处理
gr.remove_node(source) # 在图中移除这些input source
# Remove all unneeded sinks that are not used, makes code generation and
# optimization easier.
sinks = gr.sinks() # 对图的输出进行处理,移除没有用到的输出
for sink in sinks:
if sink.node_desc.startswith("__getitem__"):
gr.remove_node(sink)
antichain_gr = gr.antichain_dag() # 得到反链DAG
states = antichain_gr.topological_sort() # 拓扑排序,得到一个排序好的节点列表
###########################################################################
# 之前代码在上节分析过,我们本节从这里继续分析
###########################################################################
states_indices = {} # 为每个状态设置index
for i in range(len(states)):
states_indices[states[i]] = i
##################################### 运行时如下
#states_indices = {dict: 99}
# antichain_0 -- ['node4'] = {int} 0
# antichain_1 -- ['node5'] = {int} 1
# antichain_2 -- ['node6'] = {int} 2
# antichain_3 -- ['node7'] = {int} 3
# antichain_4 -- ['node8'] = {int} 4
# ......
# 给每个状态计算出输出激活值大小,具体是通过遍历其反链(增强反链),可以认为就是其必要前序节点给自己的输出
for i in range(len(states)):
for antichain_node in states[i].antichain:
states[i].output_activation_size += gr.nodes[antichain_node].activation_size
# 给每个状态计算其信息,比如计算时间,激活大小,参数大小等等,都是通过前置节点完成的
for i in range(len(states)):
antichain = states[i].antichain
all_predecessors = gr.all_predecessors(antichain)
states[i].compute_time = 0.0
states[i].activation_size = 0.0
states[i].parameter_size = 0.0
for predecessor in all_predecessors: # 计算所有前置节点的信息
states[i].compute_time += ((predecessor.forward_compute_time +
predecessor.backward_compute_time) / 1000.0)
states[i].activation_size += predecessor.activation_size
states[i].parameter_size += predecessor.parameter_size
gr.reset()
# 得到总体输出大小 & 所有前置节点id,后面计算分区时候需要
output_activation_sizes = [state.output_activation_size for state in states]
all_predecessor_ids = [[states_indices[predecessor] for predecessor in
antichain_gr.predecessors(states[i].node_id)]
for i in range(len(states))]
##################################### 运行时如下
# output_activation_sizes = {list: 99}
# 00 = {float} 6291456.0
# 01 = {float} 12582912.0
# 02 = {float} 12582912.0
# 03 = {float} 6553600.0
# .....
# all_predecessor_ids = {list: 99}
# 00 = {list: 0} []
# 01 = {list: 1} [0]
# 02 = {list: 2} [0, 1]
# 03 = {list: 3} [0, 1, 2]
# 04 = {list: 4} [0, 1, 2, 3]
# 05 = {list: 5} [2, 3, 4, 0, 1]
# 06 = {list: 6} [2, 3, 4, 0, 1, 5]
# 07 = {list: 7} [6, 2, 3, 4, 0, 1, 5]
# ......
compute_times = [] # 初始化计算时间
activation_sizes = [] # 初始化激活值大小
parameter_sizes = [] # 初始化参数值大小
for i in range(len(states)+1): # 具体计算每一个节点的信息,去除他之前节点的影响
compute_times_row = []
activation_sizes_row = []
parameter_sizes_row = []
for j in range(len(states)): # 去除之前的节点
if i == 0: # 列表中第一个节点
compute_times_row.append(states[j].compute_time) # i 到 j 的计算时间
activation_sizes_row.append(states[j].activation_size)
parameter_sizes_row.append(states[j].parameter_size)
else: # 列表中后续节点
if j > (i-1):
compute_times_row.append(states[j].compute_time -
states[i-1].compute_time) # i 到 j 的计算时间
activation_sizes_row.append(states[j].activation_size -
states[i-1].activation_size)
parameter_sizes_row.append(states[j].parameter_size -
states[i-1].parameter_size)
else:
compute_times_row.append(None)
activation_sizes_row.append(None)
parameter_sizes_row.append(None)
compute_times.append(compute_times_row) # 依据profile估计出系统内部的计算时间,compute_times_row 是 i 节点到 后续节点(i+1, i+2, ...)的计算时间,下面类似
