Skip to content

S
Serving

PaddlePaddle / Serving
大约 2 年 前同步成功

通知 187
Star 833
Fork 253
S
Serving
未验证 提交 b7ad237a 编写于 作者: B barriery 提交者： GitHub
浏览文件
操作

Merge pull request #769 from barrierye/pipeline-auto-batch 

Pipeline Auto-batching feature & Update Log
上级 826a7c38 2ce328c9
展开全部 显示空白变更内容
内联 并排
Showing with 2199 addition and 932 deletion
core/configure/proto/multi_lang_general_model_service.proto
浏览文件 @ b7ad237a
......@@ -37,6 +37,7 @@ message InferenceRequest {
repeated string feed_var_names = 2;
repeated string fetch_var_names = 3;
required bool is_python = 4 [ default = false ];
required uint64 log_id = 5 [ default = 0 ];
};
message InferenceResponse {
......
core/general-client/include/general_model.h
浏览文件 @ b7ad237a
......@@ -227,7 +227,8 @@ class PredictorClient {
const std::vector<std::vector<int>>& int_shape,
const std::vector<std::string>& fetch_name,
PredictorRes& predict_res_batch, // NOLINT
const int& pid);
const int& pid,
const uint64_t log_id);
int numpy_predict(
const std::vector<std::vector<py::array_t<float>>>& float_feed_batch,
......@@ -238,7 +239,8 @@ class PredictorClient {
const std::vector<std::vector<int>>& int_shape,
const std::vector<std::string>& fetch_name,
PredictorRes& predict_res_batch, // NOLINT
const int& pid);
const int& pid,
const uint64_t log_id);
private:
PredictorApi _api;
......
core/general-client/src/general_model.cpp
浏览文件 @ b7ad237a
......@@ -144,7 +144,8 @@ int PredictorClient::batch_predict(
const std::vector<std::vector<int>> &int_shape,
const std::vector<std::string> &fetch_name,
PredictorRes &predict_res_batch,
const int &pid) {
const int &pid,
const uint64_t log_id) {
int batch_size = std::max(float_feed_batch.size(), int_feed_batch.size());
predict_res_batch.clear();
......@@ -162,6 +163,7 @@ int PredictorClient::batch_predict(
VLOG(2) << "int feed name size: " << int_feed_name.size();
VLOG(2) << "max body size : " << brpc::fLU64::FLAGS_max_body_size;
Request req;
req.set_log_id(log_id);
for (auto &name : fetch_name) {
req.add_fetch_var_names(name);
}
......@@ -356,7 +358,8 @@ int PredictorClient::numpy_predict(
const std::vector<std::vector<int>> &int_shape,
const std::vector<std::string> &fetch_name,
PredictorRes &predict_res_batch,
const int &pid) {
const int &pid,
const uint64_t log_id) {
int batch_size = std::max(float_feed_batch.size(), int_feed_batch.size());
VLOG(2) << "batch size: " << batch_size;
predict_res_batch.clear();
......@@ -374,6 +377,7 @@ int PredictorClient::numpy_predict(
VLOG(2) << "int feed name size: " << int_feed_name.size();
VLOG(2) << "max body size : " << brpc::fLU64::FLAGS_max_body_size;
Request req;
req.set_log_id(log_id);
for (auto &name : fetch_name) {
req.add_fetch_var_names(name);
}
......
core/general-client/src/pybind_general_model.cpp
浏览文件 @ b7ad237a
......@@ -107,7 +107,8 @@ PYBIND11_MODULE(serving_client, m) {
const std::vector<std::vector<int>> &int_shape,
const std::vector<std::string> &fetch_name,
PredictorRes &predict_res_batch,
const int &pid) {
const int &pid,
const uint64_t log_id) {
return self.batch_predict(float_feed_batch,
float_feed_name,
float_shape,
......@@ -116,7 +117,8 @@ PYBIND11_MODULE(serving_client, m) {
int_shape,
fetch_name,
predict_res_batch,
pid);
pid,
log_id);
},
py::call_guard<py::gil_scoped_release>())
.def("numpy_predict",
......@@ -131,7 +133,8 @@ PYBIND11_MODULE(serving_client, m) {
const std::vector<std::vector<int>> &int_shape,
const std::vector<std::string> &fetch_name,
PredictorRes &predict_res_batch,
const int &pid) {
const int &pid,
const uint64_t log_id) {
return self.numpy_predict(float_feed_batch,
float_feed_name,
float_shape,
......@@ -140,7 +143,8 @@ PYBIND11_MODULE(serving_client, m) {
int_shape,
fetch_name,
predict_res_batch,
pid);
pid,
log_id);
},
py::call_guard<py::gil_scoped_release>());
}
......
core/general-server/op/general_copy_op.cpp
浏览文件 @ b7ad237a
......@@ -45,36 +45,41 @@ int GeneralCopyOp::inference() {
const std::string pre_name = pre_node_names[0];
const GeneralBlob *input_blob = get_depend_argument<GeneralBlob>(pre_name);
VLOG(2) << "precedent name: " << pre_name;
uint64_t log_id = input_blob->GetLogId();
VLOG(2) << "(logid=" << log_id << ") precedent name: " << pre_name;
const TensorVector *in = &input_blob->tensor_vector;
VLOG(2) << "input size: " << in->size();
VLOG(2) << "(logid=" << log_id << ") input size: " << in->size();
int batch_size = input_blob->GetBatchSize();
int input_var_num = 0;
GeneralBlob *res = mutable_data<GeneralBlob>();
res->SetLogId(log_id);
TensorVector *out = &res->tensor_vector;
VLOG(2) << "input batch size: " << batch_size;
VLOG(2) << "(logid=" << log_id << ") input batch size: " << batch_size;
res->SetBatchSize(batch_size);
if (!res) {
LOG(ERROR) << "Failed get op tls reader object output";
LOG(ERROR) << "(logid=" << log_id
<< ") Failed get op tls reader object output";
}
Timer timeline;
int64_t start = timeline.TimeStampUS();
VLOG(2) << "Going to init lod tensor";
VLOG(2) << "(logid=" << log_id << ") Going to init lod tensor";
for (int i = 0; i < in->size(); ++i) {
paddle::PaddleTensor lod_tensor;
CopyLod(&in->at(i), &lod_tensor);
lod_tensor.dtype = in->at(i).dtype;
lod_tensor.name = in->at(i).name;
VLOG(2) << "lod tensor [" << i << "].name = " << lod_tensor.name;
VLOG(2) << "(logid=" << log_id << ") lod tensor [" << i
<< "].name = " << lod_tensor.name;
out->push_back(lod_tensor);
}
VLOG(2) << "pack done.";
VLOG(2) << "(logid=" << log_id << ") pack done.";
for (int i = 0; i < out->size(); ++i) {
int64_t *src_ptr = static_cast<int64_t *>(in->at(i).data.data());
......@@ -86,7 +91,7 @@ int GeneralCopyOp::inference() {
}
}
VLOG(2) << "output done.";
VLOG(2) << "(logid=" << log_id << ") output done.";
timeline.Pause();
int64_t end = timeline.TimeStampUS();
......@@ -94,7 +99,7 @@ int GeneralCopyOp::inference() {
AddBlobInfo(res, start);
AddBlobInfo(res, end);
VLOG(2) << "read data from client success";
VLOG(2) << "(logid=" << log_id << ") read data from client success";
return 0;
}
......
