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
TonyTonyFun
Paddle
提交
e0c83974
P
Paddle
项目概览
TonyTonyFun
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
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
0
Issue
0
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
e0c83974
编写于
11月 07, 2018
作者:
Q
Qiao Longfei
提交者:
GitHub
11月 07, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #14257 from jacquesqiao/optimize-pserver-profiler-thread-pool
clean rpc server profiler
上级
ffc86615
3b8dd9eb
变更
9
隐藏空白更改
内联
并排
Showing
9 changed file
with
8 addition
and
56 deletion
+8
-56
paddle/fluid/operators/distributed/grpc_variable_response.cc
paddle/fluid/operators/distributed/grpc_variable_response.cc
+2
-2
paddle/fluid/operators/distributed/request_handler_impl.cc
paddle/fluid/operators/distributed/request_handler_impl.cc
+0
-1
paddle/fluid/operators/distributed/rpc_server.cc
paddle/fluid/operators/distributed/rpc_server.cc
+0
-32
paddle/fluid/operators/distributed/rpc_server.h
paddle/fluid/operators/distributed/rpc_server.h
+0
-18
paddle/fluid/operators/distributed/variable_response.cc
paddle/fluid/operators/distributed/variable_response.cc
+3
-0
paddle/fluid/operators/distributed/variable_response.h
paddle/fluid/operators/distributed/variable_response.h
+2
-0
paddle/fluid/operators/listen_and_serv_op.cc
paddle/fluid/operators/listen_and_serv_op.cc
+0
-1
paddle/fluid/platform/profiler.cc
paddle/fluid/platform/profiler.cc
+1
-1
python/paddle/fluid/__init__.py
python/paddle/fluid/__init__.py
+0
-1
未找到文件。
paddle/fluid/operators/distributed/grpc_variable_response.cc
浏览文件 @
e0c83974
...
...
@@ -286,10 +286,10 @@ int GRPCVariableResponse::Parse(Source* source) {
platform
::
EnableProfiler
(
platform
::
ProfilerState
::
kCPU
);
}
else
if
(
profiling
==
platform
::
kDisableProfiler
&&
platform
::
IsProfileEnabled
())
{
// TODO(panyx0718): Should we allow to customize file dir.
platform
::
DisableProfiler
(
platform
::
EventSortingKey
::
kDefault
,
string
::
Sprintf
(
"/tmp/profile_ps_%lld"
,
listener_id
));
string
::
Sprintf
(
"%s_%lld"
,
FLAGS_rpc_server_profile_path
,
listener_id
));
}
break
;
}
...
...
paddle/fluid/operators/distributed/request_handler_impl.cc
浏览文件 @
e0c83974
...
...
@@ -51,7 +51,6 @@ bool RequestSendHandler::Handle(const std::string& varname,
// Async
if
(
!
sync_mode_
)
{
VLOG
(
3
)
<<
"async process var: "
<<
varname
;
rpc_server_
->
Profiler
().
OneStep
();
try
{
executor_
->
RunPreparedContext
((
*
grad_to_prepared_ctx_
)[
varname
].
get
(),
scope
);
...
...
paddle/fluid/operators/distributed/rpc_server.cc
浏览文件 @
e0c83974
...
...
@@ -20,42 +20,10 @@
#include "paddle/fluid/operators/distributed/rpc_server.h"
#include "paddle/fluid/platform/profiler.h"
DEFINE_int32
(
rpc_server_profile_period
,
0
,
"the period of listen_and_serv to do profile"
);
DEFINE_string
(
rpc_server_profile_path
,
"/dev/null"
,
"the profile log file path"
);
namespace
paddle
{
namespace
operators
{
namespace
distributed
{
RPCServerProfiler
::
RPCServerProfiler
(
int
profile_period
,
const
std
::
string
&
profile_log_path
)
:
profile_period_
(
profile_period
),
profile_log_path_
(
profile_log_path
)
{
step_
=
0
;
}
void
RPCServerProfiler
::
OneStep
()
{
PADDLE_ENFORCE_LE
(
step_
,
profile_period_
,
"step_ should not be larger then "
"profile_period_"
);
if
(
profile_period_
<=
0
)
{
return
;
}
if
(
step_
==
0
)
{
auto
pf_state
=
paddle
::
platform
::
ProfilerState
::
kCPU
;
paddle
::
platform
::
EnableProfiler
(
pf_state
);
}
if
(
step_
==
profile_period_
)
{
paddle
::
platform
::
DisableProfiler
(
paddle
::
platform
::
EventSortingKey
::
kTotal
,
profile_log_path_
);
step_
=
0
;
}
else
{
step_
++
;
}
}
void
RPCServer
::
ShutDown
()
{
LOG
(
INFO
)
<<
"RPCServer ShutDown "
;
ShutDownImpl
();
...