activation_sizes.append(activation_sizes_row) # 依据profile估计出系统内部的激活值大小
parameter_sizes.append(parameter_sizes_row) # 依据profile估计出系统内部的参数大小
##################################### 运行时如下
# compute_times = {list: 100}
# 000 = {list: 99} [0.0070220000000000005, 0.012285, 0.012558, 0.021096000000,...
# 001 = {list: 99} [None, 0.005263, 0.005535999999999999, 0.014074000000000003, ...
# 002 = {list: 99} [None, None, 0.00027299999999999894, 0.008811000000000003, ...
# 003 = {list: 99} [None, None, None, 0.008538000000000004, 0.008538, ...
# 004 = {list: 99} [None, None, None, None, -3.469446951953614e-18, 0.000191999999...
counter = 1
all_As = []
num_machines_in_machine = 1 #第一个节点就是1
# all_num_machines, network_bandwidths 是用户在输入中指定
# 遍历机器集&网络带宽组合。流水线可以是straight(数目为1)或者并行(数目为num_machines)
for num_machines, network_bandwidth in zip(all_num_machines, network_bandwidths):
print("Solving optimization problem with %d machines with inter-machine bandwidth of %.2f GB/s" % (num_machines, network_bandwidth / 10**9))
import numpy as np
print(np.array(compute_times))
# 依据目前的信息,以及机器数量,网络带宽等计算分区
A = compute_partitioning(compute_times, activation_sizes, parameter_sizes,
output_activation_sizes, all_predecessor_ids,
num_machines, num_machines_in_machine,
network_bandwidth,
final_level=(counter==len(network_bandwidths)))
num_machines_in_machine = num_machines # 因为计算完了,所以设置为本阶段的机器数目
for i in range(len(compute_times)): # 遍历机器
for j in range(len(compute_times[0])): # 后续机器
compute_times[i][j] = A[i][j][-1][0] # 记录计算时间(本阶段最后一个机器的计算时间)
counter += 1
all_As.append(A) # 添加逻辑关系,就是里面包括了不同阶段的优化逻辑
print(np.array(compute_times))
# 省略后续代码
其中compute_times 是一个计算时间的二维数组,也可以认为是矩阵,具体举例如下。
[w12,w13,w14,w15], // 第一个节点到后续节点的计算时间
[None, w23,w24,w25], // 第二个节点到后续节点的计算时间
[None, None, w34, w35], // 第三个节点到后续节点的计算时间
[None, None, None, w45], // 第四个节点到后续节点的计算时间
activation_sizes 和 parameter_sizes 与之类似。
这里有一些动态规划的算法需要分析。
分割算法试图减少模型的整体训练时间。对于流水线系统,这个问题等价于最小化流水线最慢阶段所花费的时间。该问题具有最优化子问题性质;在给定机器计数的情况下,使吞吐量最大化的管道由子管道组成,这些子管道分别使自己这个子管道的吞吐量最大化。因此,我们可以用动态规划来寻找这个问题的最优解。
分区算法获取profiling步骤的输出,并计算:
1)将层划分为多个阶段,
2)每个阶段的复制因子(worker数),
3)保持训练管道繁忙的最佳动态小批量数。
PipeDream的优化器假设机器拓扑是分层的,并且可以被组织成多个级别,如下图所示。一个级别内的带宽是相同的,而跨级别的带宽是不同的。我们假设 k 级由 mk 个 k-1层组件构成 ,这些组件通过带宽为Bk的链路连接。在下图中,m2=2,m1=4。此外,我们定义m0为1。即 4 个 m0 构成一个 m1, 2个 m1 构成一个 m2。
层 0 就是绿色矩形,代表最底层的计算设备,比如GPU,4个GPU构成了一个层1(虚线矩形,代表一个服务器),2个层1构成了一个层2(就是下图全部模块)。
PipeDream的优化器从最低层到最高层逐步解决动态规划问题。直观地说,这个过程在服务器中找到最佳分区,然后使用这些分区在服务器之间最优地分割模型。
假设 A(j, m) 表示使用m台机器在第1层和第j层之间的最佳管道中,最慢阶段所用的时间。
我们算法的目标是找到 A(N,M) 和相应的划分。让T( i → j,m) 表示跨越层 i 到 j 的单级所用的时间,此时间在m台机器上复制。
其中:
max中的左项是在此阶段中所有层的总计算时间,右项是此阶段中所有层的总通信时间。
因为计算和通信可以重叠,所以不需要相加,直接取最大数值。
由1到j的由m个机器组成的最佳流水线可以是单个阶段复制m次,也可以由多个阶段组成。
当最佳管道包含多个阶段时,它可以被分解成一个最优的子管道(由从1到 i 的 由m − m′ 个机器组成)和后续的一个单独阶段(由i+1到j 的被 m’ 个机器复制组成)。因此,利用最优子问题的性质,我们得到
其中,max中:
第一项是第1层和第i层之间的最优子管道(由m-m’个机器组成)的最慢阶段所用的时间。
第二项是在层 i 和 i + 1 之间传递激活和梯度所用的时间。
第三项是最后单个阶段的时间(由 m’ 个数据并行的机器组成)。
我们具体看看如何计算,假设一个图逻辑如下:
+----------------+
+-----+ | +--------+
| +-------------> | k[m_prime] | | +-----+
| i | | | +--------->+ |
| +----+ +----------------+ | j |
+-----+ | +-------->+ |
| +----------------+ | +-----+
| | | |
+--------> | k[m-m_prime] +---------+
| |
+----------------+
在 (A [i] [k] [m-m_prime] [0], last_stage_time, output_transfer_time, input_transfer_time )之中选一个最大的:
因为传输和计算是可以重叠的,所以可以这样取最大数值。
最后得到的 A 就是动态规划优化的结果,其中每一个元素 A[i][j][m]
是个三元组 (min_pipeline_time, optimal_split, optimal_num_machines)
。 A[i][j][m]
表示节点 i 到 节点 j 之间的计算结果。三元组就是 (最小流水线时间,i 到 j 之间那个最佳分割点,最优机器数目)。
大致阶段如下图所示:
+----------------+
| i |
| |
| |
+--+------+------+
| |
| +----------+
A[i][k][m+m_prime][0] | |
| |
v v
+-----------------+-------+ +----+--------+
| k[m-m_prime] | | k[m_prime] |
| | | |
last_stage_time = compute_times[k+1][j] | | | |
+ (parameter_sizes[k+1][j]) | output_activation_sizes | | |
| | | |
| | | |
+-----------------+-------+ +-----+-------+
input_transfer_time | |
| +-----------+
| |
| |
v v
+------------+------+------+
| j |
| |
| |
| |
| output_activation_sizes |
| |
+------------------+-------+
output_transfer_time |
|
|
v
具体代码如下:
def compute_partitioning(compute_times, activation_sizes, parameter_sizes,
output_activation_sizes, all_predecessor_ids,
num_machines, num_machines_within_machine,
bandwidth, final_level=True):
# 初始化
A = []
for i in range(len(compute_times)): # 遍历所有节点
row_A = []
for j in range(len(compute_times[0])): # 所有后续节点(即第一个节点的所有后续节点)
row_row_A = []
for m in range(num_machines): # 机器数目
row_row_A.append((None, None, None))
row_A.append(row_row_A)
A.append(row_A)
# 得到计算时间
for i in range(len(compute_times)): # 遍历所有节点