core/general-server/op/general_dist_kv_infer_op.cpp
浏览文件 @ b7ad237a
......@@ -50,18 +50,20 @@ int GeneralDistKVInferOp::inference() {
const std::string pre_name = pre_node_names[0];
const GeneralBlob *input_blob = get_depend_argument<GeneralBlob>(pre_name);
VLOG(2) << "Get precedent op name: " << pre_name;
uint64_t log_id = input_blob->GetLogId();
VLOG(2) << "(logid=" << log_id << ") Get precedent op name: " << pre_name;
GeneralBlob *output_blob = mutable_data<GeneralBlob>();
if (!input_blob) {
LOG(ERROR) << "Failed mutable depended argument, op:" << pre_name;
LOG(ERROR) << "(logid=" << log_id
<< ") Failed mutable depended argument, op:" << pre_name;
return -1;
}
const TensorVector *in = &input_blob->tensor_vector;
TensorVector *out = &output_blob->tensor_vector;
int batch_size = input_blob->GetBatchSize();
VLOG(2) << "input batch size: " << batch_size;
VLOG(2) << "(logid=" << log_id << ") input batch size: " << batch_size;
std::vector<uint64_t> keys;
std::vector<rec::mcube::CubeValue> values;
int sparse_count = 0;
......@@ -96,13 +98,14 @@ int GeneralDistKVInferOp::inference() {
rec::mcube::CubeAPI *cube = rec::mcube::CubeAPI::instance();
std::vector<std::string> table_names = cube->get_table_names();
if (table_names.size() == 0) {
LOG(ERROR) << "cube init error or cube config not given.";
LOG(ERROR) << "(logid=" << log_id
<< ") cube init error or cube config not given.";
return -1;
}
int ret = cube->seek(table_names[0], keys, &values);
int64_t cube_end = timeline.TimeStampUS();
if (values.size() != keys.size() || values[0].buff.size() == 0) {
LOG(ERROR) << "cube value return null";
LOG(ERROR) << "(logid=" << log_id << ") cube value return null";
}
size_t EMBEDDING_SIZE = values[0].buff.size() / sizeof(float);
TensorVector sparse_out;
......@@ -153,14 +156,16 @@ int GeneralDistKVInferOp::inference() {
infer_in.insert(infer_in.end(), sparse_out.begin(), sparse_out.end());
output_blob->SetBatchSize(batch_size);
output_blob->SetLogId(log_id);
VLOG(2) << "infer batch size: " << batch_size;
VLOG(2) << "(logid=" << log_id << ") infer batch size: " << batch_size;
int64_t start = timeline.TimeStampUS();
if (InferManager::instance().infer(
engine_name().c_str(), &infer_in, out, batch_size)) {
LOG(ERROR) << "Failed do infer in fluid model: " << engine_name();
LOG(ERROR) << "(logid=" << log_id
<< ") Failed do infer in fluid model: " << engine_name();
return -1;
}
......
core/general-server/op/general_dist_kv_quant_infer_op.cpp
浏览文件 @ b7ad237a
......@@ -59,10 +59,13 @@ int GeneralDistKVQuantInferOp::inference() {
return -1;
}
uint64_t log_id = input_blob->GetLogId();
output_blob->SetLogId(log_id);
const TensorVector *in = &input_blob->tensor_vector;
TensorVector *out = &output_blob->tensor_vector;
int batch_size = input_blob->GetBatchSize();
VLOG(2) << "input batch size: " << batch_size;
VLOG(2) << "(logid=" << log_id << ") input batch size: " << batch_size;
std::vector<uint64_t> keys;
std::vector<rec::mcube::CubeValue> values;
int sparse_count = 0;
......@@ -94,13 +97,14 @@ int GeneralDistKVQuantInferOp::inference() {
rec::mcube::CubeAPI *cube = rec::mcube::CubeAPI::instance();
std::vector<std::string> table_names = cube->get_table_names();
if (table_names.size() == 0) {
LOG(ERROR) << "cube init error or cube config not given.";
LOG(ERROR) << "(logid=" << log_id
<< ") cube init error or cube config not given.";
return -1;
}
int ret = cube->seek(table_names[0], keys, &values);
if (values.size() != keys.size() || values[0].buff.size() == 0) {
LOG(ERROR) << "cube value return null";
LOG(ERROR) << "(logid=" << log_id << ") cube value return null";
}
TensorVector sparse_out;
......@@ -182,7 +186,7 @@ int GeneralDistKVQuantInferOp::inference() {
output_blob->SetBatchSize(batch_size);
VLOG(2) << "infer batch size: " << batch_size;
VLOG(2) << "(logid=" << log_id << ") infer batch size: " << batch_size;
Timer timeline;
int64_t start = timeline.TimeStampUS();
......@@ -190,7 +194,8 @@ int GeneralDistKVQuantInferOp::inference() {
if (InferManager::instance().infer(
engine_name().c_str(), &infer_in, out, batch_size)) {
LOG(ERROR) << "Failed do infer in fluid model: " << engine_name();
LOG(ERROR) << "(logid=" << log_id
<< ") Failed do infer in fluid model: " << engine_name();
return -1;
}
......
core/general-server/op/general_infer_helper.h
浏览文件 @ b7ad237a
......@@ -35,6 +35,7 @@ struct GeneralBlob {
std::vector<paddle::PaddleTensor> tensor_vector;
int64_t time_stamp[20];
int p_size = 0;
uint64_t _log_id = -1; // for logging
int _batch_size;
......@@ -46,9 +47,11 @@ struct GeneralBlob {
tensor_vector.clear();
}
int SetBatchSize(int batch_size) { _batch_size = batch_size; }
void SetBatchSize(int batch_size) { _batch_size = batch_size; }
void SetLogId(uint64_t log_id) { _log_id = log_id; }
int GetBatchSize() const { return _batch_size; }
uint64_t GetLogId() const { return _log_id; }
std::string ShortDebugString() const { return "Not implemented!"; }
};
......
