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
45363069
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
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
体验新版 GitCode,发现更多精彩内容 >>
提交
45363069
编写于
4月 01, 2018
作者:
Q
qiaolongfei
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix prefetch hang problem, add some more logs
上级
fdecae5f
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
24 addition
and
21 deletion
+24
-21
paddle/fluid/operators/detail/grpc_client.cc
paddle/fluid/operators/detail/grpc_client.cc
+9
-7
paddle/fluid/operators/detail/grpc_server.cc
paddle/fluid/operators/detail/grpc_server.cc
+11
-2
paddle/fluid/operators/detail/grpc_service.h
paddle/fluid/operators/detail/grpc_service.h
+2
-2
paddle/fluid/operators/listen_and_serv_op.cc
paddle/fluid/operators/listen_and_serv_op.cc
+2
-10
未找到文件。
paddle/fluid/operators/detail/grpc_client.cc
浏览文件 @
45363069
...
...
@@ -12,8 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "grpc_client.h"
#include <sys/time.h>
#include "paddle/fluid/operators/detail/grpc_client.h"
#include <limits>
#include "paddle/fluid/framework/threadpool.h"
namespace
paddle
{
...
...
@@ -52,7 +54,7 @@ bool RPCClient::AsyncSendVariable(const std::string& ep,
auto
call
=
s
->
stub_g_
.
PrepareUnaryCall
(
s
->
context_
.
get
(),
"/sendrecv.SendRecvService/SendVariable"
,
req
,
&
cq_
);
call
->
StartCall
();
call
->
Finish
(
&
s
->
reply_
,
&
s
->
status_
,
(
void
*
)
s
);
call
->
Finish
(
&
s
->
reply_
,
&
s
->
status_
,
static_cast
<
void
*>
(
s
)
);
});
req_count_
++
;
...
...
@@ -109,7 +111,7 @@ bool RPCClient::AsyncGetVariable(const std::string& ep,
auto
call
=
s
->
stub_g_
.
PrepareUnaryCall
(
s
->
context_
.
get
(),
"/sendrecv.SendRecvService/GetVariable"
,
buf
,
&
cq_
);
call
->
StartCall
();
call
->
Finish
(
&
s
->
reply_
,
&
s
->
status_
,
(
void
*
)
s
);
call
->
Finish
(
&
s
->
reply_
,
&
s
->
status_
,
static_cast
<
void
*>
(
s
)
);
});
req_count_
++
;
...
...
@@ -153,7 +155,7 @@ bool RPCClient::AsyncPrefetchVariable(const std::string& ep,
s
->
context_
.
get
(),
"/sendrecv.SendRecvService/PrefetchVariable"
,
req
,
&
cq_
);
call
->
StartCall
();
call
->
Finish
(
&
s
->
reply_
,
&
s
->
status_
,
(
void
*
)
s
);
call
->
Finish
(
&
s
->
reply_
,
&
s
->
status_
,
static_cast
<
void
*>
(
s
)
);
});
req_count_
++
;
...
...
@@ -169,7 +171,7 @@ void RPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) {
sendrecv
::
VariableMessage
req
;
req
.
set_varname
(
BATCH_BARRIER_MESSAGE
);
auto
rpc
=
s
->
stub_
->
AsyncSendVariable
(
s
->
context_
.
get
(),
req
,
&
cq_
);
rpc
->
Finish
(
&
s
->
reply_
,
&
s
->
status_
,
(
void
*
)
s
);
rpc
->
Finish
(
&
s
->
reply_
,
&
s
->
status_
,
static_cast
<
void
*>
(
s
)
);
req_count_
++
;
}
...
...
@@ -181,7 +183,7 @@ void RPCClient::AsyncSendFetchBarrier(const std::string& ep, int64_t time_out) {
sendrecv
::
VariableMessage
req
;
req
.
set_varname
(
FETCH_BARRIER_MESSAGE
);
auto
rpc
=
s
->
stub_
->
AsyncGetVariable
(
s
->
context_
.
get
(),
req
,
&
cq_
);
rpc
->
Finish
(
&
s
->
reply_
,
&
s
->
status_
,
(
void
*
)
s
);
rpc
->
Finish
(
&
s
->
reply_
,
&
s
->
status_
,
static_cast
<
void
*>
(
s
)
);
req_count_
++
;
}
...
...
paddle/fluid/operators/detail/grpc_server.cc
浏览文件 @
45363069
...
...
@@ -13,7 +13,9 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/detail/grpc_server.h"
#include <paddle/fluid/operators/detail/send_recv.pb.h>
#include <limits>
#include <string>
using
::
grpc
::
ServerAsyncResponseWriter
;
...
...