for j in range(i, len(compute_times[0])): # 所有后续节点
cum_compute_time = compute_times[i][j] # i --> j 的计算时间
cum_activation_size = activation_sizes[i][j] # i --> j 的激活大小
cum_parameter_size = parameter_sizes[i][j] # i --> j 的参数大小
max_m = 1 if straight_pipeline else num_machines # 线性还是并行流水线
for m in range(max_m): # 遍历流水线下一阶段的机器
# 存储的数据大小
stashed_data_size = math.ceil((num_machines - (m+1)) / (m+1)) * \
(cum_activation_size + cum_parameter_size)
# memory_size 是用户传进来的参数,就是每个机器有效的内存
# use_memory_constraint 也是用户传进来的参数,就是使用的内存限制
if use_memory_constraint and stashed_data_size > memory_size:
continue
# 数据并行通讯时间依据参数尺寸,带宽,下一阶段机器数量计算
data_parallel_communication_time = (4 * m * cum_parameter_size) / (bandwidth * (m+1))
# 除以本阶段机器数量,如果本阶段机器多,当然就是分开计算了
data_parallel_communication_time /= num_machines_within_machine
if cum_compute_time is None:
# 需要计算下一阶段中,每个机器的计算时间,所以还要除以(m+1)
A[i][j][m] = (None, None, None) # 直接赋值
else:
# 三元组,分别是[(计算时间 + 通信时间), None,(m+1)],对应的意义是 min_pipeline_time, optimal_split, optimal_num_machines,就对应了前面的公式 2
A[i][j][m] = (sum([cum_compute_time,
data_parallel_communication_time]) / (m+1), None, (m+1))
# 需要得到最小计算时间
min_machines = 1
max_i = len(compute_times) if not final_level else 1
for i in range(max_i): # 遍历节点
for m in range(min_machines, num_machines): # 遍历下一阶段机器的可能选择
for j in range(i+1, len(compute_times[0])): # 遍历 i 的后续节点
(min_pipeline_time, optimal_split, optimal_num_machines) = A[i][j][m]
if use_fewer_machines and m > 0 and ( # 如果设置了用尽量少的机器,则如果小于min_pipeline_time,就设置新的 min_pipeline_time
min_pipeline_time is None or A[i][j][m-1][0] < min_pipeline_time):
(min_pipeline_time, optimal_split, optimal_num_machines) = A[i][j][m-1]
# 遍历 j 节点的前置机器 k,注意,j 是 i 的后续节点之一
# 就是在 i --> k --> j 之间找到一个计算时间最小的,其中A[i][k][m-m_prime][0]已经是一个最优子问题了
for k in all_predecessor_ids[j]:
# 如果k已经在之前计算过了,就跳过
if i > 0 and k in all_predecessor_ids[i-1]:
continue
# 设置质数
max_m_prime = 2 if straight_pipeline else (m+1)
for m_prime in range(1, max_m_prime): # prime就是看看如何分割
# 输入传输时间 input_transfer_time 使用 k 的输出激活尺寸计算
input_transfer_time = (2.0 * output_activation_sizes[k]) / \
(bandwidth * m_prime)
# 输出传输时间 output_transfer_time 使用 j 的输出激活尺寸计算
output_transfer_time = None
if j < len(output_activation_sizes) -1:
output_transfer_time = (2.0 *
output_activation_sizes[j]) / (bandwidth * m_prime)
# last_stage_time 设置为 k 到 j 的计算时间, compute_times[k+1] 就对应了k的输出
last_stage_time = compute_times[k+1][j]
if last_stage_time is None:
continue
# 设置为 k 到 j 的下一阶段参数尺寸
last_stage_parameter_size = parameter_sizes[k+1][j]
# 设置为 k 到 j 的存储数据尺寸
stashed_data_size = (activation_sizes[k+1][j]) + last_stage_parameter_size
# 依据机器数据计算
stashed_data_size *= math.ceil((num_machines - (m+1)) / m_prime)
# 超过机器内存就跳过
if use_memory_constraint and stashed_data_size > memory_size:
continue
# 加上传输时间,所以 last_stage_time 是 (k 到 j 的计算时间) + 传输时间
last_stage_time = sum([last_stage_time,
((4 * (m_prime - 1) *
last_stage_parameter_size) / (bandwidth * m_prime))])
last_stage_time /= m_prime
# 如果从i到k没有边,则跳过
if A[i][k][m-m_prime][0] is None:
continue
# 如果i到k已经有计算时间,则选一个较大的
pipeline_time = max(A[i][k][m-m_prime][0], last_stage_time)
if activation_compression_ratio is not None: # 如果压缩
# 在(A[i][k][m-m_prime][0], last_stage_time, output_transfer_time, input_transfer_time 之中选一个最大的)
input_transfer_time /= activation_compression_ratio
# output_transfer_time 也压缩
if output_transfer_time is not None:
output_transfer_time /= activation_compression_ratio
# 选一个大的
pipeline_time = max(pipeline_time, input_transfer_time)
if output_transfer_time is not None:
pipeline_time = max(pipeline_time, output_transfer_time)
# 如果比min_pipeline_time小,则设定 min_pipeline_time,为了下一次循环
if min_pipeline_time is None or min_pipeline_time > pipeline_time:
optimal_split = (k, m-m_prime) # 选一个优化分割点
optimal_num_machines = m_prime
min_pipeline_time = pipeline_time
# 设置
A[i][j][m] = (min_pipeline_time, optimal_split, optimal_num_machines)
return A
all_As 就是动态规划的结果,示例如下:
all_As = {list: 2}
0 = {list: 100}
000 = {list: 99}
00 = {list: 5} [(0.0070220000000000005, None, 1), (0.1689894, None, 2), (0.14943257777777777, None, 3), (0.1258643, None, 4), (0.107310576, None, 5)]
01 = {list: 5} [(0.012285, None, 1), (0.0070220000000000005, (0, 0), 1), (0.0865995, (0, 0), 2), (0.07639255555555556, (0, 0), 3), (0.06429175000000001, (0, 0), 4)]
02 = {list: 5} [(0.012558, None, 1), (0.0070220000000000005, (0, 0), 1), (0.0070220000000000005, (1, 1), 1), (0.0070220000000000005, (1, 1), 2), (0.0070220000000000005, (1, 1), 3)]
03 = {list: 5} [(0.021096, None, 1), (0.012285, (1, 0), 1), (0.008538, (2, 1), 1), (0.008538, (2, 2), 1), (0.008538, (2, 3), 1)]
......