core/general-server/op/general_infer_op.cpp
浏览文件 @ b7ad237a
......@@ -47,22 +47,25 @@ int GeneralInferOp::inference() {
const std::string pre_name = pre_node_names[0];
const GeneralBlob *input_blob = get_depend_argument<GeneralBlob>(pre_name);
VLOG(2) << "Get precedent op name: " << pre_name;
uint64_t log_id = input_blob->GetLogId();
VLOG(2) << "(logid=" << log_id << ") Get precedent op name: " << pre_name;
GeneralBlob *output_blob = mutable_data<GeneralBlob>();
output_blob->SetLogId(log_id);
if (!input_blob) {
LOG(ERROR) << "Failed mutable depended argument, op:" << pre_name;
LOG(ERROR) << "(logid=" << log_id
<< ") Failed mutable depended argument, op:" << pre_name;
return -1;
}
const TensorVector *in = &input_blob->tensor_vector;
TensorVector *out = &output_blob->tensor_vector;
int batch_size = input_blob->GetBatchSize();
VLOG(2) << "input batch size: " << batch_size;
VLOG(2) << "(logid=" << log_id << ") input batch size: " << batch_size;
output_blob->SetBatchSize(batch_size);
VLOG(2) << "infer batch size: " << batch_size;
VLOG(2) << "(logid=" << log_id << ") infer batch size: " << batch_size;
Timer timeline;
int64_t start = timeline.TimeStampUS();
......@@ -70,7 +73,8 @@ int GeneralInferOp::inference() {
if (InferManager::instance().infer(
engine_name().c_str(), in, out, batch_size)) {
LOG(ERROR) << "Failed do infer in fluid model: " << engine_name().c_str();
LOG(ERROR) << "(logid=" << log_id
<< ") Failed do infer in fluid model: " << engine_name().c_str();
return -1;
}
......
core/general-server/op/general_reader_op.cpp
浏览文件 @ b7ad237a
......@@ -37,9 +37,9 @@ int conf_check(const Request *req,
const std::shared_ptr<PaddleGeneralModelConfig> &model_config) {
int var_num = req->insts(0).tensor_array_size();
if (var_num != model_config->_feed_type.size()) {
VLOG(2) << "var num: " << var_num;
VLOG(2) << "model config var num: " << model_config->_feed_type.size();
LOG(ERROR) << "feed var number not match.";
LOG(ERROR) << "feed var number not match: model config["
<< model_config->_feed_type.size() << "] vs. actual[" << var_num
<< "]";
return -1;
}
......@@ -72,6 +72,7 @@ int conf_check(const Request *req,
int GeneralReaderOp::inference() {
// reade request from client
const Request *req = dynamic_cast<const Request *>(get_request_message());
uint64_t log_id = req->log_id();
int batch_size = req->insts_size();
int input_var_num = 0;
......@@ -83,25 +84,28 @@ int GeneralReaderOp::inference() {
TensorVector *out = &res->tensor_vector;
res->SetBatchSize(batch_size);
res->SetLogId(log_id);
if (!res) {
LOG(ERROR) << "Failed get op tls reader object output";
LOG(ERROR) << "(logid=" << log_id
<< ") Failed get op tls reader object output";
}
Timer timeline;
int64_t start = timeline.TimeStampUS();
int var_num = req->insts(0).tensor_array_size();
VLOG(2) << "var num: " << var_num;
VLOG(2) << "(logid=" << log_id << ") var num: " << var_num;
VLOG(2) << "start to call load general model_conf op";
VLOG(2) << "(logid=" << log_id
<< ") start to call load general model_conf op";
baidu::paddle_serving::predictor::Resource &resource =
baidu::paddle_serving::predictor::Resource::instance();
VLOG(2) << "get resource pointer done.";
VLOG(2) << "(logid=" << log_id << ") get resource pointer done.";
std::shared_ptr<PaddleGeneralModelConfig> model_config =
resource.get_general_model_config();
VLOG(2) << "print general model config done.";
VLOG(2) << "(logid=" << log_id << ") print general model config done.";
// TODO(guru4elephant): how to do conditional check?
/*
......@@ -122,7 +126,8 @@ int GeneralReaderOp::inference() {
for (int i = 0; i < var_num; ++i) {
paddle::PaddleTensor lod_tensor;
elem_type[i] = req->insts(0).tensor_array(i).elem_type();
VLOG(2) << "var[" << i << "] has elem type: " << elem_type[i];
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] has elem type: " << elem_type[i];
if (elem_type[i] == 0) { // int64
elem_size[i] = sizeof(int64_t);
lod_tensor.dtype = paddle::PaddleDType::INT64;
......@@ -137,17 +142,19 @@ int GeneralReaderOp::inference() {
if (model_config->_is_lod_feed[i]) {
lod_tensor.lod.resize(1);
lod_tensor.lod[0].push_back(0);
VLOG(2) << "var[" << i << "] is lod_tensor";
VLOG(2) << "(logid=" << log_id << ") var[" << i << "] is lod_tensor";
} else {
lod_tensor.shape.push_back(batch_size);
capacity[i] = 1;
for (int k = 0; k < req->insts(0).tensor_array(i).shape_size(); ++k) {
int dim = req->insts(0).tensor_array(i).shape(k);
VLOG(2) << "shape for var[" << i << "]: " << dim;
VLOG(2) << "(logid=" << log_id << ") shape for var[" << i
<< "]: " << dim;
capacity[i] *= dim;
lod_tensor.shape.push_back(dim);
}
VLOG(2) << "var[" << i << "] is tensor, capacity: " << capacity[i];
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] is tensor, capacity: " << capacity[i];
}
lod_tensor.name = model_config->_feed_name[i];
out->push_back(lod_tensor);
......@@ -167,11 +174,12 @@ int GeneralReaderOp::inference() {
} else if (tensor.int_data_size() > 0) {
data_len = tensor.int_data_size();
}
VLOG(2) << "tensor size for var[" << i << "]: " << data_len;
VLOG(2) << "(logid=" << log_id << ") tensor size for var[" << i
<< "]: " << data_len;
tensor_size += data_len;
int cur_len = out->at(i).lod[0].back();
VLOG(2) << "current len: " << cur_len;
VLOG(2) << "(logid=" << log_id << ") current len: " << cur_len;
int sample_len = 0;
if (tensor.shape_size() == 1) {
......@@ -180,7 +188,7 @@ int GeneralReaderOp::inference() {
sample_len = tensor.shape(0);
}
out->at(i).lod[0].push_back(cur_len + sample_len);
VLOG(2) << "new len: " << cur_len + sample_len;
VLOG(2) << "(logid=" << log_id << ") new len: " << cur_len + sample_len;
}
out->at(i).data.Resize(tensor_size * elem_size[i]);
out->at(i).shape = {out->at(i).lod[0].back()};
......@@ -190,11 +198,11 @@ int GeneralReaderOp::inference() {
if (out->at(i).shape.size() == 1) {
out->at(i).shape.push_back(1);
}
VLOG(2) << "var[" << i