@@ -224,6 +226,7 @@ void AsyncGRPCServer::ShutdownQueue() {
std
::
unique_lock
<
std
::
mutex
>
lock
(
cq_mutex_
);
cq_send_
->
Shutdown
();
cq_get_
->
Shutdown
();
cq_prefetch_
->
Shutdown
();
}
// This URL explains why shutdown is complicate:
...
...
@@ -236,6 +239,7 @@ void AsyncGRPCServer::ShutDown() {
void
AsyncGRPCServer
::
TryToRegisterNewSendOne
()
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
cq_mutex_
);
if
(
is_shut_down_
)
{
VLOG
(
3
)
<<
"shutdown, do not TryToRegisterNewSendOne"
;
return
;
}
RequestSend
*
send
=
new
RequestSend
(
&
service_
,
cq_send_
.
get
(),
scope_
,
...
...
@@ -246,6 +250,7 @@ void AsyncGRPCServer::TryToRegisterNewSendOne() {
void
AsyncGRPCServer
::
TryToRegisterNewGetOne
()
{
std
::
unique_lock
<
std
::
mutex
>
lock
(
cq_mutex_
);
if
(
is_shut_down_
)
{
VLOG
(
3
)
<<
"shutdown, do not TryToRegisterNewGetOne"
;
return
;
}
RequestGet
*
get
=
new
RequestGet
(
&
service_
,
cq_get_
.
get
(),
scope_
,
dev_ctx_
,
...
...
@@ -257,6 +262,7 @@ void AsyncGRPCServer::TryToRegisterNewPrefetchOne() {
VLOG
(
4
)
<<
"TryToRegisterNewPrefetchOne in"
;
std
::
unique_lock
<
std
::
mutex
>
lock
(
cq_mutex_
);
if
(
is_shut_down_
)
{
VLOG
(
3
)
<<
"shutdown, do not TryToRegisterNewPrefetchOne"
;
return
;
}
RequestPrefetch
*
prefetch
=
...
...
@@ -274,18 +280,21 @@ void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq,
void
*
tag
=
NULL
;
bool
ok
=
false
;
while
(
true
)
{
VLOG
(
3
)
<<
"HandleRequest for "
<<
cq_name
<<
" while in"
;
if
(
!
cq
->
Next
(
&
tag
,
&
ok
))
{
LOG
(
INFO
)
<<
cq_name
<<
" CompletionQueue shutdown!"
;
break
;
}
VLOG
(
3
)
<<
"HandleRequest for "
<<
cq_name
<<
" while after Next"
;
PADDLE_ENFORCE
(
tag
);
// 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
);
RequestBase
*
base
=
(
RequestBase
*
)
tag
;
RequestBase
*
base
=
reinterpret_cast
<
RequestBase
*>
(
tag
)
;
// reference:
// https://github.com/tensorflow/tensorflow/issues/5596
// https://groups.google.com/forum/#!topic/grpc-io/xftlRy-IQwM
...
...
paddle/fluid/operators/detail/grpc_service.h
浏览文件 @
45363069
...
...
@@ -89,7 +89,7 @@ inline const char* GrpcMethodName(GrpcMethod id) {
case
GrpcMethod
::
kGetVariable
:
return
"/sendrecv.SendRecvService/GetVariable"
;
case
GrpcMethod
::
kPrefetchVariable
:
return
"/sendrecv.SendR
E
cvService/PrefetchVariable"
;
return
"/sendrecv.SendR
e
cvService/PrefetchVariable"
;
}
// Shouldn't be reached.
...
...
@@ -117,5 +117,5 @@ class GrpcService final {
};
}
// namespace detail
}
// namespace operator
}
// namespace operator
s
}
// namespace paddle
paddle/fluid/operators/listen_and_serv_op.cc
浏览文件 @
45363069
...
...
@@ -13,22 +13,13 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include <stdint.h>
#include <sys/stat.h>
#include <ostream>
#include <thread>
#include <unistd.h>
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/proto_desc.h"
#include "paddle/fluid/framework/threadpool.h"
#include "paddle/fluid/operators/detail/grpc_server.h"
#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
#include "paddle/fluid/operators/detail/simple_block_queue.h"
#include "paddle/fluid/string/printf.h"
namespace
paddle
{
namespace
operators
{
...
...
@@ -177,7 +168,8 @@ class ListenAndServOp : public framework::OperatorBase {
}
ParallelExecuteBlocks
(
parallel_blkids
,
&
executor
,
program
,
&
recv_scope
);
VLOG
(
2
)
<<
"run all blocks spent (ms) "
<<
detail
::
GetTimestamp
()
-
ts
;
VLOG
(
3
)
<<
"run all blocks spent "
<<
detail
::
GetTimestamp
()
-
ts
<<
"(ms)"
;
// Reset the received sparse variables, the sum operator would not
// sum the input sparse variables which rows is empty at the next
...
...
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