__len__ = {int} 100
1 = {list: 100}
000 = {list: 99}
00 = {list: 5} [(0.107310576, None, 1), (0.080131832, None, 2), (0.05930489777777778, None, 3), (0.046685052000000005, None, 4), (0.03840710336000001, None, 5)]
01 = {list: 5} [(0.06429175000000001, None, 1), (0.072057299, None, 2), (0.05690740466666667, None, 3), (0.0460065055, None, 4), (0.03840166136, None, 5)]
02 = {list: 5} [(0.0070220000000000005, None, 1), (0.043422424, None, 2), (0.037817488, None, 3), (0.031689068, None, 4), (0.026947711359999998, None, 5)]
03 = {list: 5} [(0.008538, None, 1), (0.0419991328, (2, 0), 1), (0.043422424, (2, 1), 1), (0.0396227304, None, 4), (0.033697556608, None, 5)]
......
__len__ = {int} 100
__len__ = {int} 2
我们接下来要分析代码作者两个相似名字变量之间的区别。
activation_sizes :某个节点所有前置节点的activation_size 之和。
for predecessor in all_predecessors:
states[i].compute_time += ((predecessor.forward_compute_time +
predecessor.backward_compute_time) / 1000.0)
states[i].activation_size += predecessor.activation_size
states[i].parameter_size += predecessor.parameter_size
用来计算stashed数据大小,用来看看是否超过了节点配置的内存额度。
stashed_data_size = (activation_sizes[k+1][j]) + last_stage_parameter_size
stashed_data_size *= math.ceil((num_machines - (m+1)) / m_prime)
if use_memory_constraint and stashed_data_size > memory_size:
continue
output_activation_sizes : 某个节点所有增强反链的activation_size之和。
for i in range(len(states)):
for antichain_node in states[i].antichain:
states[i].output_activation_size += gr.nodes[antichain_node].activation_size
用来计算输出传播时间和输入传播时间。
input_transfer_time = (2.0 * output_activation_sizes[k]) / \
(bandwidth * m_prime)
output_transfer_time = None
if j < len(output_activation_sizes) -1:
output_transfer_time = (2.0 *
output_activation_sizes[j]) / (bandwidth * m_prime)
前面计算分区只是得到了一个动态规划优化结果,需要在analyze_partitioning之中进行分析划分之后,赋予到各个层(stage)。
main函数接下来与计算分区相关的逻辑如下:
具体代码如下:
def main(all_num_machines, profile_filename, network_bandwidths, memory_size,
straight_pipeline, use_memory_constraint, use_fewer_machines,
activation_compression_ratio, output_directory,
print_configuration=True, verbose=False):
gr = graph.Graph.from_str(open(profile_filename, 'r').read())
# Zero out all metadata associated with inputs in graph, since the optimizer
# shouldn't really get a choice with where to place the input (should always
# be in the first stage).
# 排除干扰,因为input必然在第一层,没必要让优化器再来选择把输入放在哪里,所以先去除,后续会再加上。
sources = gr.sources() # 对图的输入进行处理
nodes_to_remove = OrderedDict()
for source in sources:
if source.node_desc.startswith("Input"): # 只处理input
source.forward_compute_time = 0.0
source.backward_compute_time = 0.0
source.activation_size = 0.0
source.parameter_size = 0.0
nodes_to_remove[source] = []
for out_node in gr.edges[source.node_id]:
nodes_to_remove[source].append(out_node) # 记录这些删除source对应了哪些out节点,因为后续还要处理
gr.remove_node(source) # 在图中移除这些input source
# Remove all unneeded sinks that are not used, makes code generation and
# optimization easier.
sinks = gr.sinks() # 对图的输出进行处理,移除没有用到的输出
for sink in sinks:
if sink.node_desc.startswith("__getitem__"):
gr.remove_node(sink)
antichain_gr = gr.antichain_dag() # 得到反链DAG
states = antichain_gr.topological_sort() # 拓扑排序,得到一个排序好的节点列表
###########################################################################
# 计算阶段
###########################################################################
states_indices = {} # 为每个状态设置index
for i in range(len(states)):
states_indices[states[i]] = i
##################################### 运行时如下
#states_indices = {dict: 99}
# antichain_0 -- ['node4'] = {int} 0
# antichain_1 -- ['node5'] = {int} 1
# antichain_2 -- ['node6'] = {int} 2
# antichain_3 -- ['node7'] = {int} 3
# antichain_4 -- ['node8'] = {int} 4
# ......
# 给每个状态计算出输出激活值大小,具体是通过遍历其反链(增强反链),可以认为就是其必要前序节点给自己的输出
for i in range(len(states)):
for antichain_node in states[i].antichain:
states[i].output_activation_size += gr.nodes[antichain_node].activation_size
# 给每个状态计算其信息,比如计算时间,激活大小,参数大小等等,都是通过前置节点完成的
for i in range(len(states)):
antichain = states[i].antichain
all_predecessors = gr.all_predecessors(antichain)
states[i].compute_time = 0.0
states[i].activation_size = 0.0
states[i].parameter_size = 0.0
for predecessor in all_predecessors: # 计算所有前置节点的信息
states[i].compute_time += ((predecessor.forward_compute_time +
predecessor.backward_compute_time) / 1000.0)
states[i].activation_size += predecessor.activation_size
states[i].parameter_size += predecessor.parameter_size
gr.reset()
# 得到总体输出大小 & 所有前置节点id,后面计算分区时候需要
output_activation_sizes = [state.output_activation_size for state in states]
all_predecessor_ids = [[states_indices[predecessor] for predecessor in
antichain_gr.predecessors(states[i].node_id)]
for i in range(len(states))]