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] is lod_tensor and len=" << out->at(i).lod[0].back();
} else {
out->at(i).data.Resize(batch_size * capacity[i] * elem_size[i]);
VLOG(2) << "var[" << i
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] is tensor and capacity=" << batch_size * capacity[i];
}
}
......@@ -203,8 +211,8 @@ int GeneralReaderOp::inference() {
for (int i = 0; i < var_num; ++i) {
if (elem_type[i] == 0) {
int64_t *dst_ptr = static_cast<int64_t *>(out->at(i).data.data());
VLOG(2) << "first element data in var[" << i << "] is "
<< req->insts(0).tensor_array(i).int64_data(0);
VLOG(2) << "(logid=" << log_id << ") first element data in var[" << i
<< "] is " << req->insts(0).tensor_array(i).int64_data(0);
int offset = 0;
for (int j = 0; j < batch_size; ++j) {
int elem_num = req->insts(j).tensor_array(i).int64_data_size();
......@@ -219,8 +227,8 @@ int GeneralReaderOp::inference() {
}
} else if (elem_type[i] == 1) {
float *dst_ptr = static_cast<float *>(out->at(i).data.data());
VLOG(2) << "first element data in var[" << i << "] is "
<< req->insts(0).tensor_array(i).float_data(0);
VLOG(2) << "(logid=" << log_id << ") first element data in var[" << i
<< "] is " << req->insts(0).tensor_array(i).float_data(0);
int offset = 0;
for (int j = 0; j < batch_size; ++j) {
int elem_num = req->insts(j).tensor_array(i).float_data_size();
......@@ -235,8 +243,8 @@ int GeneralReaderOp::inference() {
}
} else if (elem_type[i] == 2) {
int32_t *dst_ptr = static_cast<int32_t *>(out->at(i).data.data());
VLOG(2) << "first element data in var[" << i << "] is "
<< req->insts(0).tensor_array(i).int_data(0);
VLOG(2) << "(logid=" << log_id << ") first element data in var[" << i
<< "] is " << req->insts(0).tensor_array(i).int_data(0);
int offset = 0;
for (int j = 0; j < batch_size; ++j) {
int elem_num = req->insts(j).tensor_array(i).int_data_size();
......@@ -252,7 +260,7 @@ int GeneralReaderOp::inference() {
}
}
VLOG(2) << "output size: " << out->size();
VLOG(2) << "(logid=" << log_id << ") output size: " << out->size();
timeline.Pause();
int64_t end = timeline.TimeStampUS();
......@@ -260,7 +268,7 @@ int GeneralReaderOp::inference() {
AddBlobInfo(res, start);
AddBlobInfo(res, end);
VLOG(2) << "read data from client success";
VLOG(2) << "(logid=" << log_id << ") read data from client success";
return 0;
}
DEFINE_OP(GeneralReaderOp);
......
core/general-server/op/general_response_op.cpp
浏览文件 @ b7ad237a
......@@ -42,6 +42,9 @@ using baidu::paddle_serving::predictor::PaddleGeneralModelConfig;
int GeneralResponseOp::inference() {
const std::vector<std::string> pre_node_names = pre_names();
VLOG(2) << "pre node names size: " << pre_node_names.size();
const GeneralBlob *input_blob;
uint64_t log_id =
get_depend_argument<GeneralBlob>(pre_node_names[0])->GetLogId();
const Request *req = dynamic_cast<const Request *>(get_request_message());
// response inst with only fetch_var_names
......@@ -52,15 +55,17 @@ int GeneralResponseOp::inference() {
// timeline.Start();
int64_t start = timeline.TimeStampUS();
VLOG(2) << "start to call load general model_conf op";
VLOG(2) << "(logid=" << log_id
<< ") start to call load general model_conf op";
baidu::paddle_serving::predictor::Resource &resource =
baidu::paddle_serving::predictor::Resource::instance();
VLOG(2) << "get resource pointer done.";
VLOG(2) << "(logid=" << log_id << ") get resource pointer done.";
std::shared_ptr<PaddleGeneralModelConfig> model_config =
resource.get_general_model_config();
VLOG(2) << "max body size : " << brpc::fLU64::FLAGS_max_body_size;
VLOG(2) << "(logid=" << log_id
<< ") max body size : " << brpc::fLU64::FLAGS_max_body_size;
std::vector<int> fetch_index;
fetch_index.resize(req->fetch_var_names_size());
......@@ -69,16 +74,16 @@ int GeneralResponseOp::inference() {
model_config->_fetch_alias_name_to_index[req->fetch_var_names(i)];
}
const GeneralBlob *input_blob;
for (uint32_t pi = 0; pi < pre_node_names.size(); ++pi) {
const std::string &pre_name = pre_node_names[pi];
VLOG(2) << "pre names[" << pi << "]: " << pre_name << " ("
<< pre_node_names.size() << ")";
VLOG(2) << "(logid=" << log_id << ") pre names[" << pi << "]: " << pre_name
<< " (" << pre_node_names.size() << ")";
input_blob = get_depend_argument<GeneralBlob>(pre_name);
// fprintf(stderr, "input(%s) blob address %x\n", pre_names.c_str(),
// input_blob);
if (!input_blob) {
LOG(ERROR) << "Failed mutable depended argument, op: " << pre_name;
LOG(ERROR) << "(logid=" << log_id
<< ") Failed mutable depended argument, op: " << pre_name;
return -1;
}
......@@ -92,17 +97,19 @@ int GeneralResponseOp::inference() {
for (auto &idx : fetch_index) {
Tensor *tensor = fetch_inst->add_tensor_array();
if (model_config->_is_lod_fetch[idx]) {
VLOG(2) << "out[" << idx << "] " << model_config->_fetch_name[idx]
<< " is lod_tensor";
VLOG(2) << "(logid=" << log_id << ") out[" << idx << "] "
<< model_config->_fetch_name[idx] << " is lod_tensor";
for (int k = 0; k < in->at(idx).shape.size(); ++k) {
VLOG(2) << "shape[" << k << "]: " << in->at(idx).shape[k];
VLOG(2) << "(logid=" << log_id << ") shape[" << k
<< "]: " << in->at(idx).shape[k];
tensor->add_shape(in->at(idx).shape[k]);
}
} else {
VLOG(2) << "out[" << idx << "] " << model_config->_fetch_name[idx]
<< " is tensor";
VLOG(2) << "(logid=" << log_id << ") out[" << idx << "] "
<< model_config->_fetch_name[idx] << " is tensor";
for (int k = 0; k < in->at(idx).shape.size(); ++k) {
VLOG(2) << "shape[" << k << "]: " << in->at(idx).shape[k];
VLOG(2) << "(logid=" << log_id << ") shape[" << k
<< "]: " << in->at(idx).shape[k];
tensor->add_shape(in->at(idx).shape[k]);
}
}
......@@ -119,8 +126,8 @@ int GeneralResponseOp::inference() {
auto dtype = in->at(idx).dtype;
if (dtype == paddle::PaddleDType::INT64) {
VLOG(2) << "Prepare int64 var [" << model_config->_fetch_name[idx]
<< "].";
VLOG(2) << "(logid=" << log_id << ") Prepare int64 var ["
<< model_config->_fetch_name[idx] << "].";
int64_t *data_ptr = static_cast<int64_t *>(in->at(idx).data.data());
// from
// https://stackoverflow.com/questions/15499641/copy-a-stdvector-to-a-repeated-field-from-protobuf-with-memcpy
......@@ -130,16 +137,16 @@ int GeneralResponseOp::inference() {
fetch_p->mutable_tensor_array(var_idx)->mutable_int64_data()->Swap(
&tmp_data);
} else if (dtype == paddle::PaddleDType::FLOAT32) {
VLOG(2) << "Prepare float var [" << model_config->_fetch_name[idx]
<< "].";
VLOG(2) << "(logid=" << log_id << ") Prepare float var ["
<< model_config->_fetch_name[idx] << "].";
float *data_ptr = static_cast<float *>(in->at(idx).data.data());
google::protobuf::RepeatedField<float> tmp_data(data_ptr,
data_ptr + cap);
fetch_p->mutable_tensor_array(var_idx)->mutable_float_data()->Swap(
&tmp_data);
} else if (dtype == paddle::PaddleDType::INT32) {
VLOG(2) << "Prepare int32 var [" << model_config->_fetch_name[idx]
<< "].";
VLOG(2) << "(logid=" << log_id << ")Prepare int32 var ["
<< model_config->_fetch_name[idx] << "].";
int32_t *data_ptr = static_cast<int32_t *>(in->at(idx).data.data());
google::protobuf::RepeatedField<int32_t> tmp_data(data_ptr,
data_ptr + cap);
......@@ -154,7 +161,8 @@ int GeneralResponseOp::inference() {
}
}
VLOG(2) << "fetch var [" << model_config->_fetch_name[idx] << "] ready";
VLOG(2) << "(logid=" << log_id << ") fetch var ["
<< model_config->_fetch_name[idx] << "] ready";
var_idx++;
}
}
......@@ -167,7 +175,8 @@ int GeneralResponseOp::inference() {
// a more elegant way.
for (uint32_t pi = 0; pi < pre_node_names.size(); ++pi) {
input_blob = get_depend_argument<GeneralBlob>(pre_node_names[pi]);
VLOG(2) << "p size for input blob: " << input_blob->p_size;
VLOG(2) << "(logid=" << log_id
<< ") p size for input blob: " << input_blob->p_size;
int profile_time_idx = -1;
if (pi == 0) {
profile_time_idx = 0;
......
core/general-server/op/general_text_reader_op.cpp
浏览文件 @ b7ad237a
......@@ -35,6 +35,7 @@ using baidu::paddle_serving::predictor::PaddleGeneralModelConfig;
int GeneralTextReaderOp::inference() {
// reade request from client
const Request *req = dynamic_cast<const Request *>(get_request_message());
uint64_t log_id = req->log_id();
int batch_size = req->insts_size();
int input_var_num = 0;
......@@ -44,16 +45,18 @@ int GeneralTextReaderOp::inference() {
std::vector<int64_t> capacity;
GeneralBlob *res = mutable_data<GeneralBlob>();
TensorVector *out = &res->tensor_vector;
res->SetBatchSize(batch_size);
if (!res) {
LOG(ERROR) << "Failed get op tls reader object output";
LOG(ERROR) << "(logid=" << log_id
<< ") Failed get op tls reader object output";
}
TensorVector *out = &res->tensor_vector;
res->SetBatchSize(batch_size);
res->SetLogId(log_id);
if (batch_size <= 0) {
LOG(ERROR) << "Batch size < 0";
LOG(ERROR) << "(logid=" << log_id << ") Batch size < 0";
return -1;
}
......@@ -61,17 +64,18 @@ int GeneralTextReaderOp::inference() {
int64_t start = timeline.TimeStampUS();
int var_num = req->insts(0).tensor_array_size();
VLOG(2) << "var num: " << var_num;
VLOG(2) << "(logid=" << log_id << ") var num: " << var_num;
VLOG(2) << "start to call load general model_conf op";
VLOG(2) << "(logid=" << log_id
<< ") start to call load general model_conf op";
baidu::paddle_serving::predictor::Resource &resource =
baidu::paddle_serving::predictor::Resource::instance();
VLOG(2) << "get resource pointer done.";
VLOG(2) << "(logid=" << log_id << ") get resource pointer done.";
std::shared_ptr<PaddleGeneralModelConfig> model_config =
resource.get_general_model_config();
VLOG(2) << "print general model config done.";
VLOG(2) << "(logid=" << log_id << ") print general model config done.";
elem_type.resize(var_num);
elem_size.resize(var_num);
......@@ -79,7 +83,8 @@ int GeneralTextReaderOp::inference() {
for (int i = 0; i < var_num; ++i) {
paddle::PaddleTensor lod_tensor;
elem_type[i] = req->insts(0).tensor_array(i).elem_type();
VLOG(2) << "var[" << i << "] has elem type: " << elem_type[i];
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] has elem type: " << elem_type[i];
if (elem_type[i] == 0) { // int64
elem_size[i] = sizeof(int64_t);
lod_tensor.dtype = paddle::PaddleDType::INT64;
......@@ -91,17 +96,19 @@ int GeneralTextReaderOp::inference() {
if (req->insts(0).tensor_array(i).shape(0) == -1) {
lod_tensor.lod.resize(1);
lod_tensor.lod[0].push_back(0);
VLOG(2) << "var[" << i << "] is lod_tensor";
VLOG(2) << "(logid=" << log_id << ") var[" << i << "] is lod_tensor";
} else {
lod_tensor.shape.push_back(batch_size);
capacity[i] = 1;
for (int k = 0; k < req->insts(0).tensor_array(i).shape_size(); ++k) {
int dim = req->insts(0).tensor_array(i).shape(k);
VLOG(2) << "shape for var[" << i << "]: " << dim;
VLOG(2) << "(logid=" << log_id << ") shape for var[" << i
<< "]: " << dim;
capacity[i] *= dim;
lod_tensor.shape.push_back(dim);
}
VLOG(2) << "var[" << i << "] is tensor, capacity: " << capacity[i];
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] is tensor, capacity: " << capacity[i];
}
lod_tensor.name = model_config->_feed_name[i];
out->push_back(lod_tensor);
......@@ -117,11 +124,11 @@ int GeneralTextReaderOp::inference() {
}
out->at(i).data.Resize(out->at(i).lod[0].back() * elem_size[i]);
out->at(i).shape = {out->at(i).lod[0].back(), 1};
VLOG(2) << "var[" << i
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] is lod_tensor and len=" << out->at(i).lod[0].back();
} else {
out->at(i).data.Resize(batch_size * capacity[i] * elem_size[i]);
VLOG(2) << "var[" << i
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] is tensor and capacity=" << batch_size * capacity[i];
}
}
......@@ -163,7 +170,7 @@ int GeneralTextReaderOp::inference() {
AddBlobInfo(res, start);
AddBlobInfo(res, end);
VLOG(2) << "read data from client success";
VLOG(2) << "(logid=" << log_id << ") read data from client success";
return 0;
}
DEFINE_OP(GeneralTextReaderOp);
......
core/general-server/op/general_text_response_op.cpp
浏览文件 @ b7ad237a
......@@ -40,6 +40,9 @@ int GeneralTextResponseOp::inference() {
VLOG(2) << "Going to run inference";
const std::vector<std::string> pre_node_names = pre_names();
VLOG(2) << "pre node names size: " << pre_node_names.size();
const GeneralBlob *input_blob;
uint64_t log_id =
get_depend_argument<GeneralBlob>(pre_node_names[0])->GetLogId();
const Request *req = dynamic_cast<const Request *>(get_request_message());
// response inst with only fetch_var_names
......@@ -48,11 +51,12 @@ int GeneralTextResponseOp::inference() {
Timer timeline;
int64_t start = timeline.TimeStampUS();
VLOG(2) << "start to call load general model_conf op";
VLOG(2) << "(logid=" << log_id
<< ") start to call load general model_conf op";
baidu::paddle_serving::predictor::Resource &resource =
baidu::paddle_serving::predictor::Resource::instance();
VLOG(2) << "get resource pointer done.";
VLOG(2) << "(logid=" << log_id << ") get resource pointer done.";
std::shared_ptr<PaddleGeneralModelConfig> model_config =
resource.get_general_model_config();
......@@ -63,20 +67,20 @@ int GeneralTextResponseOp::inference() {
model_config->_fetch_alias_name_to_index[req->fetch_var_names(i)];
}
const GeneralBlob *input_blob;
for (uint32_t pi = 0; pi < pre_node_names.size(); ++pi) {
const std::string &pre_name = pre_node_names[pi];
VLOG(2) << "pre names[" << pi << "]: " << pre_name << " ("
<< pre_node_names.size() << ")";
VLOG(2) << "(logid=" << log_id << ") pre names[" << pi << "]: " << pre_name
<< " (" << pre_node_names.size() << ")";
input_blob = get_depend_argument<GeneralBlob>(pre_name);
if (!input_blob) {
LOG(ERROR) << "Failed mutable depended argument, op: " << pre_name;
LOG(ERROR) << "(logid=" << log_id
<< ") Failed mutable depended argument, op: " << pre_name;
return -1;
}
const TensorVector *in = &input_blob->tensor_vector;
int batch_size = input_blob->GetBatchSize();
VLOG(2) << "input batch size: " << batch_size;
VLOG(2) << "(logid=" << log_id << ") input batch size: " << batch_size;
ModelOutput *output = res->add_outputs();
output->set_engine_name(
......@@ -88,12 +92,13 @@ int GeneralTextResponseOp::inference() {
// currently only response float tensor or lod_tensor
tensor->set_elem_type(1);
if (model_config->_is_lod_fetch[idx]) {
VLOG(2) << "out[" << idx << " is lod_tensor";
VLOG(2) << "(logid=" << log_id << ") out[" << idx << " is lod_tensor";
tensor->add_shape(-1);
} else {
VLOG(2) << "out[" << idx << "] is tensor";
VLOG(2) << "(logid=" << log_id << ") out[" << idx << "] is tensor";
for (int k = 1; k < in->at(idx).shape.size(); ++k) {
VLOG(2) << "shape[" << k - 1 << "]: " << in->at(idx).shape[k];
VLOG(2) << "(logid=" << log_id << ") shape[" << k - 1
<< "]: " << in->at(idx).shape[k];
tensor->add_shape(in->at(idx).shape[k]);
}
}
......@@ -137,7 +142,8 @@ int GeneralTextResponseOp::inference() {
// a more elegant way.
for (uint32_t pi = 0; pi < pre_node_names.size(); ++pi) {
input_blob = get_depend_argument<GeneralBlob>(pre_node_names[pi]);
VLOG(2) << "p size for input blob: " << input_blob->p_size;
VLOG(2) << "(logid=" << log_id
<< ") p size for input blob: " << input_blob->p_size;
int profile_time_idx = -1;
if (pi == 0) {
profile_time_idx = 0;
......
core/general-server/proto/general_model_service.proto
浏览文件 @ b7ad237a
......@@ -37,6 +37,7 @@ message Request {
repeated FeedInst insts = 1;
repeated string fetch_var_names = 2;
optional bool profile_server = 3 [ default = false ];
required uint64 log_id = 4 [ default = 0 ];
};
message Response {
......
core/general-server/proto/load_general_model_service.proto
浏览文件 @ b7ad237a
......@@ -21,6 +21,7 @@ option cc_generic_services = true;
message RequestAndResponse {
required int32 a = 1;
required float b = 2;
required uint64 log_id = 3 [ default = 0 ];
};
service LoadGeneralModelService {
......
core/pdcodegen/src/pdcodegen.cpp
浏览文件 @ b7ad237a
......@@ -280,25 +280,29 @@ class PdsCodeGenerator : public CodeGenerator {
" baidu::rpc::ClosureGuard done_guard(done);\n"
" baidu::rpc::Controller* cntl = \n"
" static_cast<baidu::rpc::Controller*>(cntl_base);\n"
" uint64_t log_id = request->log_id();\n"
" cntl->set_log_id(log_id);\n"
" ::baidu::paddle_serving::predictor::InferService* svr = \n"
" "
"::baidu::paddle_serving::predictor::InferServiceManager::instance("
").item(\"$service$\");\n"
" if (svr == NULL) {\n"
" LOG(ERROR) << \"Not found service: $service$\";\n"
" LOG(ERROR) << \"(logid=\" << log_id << \") Not found service: "
"$service$\";\n"
" cntl->SetFailed(404, \"Not found service: $service$\");\n"
" return ;\n"
" }\n"
" LOG(INFO) << \" remote_side=\[\" << cntl->remote_side() << " // NOLINT
"\"\]\";\n"
" LOG(INFO) << \" local_side=\[\" << cntl->local_side() << " // NOLINT
"\"\]\";\n"
" LOG(INFO) << \" service_name=\[\" << \"$name$\" << \"\]\";\n" // NOLINT
" LOG(INFO) << \" log_id=\[\" << cntl->log_id() << \"\]\";\n" // NOLINT
" int err_code = svr->inference(request, response);\n"
" LOG(INFO) << \"(logid=\" << log_id << \") remote_side=\[\" " // NOLINT
"<< cntl->remote_side() << \"\]\";\n"
" LOG(INFO) << \"(logid=\" << log_id << \") local_side=\[\" " // NOLINT
"<< cntl->local_side() << \"\]\";\n"
" LOG(INFO) << \"(logid=\" << log_id << \") service_name=\[\" " // NOLINT
"<< \"$name$\" << \"\]\";\n"
" int err_code = svr->inference(request, response, log_id);\n"
" if (err_code != 0) {\n"
" LOG(WARNING)\n"
" << \"Failed call inferservice[$name$], name[$service$]\"\n"
" << \"(logid=\" << log_id << \") Failed call "
"inferservice[$name$], name[$service$]\"\n"
" << \", error_code: \" << err_code;\n"
" cntl->SetFailed(err_code, \"InferService inference "
"failed!\");\n"
......@@ -306,7 +310,8 @@ class PdsCodeGenerator : public CodeGenerator {
" gettimeofday(&tv, NULL);\n"
" long end = tv.tv_sec * 1000000 + tv.tv_usec;\n"
" // flush notice log\n"
" LOG(INFO) << \" tc=\[\" << (end - start) << \"\]\";\n", // NOLINT
" LOG(INFO) << \"(logid=\" << log_id << \") tc=\[\" << (end - " // NOLINT
"start) << \"\]\";\n", // NOLINT
"name",
class_name,
"service",
......@@ -317,26 +322,31 @@ class PdsCodeGenerator : public CodeGenerator {
" baidu::rpc::ClosureGuard done_guard(done);\n"
" baidu::rpc::Controller* cntl = \n"
" static_cast<baidu::rpc::Controller*>(cntl_base);\n"
" uint64_t log_id = equest->log_id();\n"
" cntl->set_log_id(log_id);\n"
" ::baidu::paddle_serving::predictor::InferService* svr = \n"
" "
"::baidu::paddle_serving::predictor::InferServiceManager::instance("
").item(\"$service$\");\n"
" if (svr == NULL) {\n"
" LOG(ERROR) << \"Not found service: $service$\";\n"
" LOG(ERROR) << \"(logid=\" << log_id << \") Not found service: "
"$service$\";\n"
" cntl->SetFailed(404, \"Not found service: $service$\");\n"
" return ;\n"
" }\n"
" LOG(INFO) << \" remote_side=\[\" << cntl->remote_side() << " // NOLINT
"\"\]\";\n"
" LOG(INFO) << \" local_side=\[\" << cntl->local_side() << " // NOLINT
"\"\]\";\n"
" LOG(INFO) << \" service_name=\[\" << \"$name$\" << \"\]\";\n" // NOLINT
" LOG(INFO) << \" log_id=\[\" << cntl->log_id() << \"\]\";\n" // NOLINT
" LOG(INFO) << \"(logid=\" << log_id << \") remote_side=\[\" " // NOLINT
"<< cntl->remote_side() << \"\]\";\n"
" LOG(INFO) << \"(logid=\" << log_id << \") local_side=\[\" " // NOLINT
"<< cntl->local_side() << \"\]\";\n"
" LOG(INFO) << \"(logid=\" << log_id << \") service_name=\[\" " // NOLINT
"<< \"$name$\" << \"\]\";\n"
" butil::IOBufBuilder debug_os;\n"
" int err_code = svr->inference(request, response, &debug_os);\n"
" int err_code = svr->inference(request, response, log_id, "
"&debug_os);\n"
" if (err_code != 0) {\n"
" LOG(WARNING)\n"
" << \"Failed call inferservice[$name$], name[$service$]\"\n"
" << \"(logid=\" << log_id << \") Failed call "
"inferservice[$name$], name[$service$]\"\n"
" << \", error_code: \" << err_code;\n"
" cntl->SetFailed(err_code, \"InferService inference "
"failed!\");\n"
......@@ -345,9 +355,11 @@ class PdsCodeGenerator : public CodeGenerator {
" gettimeofday(&tv, NULL);\n"
" long end = tv.tv_sec * 1000000 + tv.tv_usec;\n"
" // flush notice log\n"
" LOG(INFO) << \" tc=\[\" << (end - start) << \"\]\";\n" // NOLINT
" LOG(INFO) << \"(logid=\" << log_id << \") tc=\[\" << (end - " // NOLINT
"start) << \"\]\";\n"
" LOG(INFO)\n"
" << \"TC=[\" << (end - start) << \"] Received debug "
" << \"(logid=\" << log_id << \") TC=[\" << (end - start) << "
"\"] Received debug "
"request[log_id=\" << cntl->log_id()\n"
" << \"] from \" << cntl->remote_side()\n"
" << \" to \" << cntl->local_side();\n",
......@@ -1011,25 +1023,31 @@ class PdsCodeGenerator : public CodeGenerator {
" brpc::ClosureGuard done_guard(done);\n"
" brpc::Controller* cntl = \n"
" static_cast<brpc::Controller*>(cntl_base);\n"
" uint64_t log_id = request->log_id();\n"
" cntl->set_log_id(log_id);\n"
" ::baidu::paddle_serving::predictor::InferService* svr = \n"
" "
"::baidu::paddle_serving::predictor::InferServiceManager::instance("
").item(\"$service$\");\n"
" if (svr == NULL) {\n"
" LOG(ERROR) << \"Not found service: $service$\";\n"
" LOG(ERROR) << \"(logid=\" << log_id << \") Not found service: "
"$service$\";\n"
" cntl->SetFailed(404, \"Not found service: $service$\");\n"
" return ;\n"
" }\n"
" LOG(INFO) << \" remote_side=\[\" << cntl->remote_side() << " // NOLINT
" LOG(INFO) << \"(logid=\" << log_id << \") "
"remote_side=\[\" << cntl->remote_side() << " // NOLINT
"\"\]\";\n"
" LOG(INFO) << \" local_side=\[\" << cntl->local_side() << " // NOLINT
" LOG(INFO) << \"(logid=\" << log_id << \") "
"local_side=\[\" << cntl->local_side() << " // NOLINT
"\"\]\";\n"
" LOG(INFO) << \" service_name=\[\" << \"$name$\" << \"\]\";\n" // NOLINT
" LOG(INFO) << \" log_id=\[\" << cntl->log_id() << \"\]\";\n" // NOLINT
" int err_code = svr->inference(request, response);\n"
" LOG(INFO) << \"(logid=\" << log_id << \") "
"service_name=\[\" << \"$name$\" << \"\]\";\n" // NOLINT
" int err_code = svr->inference(request, response, log_id);\n"
" if (err_code != 0) {\n"
" LOG(WARNING)\n"
" << \"Failed call inferservice[$name$], name[$service$]\"\n"
" << \"(logid=\" << log_id << \") Failed call "
"inferservice[$name$], name[$service$]\"\n"
" << \", error_code: \" << err_code;\n"
" cntl->SetFailed(err_code, \"InferService inference "
"failed!\");\n"
......@@ -1037,7 +1055,8 @@ class PdsCodeGenerator : public CodeGenerator {
" gettimeofday(&tv, NULL);\n"
" long end = tv.tv_sec * 1000000 + tv.tv_usec;\n"
" // flush notice log\n"
" LOG(INFO) << \" tc=\[\" << (end - start) << \"\]\";\n", // NOLINT
" LOG(INFO) << \"(logid=\" << log_id << \") tc=\[\" << (end - " // NOLINT
"start) << \"\]\";\n", // NOLINT
"name",
class_name,
"service",
......@@ -1048,26 +1067,31 @@ class PdsCodeGenerator : public CodeGenerator {
" brpc::ClosureGuard done_guard(done);\n"
" brpc::Controller* cntl = \n"
" static_cast<brpc::Controller*>(cntl_base);\n"
" uint64_t log_id = request->log_id();\n"
" cntl->set_log_id(log_id);\n"
" ::baidu::paddle_serving::predictor::InferService* svr = \n"
" "
"::baidu::paddle_serving::predictor::InferServiceManager::instance("
").item(\"$service$\");\n"
" if (svr == NULL) {\n"
" LOG(ERROR) << \"Not found service: $service$\";\n"
" LOG(ERROR) << \"(logid=\" << log_id << \") Not found service: "
"$service$\";\n"
" cntl->SetFailed(404, \"Not found service: $service$\");\n"
" return ;\n"
" }\n"
" LOG(INFO) << \" remote_side=\[\" << cntl->remote_side() << " // NOLINT
"\"\]\";\n"
" LOG(INFO) << \" local_side=\[\" << cntl->local_side() << " // NOLINT
"\"\]\";\n"
" LOG(INFO) << \" service_name=\[\" << \"$name$\" << \"\]\";\n" // NOLINT
" LOG(INFO) << \" log_id=\[\" << cntl->log_id() << \"\]\";\n" // NOLINT
" LOG(INFO) << \"(logid=\" << log_id << \") remote_side=\[\" " // NOLINT
" << cntl->remote_side() << \"\]\";\n"
" LOG(INFO) << \"(logid=\" << log_id << \") local_side=\[\" " // NOLINT
"<< cntl->local_side() << \"\]\";\n"
" LOG(INFO) << \"(logid=\" << log_id << \") service_name=\[\" " // NOLINT
"<< \"$name$\" << \"\]\";\n"
" butil::IOBufBuilder debug_os;\n"
" int err_code = svr->inference(request, response, &debug_os);\n"
" int err_code = svr->inference(request, response, log_id, "
"&debug_os);\n"
" if (err_code != 0) {\n"
" LOG(WARNING)\n"
" << \"Failed call inferservice[$name$], name[$service$]\"\n"
" << \"(logid=\" << log_id << \") Failed call "
"inferservice[$name$], name[$service$]\"\n"
" << \", error_code: \" << err_code;\n"
" cntl->SetFailed(err_code, \"InferService inference "
"failed!\");\n"
......@@ -1076,9 +1100,11 @@ class PdsCodeGenerator : public CodeGenerator {
" gettimeofday(&tv, NULL);\n"
" long end = tv.tv_sec * 1000000 + tv.tv_usec;\n"
" // flush notice log\n"
" LOG(INFO) << \" tc=\[\" << (end - start) << \"\]\";\n" // NOLINT
" LOG(INFO) << \"(logid=\" << log_id << \") tc=\[\" << (end - " // NOLINT
"start) << \"\]\";\n" // NOLINT
" LOG(INFO)\n"
" << \"TC=[\" << (end - start) << \"] Received debug "
" << \"(logid=\" << log_id << \") TC=[\" << (end - start) << "
"\"] Received debug "
"request[log_id=\" << cntl->log_id()\n"
" << \"] from \" << cntl->remote_side()\n"
" << \" to \" << cntl->local_side();\n",
......
core/predictor/framework/channel.h
浏览文件 @ b7ad237a
......@@ -72,9 +72,10 @@ class Channel {
const std::string& op() { return _op; }
int share_to_bus(Bus* bus) {
int share_to_bus(Bus* bus, const uint64_t log_id) {
if (bus->regist(_op, this) != 0) {
LOG(ERROR) << "Failed regist channel[" << _op << "] to bus!";
LOG(ERROR) << "(logid=" << log_id << ") Failed regist channel[" << _op
<< "] to bus!";
return -1;
}
......
core/predictor/framework/dag.cpp
浏览文件 @ b7ad237a
......@@ -155,13 +155,11 @@ int Dag::init(const configure::Workflow& conf, const std::string& name) {
}
if (FLAGS_el_log_level == 16) {
LOG(INFO) << "DAG: " << _dag_name;
LOG(INFO) << ", Op Num: " << _index_nodes.size();
LOG(INFO) << "DAG: " << _dag_name << ", Op Num: " << _index_nodes.size();
for (uint32_t nid = 0; nid < _index_nodes.size(); nid++) {
DagNode* node = _index_nodes[nid];
LOG(INFO) << ", OP-" << node->id << "-" << node->name << "-"
<< node->type;
LOG(INFO) << " depends: " << node->depends.size();
LOG(INFO) << "OP-" << node->id << "-" << node->name << "-" << node->type
<< " depends: " << node->depends.size();
boost::unordered_map<std::string, EdgeMode>::iterator it;
for (it = node->depends.begin(); it != node->depends.end(); it++) {
......@@ -214,7 +212,7 @@ int Dag::topo_sort() {
}
}
for (int i = 0; i < in_degree.size(); ++i) {
LOG(INFO) << "(" << _index_nodes[i]->name << ") in_degree[" << i
VLOG(2) << "(" << _index_nodes[i]->name << ") in_degree[" << i
<< "]: " << in_degree[i];
}
int sorted_num = 0;
......
core/predictor/framework/dag_view.cpp
浏览文件 @ b7ad237a
......@@ -26,7 +26,9 @@ namespace baidu {
namespace paddle_serving {
namespace predictor {
int DagView::init(Dag* dag, const std::string& service_name) {
int DagView::init(Dag* dag,
const std::string& service_name,
const uint64_t log_id) {
_name = dag->name();
_full_name = service_name + NAME_DELIMITER + dag->name();
_bus = butil::get_object<Bus>();
......@@ -36,17 +38,20 @@ int DagView::init(Dag* dag, const std::string& service_name) {
for (uint32_t si = 0; si < stage_size; si++) {
const DagStage* stage = dag->stage_by_index(si);
if (stage == NULL) {
LOG(ERROR) << "Failed get stage by index:" << si;
LOG(ERROR) << "(logid=" << log_id << ") