Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
一WILLPOWER一
Paddle
提交
faebadd9
P
Paddle
项目概览
一WILLPOWER一
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
列表
看板
标记
里程碑
合并请求
0
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
1
Issue
1
列表
看板
标记
里程碑
合并请求
0
合并请求
0
Pages
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
提交
Issue看板
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
faebadd9
编写于
5月 02, 2018
作者:
Q
Qiao Longfei
提交者:
GitHub
5月 02, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #10228 from jacquesqiao/use-multi-thread-todo-update
Use multi thread to do update
上级
ff99d941
d86626df
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
61 addition
and
36 deletion
+61
-36
paddle/fluid/operators/detail/grpc_server.cc
paddle/fluid/operators/detail/grpc_server.cc
+11
-6
paddle/fluid/operators/listen_and_serv_op.cc
paddle/fluid/operators/listen_and_serv_op.cc
+46
-26
paddle/fluid/operators/listen_and_serv_op.h
paddle/fluid/operators/listen_and_serv_op.h
+1
-3
python/paddle/fluid/layers/io.py
python/paddle/fluid/layers/io.py
+3
-1
未找到文件。
paddle/fluid/operators/detail/grpc_server.cc
浏览文件 @
faebadd9
...
...
@@ -82,7 +82,9 @@ class RequestSend final : public RequestBase {
virtual
std
::
string
GetReqName
()
{
return
request_
->
Varname
();
}
virtual
void
Process
()
{
queue_
->
Push
(
std
::
make_pair
(
request_
->
Varname
(),
request_
));
std
::
string
var_name
=
GetReqName
();
VLOG
(
3
)
<<
"RequestSend "
<<
var_name
;
queue_
->
Push
(
std
::
make_pair
(
var_name
,
request_
));
sendrecv
::
VoidMessage
reply
;
responder_
.
Finish
(
reply
,
::
grpc
::
Status
::
OK
,
this
);
...
...
@@ -106,7 +108,7 @@ class RequestGet final : public RequestBase {
responder_
(
&
ctx_
),
scope_
(
scope
),
queue_
(
queue
)
{
int
method_id
=
static_cast
<
int
>
(
detail
::
GrpcMethod
::
kGetVariable
);
auto
method_id
=
static_cast
<
int
>
(
detail
::
GrpcMethod
::
kGetVariable
);
service_
->
RequestAsyncUnary
(
method_id
,
&
ctx_
,
&
request_
,
&
responder_
,
cq_
,
cq_
,
this
);
}
...
...
@@ -118,6 +120,7 @@ class RequestGet final : public RequestBase {
virtual
void
Process
()
{
// proc request.
std
::
string
var_name
=
request_
.
varname
();
VLOG
(
3
)
<<
"RequestGet "
<<
var_name
;
auto
*
var
=
scope_
->
FindVar
(
var_name
);
::
grpc
::
ByteBuffer
reply
;
...
...
@@ -176,7 +179,7 @@ class RequestPrefetch final : public RequestBase {
::
grpc
::
ByteBuffer
reply
;
std
::
string
var_name
=
request_
->
OutVarname
();
VLOG
(
3
)
<<
"
prefetch var
"
<<
var_name
;
VLOG
(
3
)
<<
"
RequestPrefetch
"
<<
var_name
;
auto
var_desc
=
program_
->
Block
(
0
).
FindVar
(
var_name
);
framework
::
Scope
*
local_scope
=
&
scope_
->
NewScope
();
auto
*
var
=
local_scope
->
FindVar
(
var_name
);
...
...
@@ -307,18 +310,20 @@ void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq,
bool
ok
=
false
;
while
(
true
)
{
VLOG
(
3
)
<<
"HandleRequest for "
<<
cq_name
<<
" w
hile in
"
;
VLOG
(
3
)
<<
"HandleRequest for "
<<
cq_name
<<
" w
ait Next
"
;
if
(
!
cq
->
Next
(
&
tag
,
&
ok
))
{
LOG
(
INFO
)
<<
cq_name
<<
" CompletionQueue shutdown!"
