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
PaddlePaddle
Paddle
提交
88d79dfe
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
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
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
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
未验证
提交
88d79dfe
编写于
5月 03, 2018
作者:
W
Wu Yi
提交者:
GitHub
5月 03, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #10292 from typhoonzero/fix_grpc_server_ready_condition
Fix grpc server ready condition
上级
6084af47
6422c0e4
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
66 addition
and
22 deletion
+66
-22
paddle/fluid/operators/detail/grpc_server.cc
paddle/fluid/operators/detail/grpc_server.cc
+11
-0
paddle/fluid/operators/detail/grpc_server.h
paddle/fluid/operators/detail/grpc_server.h
+6
-1
paddle/fluid/operators/listen_and_serv_op.cc
paddle/fluid/operators/listen_and_serv_op.cc
+23
-12
paddle/fluid/operators/listen_and_serv_op.h
paddle/fluid/operators/listen_and_serv_op.h
+12
-2
paddle/fluid/operators/send_recv_op_test.cc
paddle/fluid/operators/send_recv_op_test.cc
+13
-6
python/paddle/fluid/tests/unittests/test_dist_train.py
python/paddle/fluid/tests/unittests/test_dist_train.py
+1
-1
未找到文件。
paddle/fluid/operators/detail/grpc_server.cc
浏览文件 @
88d79dfe
...
@@ -211,6 +211,11 @@ void AsyncGRPCServer::WaitClientGet(int count) {
...
@@ -211,6 +211,11 @@ void AsyncGRPCServer::WaitClientGet(int count) {
}
}
}
}
void
AsyncGRPCServer
::
WaitServerReady
()
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
this
->
mutex_ready_
);
condition_ready_
.
wait
(
lock
,
[
=
]
{
return
this
->
ready_
==
1
;
});
}
void
AsyncGRPCServer
::
RunSyncUpdate
()
{
void
AsyncGRPCServer
::
RunSyncUpdate
()
{
::
grpc
::
ServerBuilder
builder
;
::
grpc
::
ServerBuilder
builder
;
builder
.
AddListeningPort
(
address_
,
::
grpc
::
InsecureServerCredentials
(),
builder
.
AddListeningPort
(
address_
,
::
grpc
::
InsecureServerCredentials
(),
...
@@ -244,6 +249,12 @@ void AsyncGRPCServer::RunSyncUpdate() {
...
@@ -244,6 +249,12 @@ void AsyncGRPCServer::RunSyncUpdate() {
t_prefetch_
.
reset
(
new
std
::
thread
(
t_prefetch_
.
reset
(
new
std
::
thread
(
std
::
bind
(
&
AsyncGRPCServer
::
HandleRequest
,
this
,
cq_prefetch_
.
get
(),
std
::
bind
(
&
AsyncGRPCServer
::
HandleRequest
,
this
,
cq_prefetch_
.
get
(),
"cq_prefetch"
,
prefetch_register
)));
"cq_prefetch"
,
prefetch_register
)));
{
std
::
lock_guard
<
std
::
mutex
>
lock
(
this
->
mutex_ready_
);
ready_
=
1
;
}
condition_ready_
.
notify_all
();
// wait server
// wait server
server_
->
Wait
();
server_
->
Wait
();
t_send_
->
join
();
t_send_
->
join
();
...
...
paddle/fluid/operators/detail/grpc_server.h
浏览文件 @
88d79dfe
...
@@ -45,8 +45,9 @@ class RequestBase;
...
@@ -45,8 +45,9 @@ class RequestBase;
class
AsyncGRPCServer
final
{
class
AsyncGRPCServer
final
{
public:
public:
explicit
AsyncGRPCServer
(
const
std
::
string
&
address
,
bool
sync_mode
)
explicit
AsyncGRPCServer
(
const
std
::
string
&
address
,
bool
sync_mode
)
:
address_
(
address
),
sync_mode_
(
sync_mode
)
{}
:
address_
(
address
),
sync_mode_
(
sync_mode
)
,
ready_
(
0
)
{}
void
WaitServerReady
();
void
RunSyncUpdate
();
void
RunSyncUpdate
();
// functions to sync server barrier status.