##################################### 运行时如下
# output_activation_sizes = {list: 99}
# 00 = {float} 6291456.0
# 01 = {float} 12582912.0
# 02 = {float} 12582912.0
# 03 = {float} 6553600.0
# .....
# all_predecessor_ids = {list: 99}
# 00 = {list: 0} []
# 01 = {list: 1} [0]
# 02 = {list: 2} [0, 1]
# 03 = {list: 3} [0, 1, 2]
# 04 = {list: 4} [0, 1, 2, 3]
# 05 = {list: 5} [2, 3, 4, 0, 1]
# 06 = {list: 6} [2, 3, 4, 0, 1, 5]
# 07 = {list: 7} [6, 2, 3, 4, 0, 1, 5]
# ......
compute_times = [] # 初始化计算时间
activation_sizes = [] # 初始化激活值大小
parameter_sizes = [] # 初始化参数值大小
for i in range(len(states)+1): # 具体计算每一个节点的信息,去除他之前节点的影响
compute_times_row = []
activation_sizes_row = []
parameter_sizes_row = []
for j in range(len(states)): # 去除之前的节点
if i == 0: # 列表中第一个节点
compute_times_row.append(states[j].compute_time) # i 到 j 的计算时间
activation_sizes_row.append(states[j].activation_size)
parameter_sizes_row.append(states[j].parameter_size)
else: # 列表中后续节点
if j > (i-1):
compute_times_row.append(states[j].compute_time -
states[i-1].compute_time) # i 到 j 的计算时间
activation_sizes_row.append(states[j].activation_size -
states[i-1].activation_size)
parameter_sizes_row.append(states[j].parameter_size -
states[i-1].parameter_size)
else:
compute_times_row.append(None)
activation_sizes_row.append(None)
parameter_sizes_row.append(None)
compute_times.append(compute_times_row) # 依据profile估计出系统内部的计算时间,compute_times_row 是 i 节点到 后续节点(i+1, i+2, ...)的计算时间,下面类似
activation_sizes.append(activation_sizes_row) # 依据profile估计出系统内部的激活值大小
parameter_sizes.append(parameter_sizes_row) # 依据profile估计出系统内部的参数大小
##################################### 运行时如下
# compute_times = {list: 100}
# 000 = {list: 99} [0.0070220000000000005, 0.012285, 0.012558, 0.021096000000,...
# 001 = {list: 99} [None, 0.005263, 0.005535999999999999, 0.014074000000000003, ...
# 002 = {list: 99} [None, None, 0.00027299999999999894, 0.008811000000000003, ...
# 003 = {list: 99} [None, None, None, 0.008538000000000004, 0.008538, ...
# 004 = {list: 99} [None, None, None, None, -3.469446951953614e-18, 0.000191999999...
counter = 1
all_As = []
num_machines_in_machine = 1 #第一个节点就是1
# all_num_machines, network_bandwidths 是用户在输入中指定
# 遍历机器集&网络带宽组合。流水线可以是straight(数目为1)或者并行(数目为num_machines)
for num_machines, network_bandwidth in zip(all_num_machines, network_bandwidths):
print("Solving optimization problem with %d machines with inter-machine bandwidth of %.2f GB/s" % (num_machines, network_bandwidth / 10**9))
import numpy as np
print(np.array(compute_times))
# 依据目前的信息,以及机器数量,网络带宽等计算分区
A = compute_partitioning(compute_times, activation_sizes, parameter_sizes,
output_activation_sizes, all_predecessor_ids,
num_machines, num_machines_in_machine,
network_bandwidth,
final_level=(counter==len(network_bandwidths)))
num_machines_in_machine = num_machines # 因为计算完了,所以设置为本阶段的机器数目
for i in range(len(compute_times)): # 遍历机器
for j in range(len(compute_times[0])): # 后续机器
compute_times[i][j] = A[i][j][-1][0] # 记录计算时间(本阶段最后一个机器的计算时间)
counter += 1
all_As.append(A) # 添加逻辑关系,就是里面包括了不同阶段的优化逻辑
print(np.array(compute_times))
###########################################################################
# 我们从这里继续分析
###########################################################################
# 分析阶段
# 在 analyze_partitioning 内部做了具体分析
# 这里最重要的是对 gr.all_predecessors 做设置,就是设置 gr 之中每个node的stage_id,这样就是利用stage_id把初始流水线重新划分
splits = [(0, len(states))] # 如何分割,states是反链DAG的结果,所以 splits 初始化时候就只有一个二元组元素:最初的划分 (0, len(states))
i = len(all_As) - 1 # all_As 就是动态规划得到的优化结果
while i >= 0: # 遍历优化的出来的各个逻辑关系
print("======================================")
print("Level %d" % (i+1))
print("======================================")
new_splits = []
stage_id = 0 # 在后续的convert_graph_to_model.py 之中会使用到
for (start, end) in splits: # 在分割中遍历,splits会逐步更新
# 依据新的splits中的二元组重新计算
partial_splits = \
analyze_partitioning(all_As[i], states, start, end,
network_bandwidths[i], all_num_machines[i],
activation_compression_ratio,
print_configuration, verbose)
start_point = start # 起始点
for split in partial_splits: # 遍历分析得出的节点
new_splits.append((start_point, split)) # 添加一个新的二元祖
if i == 0:
predecessors = gr.all_predecessors(states[split-1].antichain)
for predecessor in predecessors:
if predecessor.stage_id is None:
predecessor.set_stage_id(stage_id) # 设置所在阶段
start_point = split # 下一个阶段
stage_id += 1 # 增加所在阶段
new_splits.append((start_point, end)) # 添加一个新的二元祖
if i == 0:
predecessors = gr.all_predecessors(states[end-1].antichain)
for predecessor in predecessors:
if predecessor.stage_id is None:
predecessor.set_stage_id(stage_id) # 设置所在阶段
stage_id += 1 # 增加所在阶段
print("Total number of stages: %d" % stage_id)
splits = new_splits # 加入新的分割
i -= 1
# 以下是为了把图写到文件之中。后续convert_graph_to_model.py会把这个文件转换成模型
for source in nodes_to_remove: # 之前移除了input节点,现在需要加回到图中