...
paddle/fluid/operators/distributed/rpc_server.h
浏览文件 @
e0c83974
...
...
@@ -23,30 +23,14 @@
#include "paddle/fluid/operators/distributed/request_handler.h"
DECLARE_int32
(
rpc_server_profile_period
);
DECLARE_string
(
rpc_server_profile_path
);
namespace
paddle
{
namespace
operators
{
namespace
distributed
{
class
RPCServerProfiler
{
public:
RPCServerProfiler
(
int
profile_period
,
const
std
::
string
&
profile_log_path
);
void
OneStep
();
private:
const
int
profile_period_
;
std
::
string
profile_log_path_
;
int
step_
;
};
class
RPCServer
{
public:
explicit
RPCServer
(
const
std
::
string
&
address
,
int
client_num
)
:
cur_cond_
(
0
),
profiler_
(
FLAGS_rpc_server_profile_period
,
FLAGS_rpc_server_profile_path
),
bind_address_
(
address
),
exit_flag_
(
false
),
selected_port_
(
0
),
...
...
@@ -86,7 +70,6 @@ class RPCServer {
void
Complete
();
void
ResetBarrierCounter
();
RPCServerProfiler
&
Profiler
()
{
return
profiler_
;
}
bool
NeedResetAllVars
();
...
...
@@ -101,7 +84,6 @@ class RPCServer {
std
::
unordered_map
<
std
::
string
,
int
>
rpc_cond_map_
;
std
::
atomic
<
int
>
cur_cond_
;
std
::
condition_variable
rpc_cond_
;
RPCServerProfiler
profiler_
;
protected:
std
::
string
bind_address_
;
...
...
paddle/fluid/operators/distributed/variable_response.cc
浏览文件 @
e0c83974
...
...
@@ -16,6 +16,9 @@
#include <vector>
#include "paddle/fluid/operators/distributed/sendrecvop_utils.h"
DEFINE_string
(
rpc_server_profile_path
,
"./profile_ps"
,
"the profile log file path"
);
namespace
paddle
{
namespace
operators
{
namespace
distributed
{
...
...
paddle/fluid/operators/distributed/variable_response.h
浏览文件 @
e0c83974
...
...
@@ -27,6 +27,8 @@
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/operators/distributed/send_recv.pb.h"
DECLARE_string
(
rpc_server_profile_path
);
namespace
paddle
{
namespace
operators
{
namespace
distributed
{
...
...
paddle/fluid/operators/listen_and_serv_op.cc
浏览文件 @
e0c83974
...
...
@@ -134,7 +134,6 @@ void ListenAndServOp::RunSyncLoop(
rpc_service_
->
ResetBarrierCounter
();
while
(
true
)
{
rpc_service_
->
Profiler
().
OneStep
();
// Get from multiple trainers, we don't care about the order in which
// the gradients arrives, just add suffix 0~n and merge the gradient.
rpc_service_
->
SetCond
(
distributed
::
kRequestSend
);
...
...
paddle/fluid/platform/profiler.cc
浏览文件 @
e0c83974
...
...
@@ -226,7 +226,7 @@ RecordBlock::~RecordBlock() {
void
EnableProfiler
(
ProfilerState
state
)
{
PADDLE_ENFORCE
(
state
!=
ProfilerState
::
kDisabled
,
"Can't en
bale prof
ling, since the input state is "
,
"Can't en
able profi
ling, since the input state is "
,
"ProfilerState::kDisabled"
);
std
::
lock_guard
<
std
::
mutex
>
l
(
profiler_mu
);
...
...
python/paddle/fluid/__init__.py
浏览文件 @
e0c83974
...
...
@@ -118,7 +118,6 @@ def __bootstrap__():
]
if
core
.
is_compiled_with_dist
():
read_env_flags
.
append
(
'rpc_deadline'
)
read_env_flags
.
append
(
'rpc_server_profile_period'
)
read_env_flags
.
append
(
'rpc_server_profile_path'
)
read_env_flags
.
append
(
'enable_rpc_profiler'
)
read_env_flags
.
append
(
'rpc_send_thread_num'
)
...
...
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