// functions to sync server barrier status.
...
@@ -118,6 +119,10 @@ class AsyncGRPCServer final {
...
@@ -118,6 +119,10 @@ class AsyncGRPCServer final {
framework
::
ProgramDesc
*
program_
;
framework
::
ProgramDesc
*
program_
;
framework
::
Executor
*
executor_
;
framework
::
Executor
*
executor_
;
int
selected_port_
;
int
selected_port_
;
std
::
mutex
mutex_ready_
;
std
::
condition_variable
condition_ready_
;
int
ready_
;
};
};
};
// namespace detail
};
// namespace detail
...
...
paddle/fluid/operators/listen_and_serv_op.cc
浏览文件 @
88d79dfe
...
@@ -66,12 +66,7 @@ static void ParallelExecuteBlocks(
...
@@ -66,12 +66,7 @@ static void ParallelExecuteBlocks(
for
(
size_t
i
=
0
;
i
<
fs
.
size
();
++
i
)
fs
[
i
].
wait
();
for
(
size_t
i
=
0
;
i
<
fs
.
size
();
++
i
)
fs
[
i
].
wait
();
}
}
static
void
SavePort
(
std
::
shared_ptr
<
detail
::
AsyncGRPCServer
>
rpc_service
)
{
std
::
atomic_int
ListenAndServOp
::
selected_port_
{
0
};
std
::
ofstream
port_file
;
port_file
.
open
(
"/tmp/paddle.selected_port"
);
port_file
<<
rpc_service
->
GetSelectedPort
();
port_file
.
close
();
}
ListenAndServOp
::
ListenAndServOp
(
const
std
::
string
&
type
,
ListenAndServOp
::
ListenAndServOp
(
const
std
::
string
&
type
,
const
framework
::
VariableNameMap
&
inputs
,
const
framework
::
VariableNameMap
&
inputs
,
...
@@ -79,15 +74,27 @@ ListenAndServOp::ListenAndServOp(const std::string &type,
...
@@ -79,15 +74,27 @@ ListenAndServOp::ListenAndServOp(const std::string &type,
const
framework
::
AttributeMap
&
attrs
)
const
framework
::
AttributeMap
&
attrs
)
:
OperatorBase
(
type
,
inputs
,
outputs
,
attrs
)
{}
:
OperatorBase
(
type
,
inputs
,
outputs
,
attrs
)
{}
int
ListenAndServOp
::
GetSelectedPort
()
const
{
return
rpc_service_
->
GetSelectedPort
();
}
void
ListenAndServOp
::
Stop
()
{
void
ListenAndServOp
::
Stop
()
{
rpc_service_
->
Push
(
LISTEN_TERMINATE_MESSAGE
);
rpc_service_
->
Push
(
LISTEN_TERMINATE_MESSAGE
);
server_thread_
->
join
();
server_thread_
->
join
();
}
}
void
ListenAndServOp
::
SavePort
(
const
std
::
string
&
file_path
)
const
{
// NOTE: default write file to /tmp/paddle.selected_port
selected_port_
=
rpc_service_
->
GetSelectedPort
();
std
::
ofstream
port_file
;
port_file
.
open
(
file_path
);
port_file
<<
selected_port_
.
load
();
port_file
.
close
();
VLOG
(
4
)
<<
"selected port written to "
<<
file_path
;
}
void
ListenAndServOp
::
WaitServerReady
()
{
while
(
selected_port_
.
load
()
==
0
)
{
}
}
void
ListenAndServOp
::
RunSyncLoop
(
framework
::
Executor
*
executor
,
void
ListenAndServOp
::
RunSyncLoop
(
framework
::
Executor
*
executor
,
framework
::
ProgramDesc
*
program
,
framework
::
ProgramDesc
*
program
,
framework
::
Scope
*
recv_scope
,
framework
::
Scope
*
recv_scope
,
...
@@ -318,9 +325,13 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
...
@@ -318,9 +325,13 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
// start the server listening after all member initialized.
// start the server listening after all member initialized.
server_thread_
.
reset
(
new
std
::
thread
(
RunServer
,
rpc_service_
));
server_thread_
.
reset
(
new
std
::
thread
(
RunServer
,
rpc_service_
));
VLOG
(
3
)
<<
"wait server thread to become ready..."
;
VLOG
(
3
)
<<
"wait server thread to become ready..."
;
sleep
(
5
);
rpc_service_
->
WaitServerReady
();
// Write to a file of server selected port for python use.
// Write to a file of server selected port for python use.
SavePort
(
rpc_service_
);
std
::
string
file_path
=
string
::
Sprintf
(
"/tmp/paddle.%d.selected_port"
,
static_cast
<
int
>
(
::
getpid
()));
SavePort
(
file_path
);
if
(
sync_mode
)
{
if
(
sync_mode
)
{
RunSyncLoop
(
&
executor
,
program
,
&
recv_scope
,
prefetch_block
);
RunSyncLoop
(
&
executor
,
program
,
&
recv_scope
,
prefetch_block
);
}
else
{
}
else
{
...
...
paddle/fluid/operators/listen_and_serv_op.h
浏览文件 @
88d79dfe
...
@@ -15,6 +15,7 @@ limitations under the License. */
...
@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once
#pragma once
#include <stdint.h>
#include <stdint.h>
#include <atomic>
#include <ostream>
#include <ostream>
#include <string>
#include <string>
...
@@ -39,8 +40,6 @@ class ListenAndServOp : public framework::OperatorBase {
...
@@ -39,8 +40,6 @@ class ListenAndServOp : public framework::OperatorBase {
const
framework
::
VariableNameMap
&
outputs
,
const
framework
::
VariableNameMap
&
outputs
,
const
framework
::
AttributeMap
&
attrs
);
const
framework
::
AttributeMap
&
attrs
);
int
GetSelectedPort
()
const
;
void
RunSyncLoop
(
framework
::
Executor
*
executor
,
void
RunSyncLoop
(
framework
::
Executor
*
executor
,
framework
::
ProgramDesc
*
program
,
framework
::
ProgramDesc
*
program
,
framework
::
Scope
*
recv_scope
,
framework
::
Scope
*
recv_scope
,
...
@@ -49,14 +48,25 @@ class ListenAndServOp : public framework::OperatorBase {
...
@@ -49,14 +48,25 @@ class ListenAndServOp : public framework::OperatorBase {
void
RunAsyncLoop
(
framework
::
Executor
*
executor
,
void
RunAsyncLoop
(
framework
::
Executor
*
executor
,
framework
::
ProgramDesc
*
program
)
const
;
framework
::
ProgramDesc
*
program
)
const
;
void
SavePort
(
const
std
::
string
&
file_path
=
"/tmp/paddle.selected_port"
)
const
;
void
WaitServerReady
();
int
GetSelectedPort
()
{
return
selected_port_
;
}
void
Stop
()
override
;
void
Stop
()
override
;
void
RunImpl
(
const
framework
::
Scope
&
scope
,
void
RunImpl
(
const
framework
::
Scope
&
scope
,
const
platform
::
Place
&
dev_place
)
const
override
;
const
platform
::
Place
&
dev_place
)
const
override
;
static
void
ResetPort
()
{
selected_port_
=
0
;
}
protected:
protected:
mutable
std
::
shared_ptr
<
detail
::
AsyncGRPCServer
>
rpc_service_
;
mutable
std
::
shared_ptr
<
detail
::
AsyncGRPCServer
>
rpc_service_
;
mutable
std
::
shared_ptr
<
std
::
thread
>
server_thread_
;
mutable
std
::
shared_ptr
<
std
::
thread
>
server_thread_
;
// FIXME(wuyi): it's static so that the operator can be cloned.
static
std
::
atomic_int
selected_port_
;
};
};
}
// namespace operators
}
// namespace operators
...
...
paddle/fluid/operators/send_recv_op_test.cc
浏览文件 @
88d79dfe
...
@@ -116,6 +116,7 @@ void AddOp(const std::string &type, const f::VariableNameMap &inputs,
...
@@ -116,6 +116,7 @@ void AddOp(const std::string &type, const f::VariableNameMap &inputs,
void
StartServerNet
(
bool
is_sparse
,
std
::
atomic
<
bool
>
*
initialized
)
{
void
StartServerNet
(
bool
is_sparse
,
std
::
atomic
<
bool
>
*
initialized
)
{
f
::
Scope
scope
;
f
::
Scope
scope
;
p
::
CPUPlace
place
;
p
::
CPUPlace
place
;
VLOG
(
4
)
<<
"before init tensor"
;
if
(
is_sparse
)
{
if
(
is_sparse
)
{
InitSelectedRowsInScope
(
place
,
&
scope
);
InitSelectedRowsInScope
(
place
,
&
scope
);
}
else
{
}
else
{
...
@@ -137,6 +138,7 @@ void StartServerNet(bool is_sparse, std::atomic<bool> *initialized) {
...
@@ -137,6 +138,7 @@ void StartServerNet(bool is_sparse, std::atomic<bool> *initialized) {
attrs
.
insert
({
"PrefetchBlock"
,
prefetch_block
});
attrs
.
insert
({
"PrefetchBlock"
,
prefetch_block
});
attrs
.
insert
({
"grad_to_block_id"
,
std
::
vector
<
std
::
string
>
({
""
})});
attrs
.
insert
({
"grad_to_block_id"
,
std
::
vector
<
std
::
string
>
({
""
})});
attrs
.
insert
({
"sync_mode"
,
true
});
attrs
.
insert
({
"sync_mode"
,
true
});
VLOG
(
4
)
<<
"before init op"
;
listen_and_serv_op
=
listen_and_serv_op
=
f
::
OpRegistry
::
CreateOp
(
"listen_and_serv"
,
{{
"X"
,
{
"x1"
}}},
{},
attrs
);
f
::
OpRegistry
::
CreateOp
(
"listen_and_serv"
,
{{
"X"
,
{
"x1"
}}},
{},
attrs
);
*
initialized
=
true
;
*
initialized
=
true
;
...
@@ -149,7 +151,9 @@ TEST(SendRecvOp, CPUDense) {
...
@@ -149,7 +151,9 @@ TEST(SendRecvOp, CPUDense) {
std
::
thread
server_thread
(
StartServerNet
,
false
,
&
initialized
);
std
::
thread
server_thread
(
StartServerNet
,
false
,
&
initialized
);
while
(
!
initialized
)
{
while
(
!
initialized
)
{
}
}
sleep
(
5
);
// wait server to start
static_cast
<
paddle
::
operators
::
ListenAndServOp
*>
(
listen_and_serv_op
.
get
())
->
WaitServerReady
();
// local net
// local net
f
::
Scope
scope
;
f
::
Scope
scope
;
p
::
CPUPlace
place
;
p
::
CPUPlace
place
;
...
@@ -185,6 +189,7 @@ TEST(SendRecvOp, CPUDense) {
...
@@ -185,6 +189,7 @@ TEST(SendRecvOp, CPUDense) {
listen_and_serv_op
->
Stop
();
listen_and_serv_op
->
Stop
();
server_thread
.
join
();
server_thread
.
join
();
listen_and_serv_op
.
reset
(
nullptr
);
listen_and_serv_op
.
reset
(
nullptr
);
paddle
::
operators
::
ListenAndServOp
::
ResetPort
();
}
}
TEST
(
SendRecvOp
,
CPUSparse
)
{
TEST
(
SendRecvOp
,
CPUSparse
)
{
...
@@ -193,7 +198,12 @@ TEST(SendRecvOp, CPUSparse) {
...
@@ -193,7 +198,12 @@ TEST(SendRecvOp, CPUSparse) {
std
::
thread
server_thread
(
StartServerNet
,
true
,
&
initialized
);
std
::
thread
server_thread
(
StartServerNet
,
true
,
&
initialized
);
while
(
!
initialized
)
{
while
(
!
initialized
)
{
}
}
sleep
(
5
);
// wait server to start
auto
*
listen_and_serv_op_ptr
=
static_cast
<
paddle
::
operators
::
ListenAndServOp
*>
(
listen_and_serv_op
.
get
());
ASSERT_TRUE
(
listen_and_serv_op_ptr
!=
nullptr
);
listen_and_serv_op_ptr
->
WaitServerReady
();
// local net
// local net
f
::
Scope
scope
;
f
::
Scope
scope
;
p
::
CPUPlace
place
;
p
::
CPUPlace
place
;
...
@@ -201,10 +211,6 @@ TEST(SendRecvOp, CPUSparse) {
...
@@ -201,10 +211,6 @@ TEST(SendRecvOp, CPUSparse) {
InitSelectedRowsInScope
(
place
,
&
scope
);
InitSelectedRowsInScope
(
place
,
&
scope
);
scope
.
Var
(
"RPC_CLIENT_VAR"
);
scope
.
Var
(
"RPC_CLIENT_VAR"
);
f
::
AttributeMap
attrs
;
f
::
AttributeMap
attrs
;
auto
*
listen_and_serv_op_ptr
=
static_cast
<
paddle
::
operators
::
ListenAndServOp
*>
(
listen_and_serv_op
.
get
());
ASSERT_TRUE
(
listen_and_serv_op_ptr
!=
nullptr
);
selected_port
=
listen_and_serv_op_ptr
->
GetSelectedPort
();
selected_port
=
listen_and_serv_op_ptr
->
GetSelectedPort
();
std
::
string
endpoint
=
paddle
::
string
::
Sprintf
(
"127.0.0.1:%d"
,
selected_port
);
std
::
string
endpoint
=
paddle
::
string
::
Sprintf
(
"127.0.0.1:%d"
,
selected_port
);
attrs
.
insert
({
"endpoints"
,
std
::
vector
<
std
::
string
>
({
endpoint
})});
attrs
.
insert
({
"endpoints"
,
std
::
vector
<
std
::
string
>
({
endpoint
})});
...
@@ -236,4 +242,5 @@ TEST(SendRecvOp, CPUSparse) {
...
@@ -236,4 +242,5 @@ TEST(SendRecvOp, CPUSparse) {
listen_and_serv_op
->
Stop
();
listen_and_serv_op
->
Stop
();
server_thread
.
join
();
server_thread
.
join
();
listen_and_serv_op
.
reset
();
listen_and_serv_op
.
reset
();
paddle
::
operators
::
ListenAndServOp
::
ResetPort
();
}
}
python/paddle/fluid/tests/unittests/test_dist_train.py
浏览文件 @
88d79dfe
...
@@ -34,7 +34,7 @@ class TestSendOp(unittest.TestCase):
...
@@ -34,7 +34,7 @@ class TestSendOp(unittest.TestCase):
p
.
start
()
p
.
start
()
time
.
sleep
(
10
)
time
.
sleep
(
10
)
with
open
(
"/tmp/paddle.
selected_port"
,
"r"
)
as
fn
:
with
open
(
"/tmp/paddle.
%d.selected_port"
%
p
.
pid
,
"r"
)
as
fn
:
selected_port
=
int
(
fn
.
readlines
()[
0
])
selected_port
=
int
(
fn
.
readlines
()[
0
])
self
.
init_client
(
place
,
selected_port
)
self
.
init_client
(
place
,
selected_port
)
...
...
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