;
break
;
}
VLOG
(
3
)
<<
"HandleRequest for "
<<
cq_name
<<
"
while after
Next"
;
VLOG
(
3
)
<<
"HandleRequest for "
<<
cq_name
<<
"
get
Next"
;
PADDLE_ENFORCE
(
tag
);
if
(
sync_mode_
)
{
// FIXME(typhoonzero): de-couple the barriers with recv_op
if
(
!
is_shut_down_
&&
cq_name
==
"cq_get"
)
WaitCond
(
1
);
if
(
!
is_shut_down_
&&
cq_name
==
"cq_send"
)
WaitCond
(
0
);
VLOG
(
3
)
<<
"HandleRequest for "
<<
cq_name
<<
" after WaitCond"
;
}
RequestBase
*
base
=
reinterpret_cast
<
RequestBase
*>
(
tag
);
...
...
@@ -336,9 +341,9 @@ void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq,
switch
(
base
->
Status
())
{
case
PROCESS
:
{
VLOG
(
4
)
<<
cq_name
<<
" PROCESS status:"
<<
base
->
Status
();
TryToRegisterNewOne
();
base
->
Process
();
VLOG
(
4
)
<<
cq_name
<<
" PROCESS status:"
<<
base
->
Status
();
break
;
}
case
FINISH
:
{
...
...
paddle/fluid/operators/listen_and_serv_op.cc
浏览文件 @
faebadd9
...
...
@@ -45,20 +45,6 @@ static void split(const std::string &str, char sep,
}
}
static
void
AsyncExecuteBlock
(
framework
::
Executor
*
executor
,
framework
::
ExecutorPrepareContext
*
prepared
,
framework
::
Scope
*
scope
)
{
std
::
future
<
void
>
future
=
framework
::
Async
([
&
executor
,
&
prepared
,
&
scope
]()
{
try
{
executor
->
RunPreparedContext
(
prepared
,
scope
,
false
,
false
);
}
catch
(
std
::
exception
&
e
)
{
LOG
(
ERROR
)
<<
"run sub program error "
<<
e
.
what
();
}
});
// TODO(qiao) maybe we can remove this
future
.
wait
();
}
static
void
ParallelExecuteBlocks
(
const
std
::
vector
<
size_t
>
&
parallel_blkids
,
framework
::
Executor
*
executor
,
const
std
::
vector
<
std
::
shared_ptr
<
framework
::
ExecutorPrepareContext
>>
...
...
@@ -201,14 +187,40 @@ void ListenAndServOp::RunSyncLoop(framework::Executor *executor,
}
// while(true)
}
static
void
AsyncUpdateThread
(
const
std
::
string
&
var_name
,
const
bool
&
exit_flag
,
const
std
::
shared_ptr
<
detail
::
ReceivedQueue
>
&
queue
,
framework
::
Executor
*
executor
,
framework
::
ExecutorPrepareContext
*
prepared
)
{
VLOG
(
3
)
<<
"update thread for "
<<
var_name
<<
" started"
;
while
(
!
exit_flag
)
{
const
detail
::
ReceivedMessage
v
=
queue
->
Pop
();
auto
recv_var_name
=
v
.
first
;
auto
var
=
v
.
second
->
GetVar
();
if
(
var
==
nullptr
)
{
LOG
(
ERROR
)
<<
"Can not find server side var: "
<<
recv_var_name
;
PADDLE_THROW
(
"Can not find server side var"
);
}
auto
fs
=
framework
::
Async
([
var_name
,
&
executor
,
&
v
,
prepared
]
{
try
{
executor
->
RunPreparedContext
(
prepared
,
v
.
second
->
GetMutableLocalScope
(),
false
,
false
);
}
catch
(
std
::
exception
&
e
)
{
LOG
(
ERROR
)
<<
"run sub program error "
<<
e
.
what
();
}
});
fs
.
wait
();
}
}
void
ListenAndServOp
::
RunAsyncLoop
(
framework
::
Executor
*
executor
,
framework
::
ProgramDesc
*
program
,
framework
::
Scope
*
recv_scope
,
framework
::
BlockDesc
*
prefetch_block
)
const
{
framework
::
ProgramDesc
*
program
)
const
{
VLOG
(
3
)
<<
"RunAsyncLoop in"
;
// grad name to block id
std
::
unordered_map
<
std
::
string
,
int32_t
>
grad_to_block_id
;
std
::
unordered_map
<
int32_t
,
std
::
string
>
id_to_grad
;
std
::
unordered_map
<
std
::
string
,
std
::
shared_ptr
<
detail
::
ReceivedQueue
>>
grad_to_queue
;
auto
grad_to_block_id_str
=
Attr
<
std
::
vector
<
std
::
string
>>
(
"grad_to_block_id"
);
...
...
@@ -220,6 +232,7 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor,
PADDLE_ENFORCE_EQ
(
grad_to_block_id
.
count
(
pieces
[
0
]),
0
);
int
block_id
=
std
::
stoi
(
pieces
[
1
]);
grad_to_block_id
[
pieces
[
0
]]
=
block_id
;
grad_to_queue
[
pieces
[
0
]]
=
std
::
make_shared
<
detail
::
ReceivedQueue
>
();
id_to_grad
[
block_id
]
=
pieces
[
0
];
}
size_t
num_blocks
=
program
->
Size
();
...
...
@@ -238,8 +251,21 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor,
grad_to_prepared_ctx
[
id_to_grad
[
block_list
[
i
]]]
=
optimize_prepared
[
i
];
}
VLOG
(
3
)
<<
"RunAsyncLoop into while"
;
bool
exit_flag
=
false
;
VLOG
(
3
)
<<
"start async optimize threads"
;
std
::
vector
<
std
::
future
<
void
>>
fs
;
for
(
auto
iter
=
grad_to_queue
.
begin
();
iter
!=
grad_to_queue
.
end
();
iter
++
)
{
std
::
string
grad_name
=
iter
->
first
;
VLOG
(
3
)
<<
"create async update thread for "
<<
grad_name
;
fs
.
push_back
(
framework
::
AsyncIO
([
grad_name
,
&
exit_flag
,
&
executor
,
&
grad_to_queue
,
&
grad_to_prepared_ctx
]()
{
AsyncUpdateThread
(
grad_name
,
exit_flag
,
grad_to_queue
[
grad_name
],
executor
,
grad_to_prepared_ctx
[
grad_name
].
get
());
}));
}
VLOG
(
3
)
<<
"RunAsyncLoop into while"
;
while
(
!
exit_flag
)
{
const
detail
::
ReceivedMessage
v
=
rpc_service_
->
Get
();
auto
recv_var_name
=
v
.
first
;
...
...
@@ -249,13 +275,7 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor,
break
;
}
else
{
VLOG
(
3
)
<<
"received grad: "
<<
recv_var_name
;
auto
var
=
v
.
second
->
GetVar
();
if
(
var
==
nullptr
)
{
LOG
(
ERROR
)
<<
"Can not find server side var: "
<<
recv_var_name
;
PADDLE_THROW
(
"Can not find server side var"
);
}
AsyncExecuteBlock
(
executor
,
grad_to_prepared_ctx
[
recv_var_name
].
get
(),
v
.
second
->
GetMutableLocalScope
());
grad_to_queue
[
recv_var_name
]
->
Push
(
v
);
}
if
(
exit_flag
)
{
...
...
@@ -304,7 +324,7 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
if
(
sync_mode
)
{
RunSyncLoop
(
&
executor
,
program
,
&
recv_scope
,
prefetch_block
);
}
else
{
RunAsyncLoop
(
&
executor
,
program
,
&
recv_scope
,
prefetch_block
);
RunAsyncLoop
(
&
executor
,
program
);
}
}
...
...
paddle/fluid/operators/listen_and_serv_op.h
浏览文件 @
faebadd9
...
...
@@ -47,9 +47,7 @@ class ListenAndServOp : public framework::OperatorBase {
framework
::
BlockDesc
*
prefetch_block
)
const
;
void
RunAsyncLoop
(
framework
::
Executor
*
executor
,
framework
::
ProgramDesc
*
program
,
framework
::
Scope
*
recv_scope
,
framework
::
BlockDesc
*
prefetch_block
)
const
;
framework
::
ProgramDesc
*
program
)
const
;
void
Stop
()
override
;
...
...
python/paddle/fluid/layers/io.py
浏览文件 @
faebadd9
...
...
@@ -168,7 +168,9 @@ class ListenAndServ(object):
'endpoint'
:
self
.
endpoint
,
'Fanin'
:
self
.
fan_in
,
'OptimizeBlock'
:
current_block
,
'PrefetchBlock'
:
empty_block
'PrefetchBlock'
:
empty_block
,
'sync_mode'
:
True
,
# did not support async now in layers
'grad_to_block_id'
:
[
""
]
})
...
...
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录