for out_node in nodes_to_remove[source]: # input对应的哪些输出
source.stage_id = 0
gr.add_edge(source, out_node)
if output_directory is not None:
total_num_machines = 1
for num_machines in all_num_machines:
total_num_machines *= num_machines
gr.to_dot(os.path.join(output_directory, "gpus=%d" % total_num_machines))
gr_str = str(gr)
with open(os.path.join(output_directory, "gpus=%d.txt" % total_num_machines), 'w') as f:
f.write(gr_str)
# 以下是为了做分析对比
# 计算数据并行需要的时间,以便接下来做比较,这个时间要比动态规划时间长。
total_time = states[-1].compute_time # 最后一个阶段的计算时间,是没有经过优化的最初计算时间
total_parameter_size = states[-1].parameter_size
data_parallel_total_time = total_time # 先赋值为最后一阶段的计算时间
num_machines_in_machine = 1 # 本阶段的机器数目
# 遍历流水线上各个阶段,因为没有优化,所以就是严格按照用户原始配置的流水线阶段来逐一计算
for (num_machines, network_bandwidth) in zip(all_num_machines, network_bandwidths):
# 计算传输时间。num_machines是下一阶段流水线机器数目,所以带宽需要乘以这个数字
data_parallel_communication_time = (
(4 * (num_machines - 1) * total_parameter_size) /
(network_bandwidth * num_machines)) / num_machines_in_machine
# 总时间需要加上传输时间
data_parallel_total_time = sum(
[data_parallel_total_time, data_parallel_communication_time]) / num_machines
# 下个迭代中,本阶段的机器数目需要设置为num_machines
num_machines_in_machine = num_machines
# 这个是用动态规划算法得出来的优化时间
pipeline_parallel_total_time = A[0][len(states)-1][num_machines-1][0]
# 可以看到用户需要注意哪些数据
if verbose:
print()
print("Time taken by single-stage pipeline:", total_time)
print("Time per stage in pipeline:", pipeline_parallel_total_time)
print("Throughput increase (compared to single machine):",
total_time / pipeline_parallel_total_time)
dp_str = ",".join([str(elem) for elem in all_num_machines])
print(("[Note that single-machine and (%s)-machine DP might not fit "
"given memory constraints]") % dp_str)
print("Throughput increase of (%s)-machine DP compared to single "
"machine:" % dp_str, total_time / data_parallel_total_time)
print("Throughput increase (compared to (%s)-machine DP):" % dp_str,
data_parallel_total_time / pipeline_parallel_total_time)
return pipeline_parallel_total_time, data_parallel_total_time
分析阶段具体可以参见下面注释。
def analyze_partitioning(A, states, start, end, network_bandwidth, num_machines,
activation_compression_ratio, print_configuration, verbose):
# start,end 是本组节点的起始点,终止点
metadata = A[start][end-1][num_machines-1] # 这是个三元组 (min_pipeline_time, optimal_split, optimal_num_machines)
next_split = metadata[1] # metadata[1] 是 optimal_split,即 (k, m-m_prime)
remaining_machines_left = num_machines
splits = []
replication_factors = []
prev_split = end - 1 # 前一个分割点
while next_split is not None: #是否继续分割
num_machines_used = metadata[2] # optimal_num_machines
if verbose:
print("-------------------------------------")
print("Number of machines used: %d..." % num_machines_used)
print("Split between layers %d and %d..." % (next_split[0], next_split[0] + 1))
print("Split before antichain %s..." % (states[next_split[0]+1].antichain))
splits.append(next_split[0]+1) # 得到了 k + 1,这是关键点,因为最后返回的是splits
compute_time = states[prev_split-1].compute_time - \
states[next_split[0]].compute_time
parameter_size = states[prev_split-1].parameter_size - \
states[next_split[0]].parameter_size
dp_communication_time = (4 * (num_machines_used - 1) * parameter_size) \
/ (network_bandwidth * num_machines_used)
pp_communication_time_input = ( # 下个阶段的数据输入时间
2.0 * states[next_split[0]].output_activation_size *
(1.0 / float(num_machines_used))) / network_bandwidth
pp_communication_time_output = ( # 上个阶段的数据输出时间
2.0 * states[prev_split-1].output_activation_size *
(1.0 / float(num_machines_used))) / network_bandwidth
# 如果需要压缩,就进行压缩
if activation_compression_ratio is not None:
pp_communication_time_input /= activation_compression_ratio
pp_communication_time_output /= activation_compression_ratio
if activation_compression_ratio is None:
pp_communication_time_input = 0.0
pp_communication_time_output = 0.0
compute_time /= num_machines_used # 本阶段计算时间
dp_communication_time /= num_machines_used # 数据并行时间
if verbose:
print(("Compute time = %f, Data-parallel communication time = %f, "
"Pipeline-parallel communication time = %f...") % (
compute_time, dp_communication_time,
max(pp_communication_time_input, pp_communication_time_output)))
prev_split = splits[-1] # 设定新的前一分割点
# next_split 格式是 (k, m-m_prime),就是 optimal_split 的格式
# A[i][j][m] 格式是 (min_pipeline_time, optimal_split, optimal_num_machines)
metadata = A[start][next_split[0]][next_split[1]]
next_split = metadata[1] # 设定新的下一次分割点,就是 optimal_split
replication_factors.append(num_machines_used) # 每个阶段的 replication factor
remaining_machines_left -= num_machines_used # 剩余机器
if verbose:
print("-------------------------------------")
print("Number of machines used: %d..." % metadata[2])
#
num_machines_used = metadata[2]
remaining_machines_left -= num_machines_used # 剩余的机器
compute_time = states[prev_split-1].compute_time
parameter_size = states[prev_split-1].parameter_size
dp_communication_time = ((4 * (num_machines_used - 1) * parameter_size) /
(network_bandwidth * num_machines_used))
compute_time /= num_machines_used # 计算时间
dp_communication_time /= num_machines_used # 数据并行通信时间
if verbose:
print("Compute time = %f, Data-parallel communication time = %f..." %
(compute_time, dp_communication_time))
print("-------------------------------------")
if print_configuration:
print("Number of machines in budget not used: %d..." %
remaining_machines_left)
print()
print("(Split start, split end) / compute time taken per stage "
"/ replication factor per stage:")
# 下面就是打印 (Split start, split end) / compute time taken per stage / replication factor per stage
prev_split = start
splits.reverse() #
splits.append(end)
replication_factors.append(num_machines_used)
replication_factors.reverse()
for i in range(len(splits)):
time = 0.0
if prev_split > 0:
time = states[splits[i]-1].compute_time - states[prev_split-1].compute_time
else:
time = states[splits[i]-1].compute_time
if print_configuration:
print((prev_split, splits[i]), time, replication_factors[i])
prev_split = splits[i]
if print_configuration:
print()
return splits[:-1] # 最后一个不返回
我们还是用样例进行说明。
这里是从后面进行分割,举例分析如下,这里设定了总机器数目为10:
回忆在计算分区之中,A[i][j][m] = (min_pipeline_time, optimal_split, optimal_num_machines),optimal_split = (k, m-m_prime)
是一个本阶段优化点。
所以在本函数之中,start = 0, end = 99,所以 metadata 为A[0][99][10]
,即 (0.01903199999999998, (95, 8), 1),next_split = (95, 8),prev_split = end - 1 = 98。
next_split 就是下一个分割点,splits 是目前的分割序列。
第一轮while循环:
因为next_split = (95, 8),所以 splits = append(next_split[0]+1) = [96],因此计算 states[prev_split-1] - states[next_split[0]] = state[97] - state[95]。这样把0~99分成了 0 ~95 和 96 ~ 99。
然后 prev_split = 96,去找A[ 0 ] [ 95] [8] 得到 meta = (0.019031999999999993, (78, 7), 1),next_split = (78, 7)。
所以下一轮从78这个分割点开始分割。
第二轮while循环:
因为next_split = (78, 7),所以 splits = [96, 79],这就是新的分割序列。,因此计算 states[96-1] - states[next_split[0]] = state[96] - state[78]。这样就使用 splits = [96, 79] 把0~99分成了 0 ~78,79 ~ 95 和 96 ~ 99。
然后 prev_split =79,去找A[ 0 ] [ 78 ] [ 7 ] 得到 meta = (0.011081, (48, 6), 1),next_split = (48, 6)。
所以下一轮从 48 这个分割点开始分割,以此类推。
while循环之后,得到 splits = [96, 79, 49, 15, 12, 7, 5, 3, 1]。
于是下面代码需要把顺序调整过来。
prev_split = start
splits.reverse()
splits.append(end)
replication_factors.append(num_machines_used)
replication_factors.reverse()
得到:splits = { 1,3,5,7,12,15,49,79,96 }。然后加上 end = 99。
最后返回 splits[:-1],即返回 { 1,3,5,7,12,15,49,79,96 },去掉刚刚添加的end。
而依据 { 1,3,5,7,12,15,49,79,96 } 得到的最终分割序列 是 [(0, 1), (1, 3), (3, 5), (5, 7), (7, 12), (12, 15), (15, 49), (49, 79), (79, 96), (96, 99)],这个列表会在后续"设定stage"之中会用到。
目前我们得到了一个理想分割序列,但是事情没有结束,我们回忆一下分区算法的目的:依据profile结果确定所有层的运行时间,然后使用动态规划对模型进行划分,将模型划分为不同的stage,以及得到每个stage的replication数。
所以,分析的最终目的是给模型的每一个子层分配一个stage,如果某些子层属于同一个stage,这些子层最终就被分配到同一个worker(节点)上执行。
因为这里涉及到多个子网,所以我们依然用实例来分析。
如果分成了两个子网,假设:
all_num_machines = [5,5]
network_bandwidths = [800000000, 1000000000]
初始化 splits = [0,99]。
第一轮 while 中,i = 1,
对于 splits 结果[(0, 99)] 遍历,每一段应用analyze_partitioning,得到 partial_splits 为 [3, 6, 30, 75, 99]。
最后,splits 更新为:[(0, 3), (3, 6), (6, 30), (30, 75), (75, 99)]。
此时不会设置stage_id。
第二轮 while 中,i = 0,
对于第一轮的 splits 结果 [(0, 3), (3, 6), (6, 30), (30, 75), (75, 99)] 进行遍历,对于这里的每一段也应用 analyze_partitioning,比如对 (0,3) 应用analyze_partitioning,对 (3,6) 应用 analyze_partitioning,对(6,30) 也应用 analyze_partitioning,…,最后得到新的 partial_splits 为 [1, 2, 3, 4, 5, 6, 8, 10, 13, 28, 30, 45, 49, 51, 75, 79, 96, 99]。
最后,splits 更新为:[(0, 1), (1, 2), (2, 3), (3, 4), (4, 5), (5, 6), (6, 8), (8, 10), (10, 13), (13, 28), (28, 30), (30, 45), (45, 49), (49, 51), (51, 75), (75, 79), (79, 96), (96, 99)]。
这个列表就是理想分割序列。
在此轮中,得到了partial_splits之后,会遍历 for split in partial_splits:
然后对于每一个 split,利用
states[split-1].antichain
获取其增强反链的所有前置节点,给这些节点打上 split 对应的 stage_id。
回忆一下增强反链的意义:
所以,对于 split = 1,1 - 1 = 0,于是就得到 states[0].antichain
,就是 ‘node4’,那么 ‘node4’ 自己被打上了一个stage_id=0,说明 ‘node4’ 被分到了一个 “与stage_id=0 所对应” 的 worker 节点上训练。
如果有疑问,我们回忆一下state如何构建,就是有序的 “节点组合”。
antichain_gr = gr.antichain_dag()
states = antichain_gr.topological_sort()
具体如下。
states = {list: 99}
00 = {AntichainNode} antichain_0 -- ['node4'] # states[0].antichain
01 = {AntichainNode} antichain_1 -- ['node5']
02 = {AntichainNode} antichain_2 -- ['node6']
03 = {AntichainNode} antichain_3 -- ['node7']
04 = {AntichainNode} antichain_4 -- ['node8']
05 = {AntichainNode} antichain_5 -- ['node8', 'node10']
06 = {AntichainNode} antichain_7 -- ['node8', 'node11']
07 = {AntichainNode} antichain_10 -- ['node8', 'node12']
08 = {AntichainNode} antichain_6 -- ['node14']
09 = {AntichainNode} antichain_8 -- ['node14', 'node15']
10 = {AntichainNode} antichain_11 -- ['node14', 'node16']
11 = {AntichainNode} antichain_13 -- ['node14', 'node17']
12 = {AntichainNode} antichain_9 -- ['node19']
13 = {AntichainNode} antichain_12 -- ['node20', 'node23']
14 = {AntichainNode} antichain_18 -- ['node23', 'node20', 'node26']
15 = {AntichainNode} antichain_17 -- ['node23', 'node20', 'node24']
16 = {AntichainNode} antichain_32 -- ['node23', 'node20', 'node28']
17 = {AntichainNode} antichain_31 -- ['node23', 'node20', 'node26', 'node24']
18 = {AntichainNode} antichain_63 -- ['node23', 'node20', 'node26', 'node28']
19 = {AntichainNode} antichain_33 -- ['node20', 'node26', 'node29']
20 = {AntichainNode} antichain_16 -- ['node20', 'node43', 'node23']
21 = {AntichainNode} antichain_30 -- ['node23', 'node20', 'node43', 'node26']
22 = {AntichainNode} antichain_29 -- ['node23', 'node20', 'node43', 'node24']
23 = {AntichainNode} antichain_59 -- ['node23', 'node20', 'node43', 'node28']
设定stage 具体代码如下:
splits = [(0, len(states))]
i = len(all_As) - 1
while i >= 0:
new_splits = []
stage_id = 0
for (start, end) in splits:
partial_splits = \
analyze_partitioning(all_As[i], states, start, end,
network_bandwidths[i], all_num_machines[i],
activation_compression_ratio,
print_configuration, verbose)
start_point = start
for split in partial_splits: # 遍历这个偏序列表
new_splits.append((start_point, split))
if i == 0: # 最终的while
# 针对每个节点,找到每个节点的所有反链
predecessors = gr.all_predecessors(states[split-1].antichain)
for predecessor in predecessors:
if predecessor.stage_id is None:
predecessor.set_stage_id(stage_id) # 打上stage id
start_point = split
stage_id += 1
new_splits.append((start_point, end))
if i == 0: # 最终的while
predecessors = gr.all_predecessors(states[end-1].antichain)
for predecessor in predecessors:
if predecessor.stage_id is None:
predecessor.set_stage_id(stage_id) # 打上stage id
stage_id += 1
splits = new_splits
i -= 1
我们总结一下计算分区和分析分区所做的工作:
反链DAG图已经被分割成若干状态(states),每个状态很重要的一个属性是其增强反链。states 就是对增强反链进行拓扑排序之后的结果,按照这个顺序进行训练是符合逻辑的。
compute_partitioning 是使用动态规划算法对于这些 states 状态得出一个最优化结果,但是这个计算分区只是得到了一个动态规划优化结果,需要在analyze_partitioning之中进行分析划分之后,赋予到各个层(stage)。
analyze_partitioning 是利用动态规划算法的最优化结果来做具体分区,排序后得到了一个偏序结果,就是理想分割序列。
依据 analyze_partitioning 的结果,给模型的每一个子层分配一个stage,如果某些子层属于同一个stage,这些子层最终就被分配到同一个worker(节点)上执行。
输出文件如下(摘录部分),可以看到,关键之处在于给每一个节点加上了stage,具体如何使用我们将在下一篇进行分析。比如:
stage_id=0 对应的是 node4。
stage_id=1 对应的是 node5,node6。
stage_id=2 对应的是 node7。
stage_id=3 对应的是 node8,node10,node11,node12。
…
具体如下:
node4 -- Embedding(32320, 1024, padding_idx=0) -- forward_compute_time=0.073, backward_compute_time=6.949, activation_size=6291456.0, parameter_size=132382720.000 -- stage_id=0
node5 -- EmuBidirLSTM( (bidir): LSTM(1024, 1024, bidirectional=True) (layer1): LSTM(1024, 1024) (layer2): LSTM(1024, 1024)) -- forward_compute_time=5.247, backward_compute_time=0.016, activation_size=12582912.0, parameter_size=67174400.000 -- stage_id=1
node6 -- Dropout(p=0.2) -- forward_compute_time=0.077, backward_compute_time=0.196, activation_size=12582912.0, parameter_size=0.000 -- stage_id=1
node7 -- LSTM(2048, 1024) -- forward_compute_time=3.190, backward_compute_time=5.348, activation_size=6553600.0, parameter_size=50364416.000 -- stage_id=2
node8 -- __getitem__(0) -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=6291456.0, parameter_size=0.000 -- stage_id=3
node10 -- Dropout(p=0.2) -- forward_compute_time=0.064, backward_compute_time=0.128, activation_size=6291456.0, parameter_size=0.000 -- stage_id=3
node11 -- LSTM(1024, 1024) -- forward_compute_time=2.491, backward_compute_time=4.203, activation_size=6553600.0, parameter_size=33587200.000 -- stage_id=3
node12 -- __getitem__(0) -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=6291456.0, parameter_size=0.000 -- stage_id=3
node14 -- Add -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=6291456.0, parameter_size=0.000 -- stage_id=4
node15 -- Dropout(p=0.2) -- forward_compute_time=0.059, backward_compute_time=0.121, activation_size=6291456.0, parameter_size=0.000 -- stage_id=4
node16 -- LSTM(1024, 1024) -- forward_compute_time=2.492, backward_compute_time=4.201, activation_size=6553600.0, parameter_size=33587200.000 -- stage_id=4
node17 -- __getitem__(0) -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=6291456.0, parameter_size=0.000 -- stage_id=5
node19 -- Add -- forward_compute_time=0.000, backward_compute_time=0.000, activation_size=6291456.0, parameter_size=0.000 -- stage_id=5
node1 -- node4
node4 -- node5
node2 -- node5
node5 -- node6
node6 -- node7
node7 -- node8
node8 -- node10
node10 -- node11
node11 -- node12
node12 -- node14
node8 -- node14
node14 -- node15
node15 -- node16
node16 -- node17
node17 -- node19
★★★★★★关于生活和技术的思考★★★★★★
微信公众账号:罗西的思考
如果您想及时得到个人撰写文章的消息推送,或者想看看个人推荐的技术资料,敬请关注。
[源码解析] 深度学习流水线并行之PipeDream(1)— Profile阶段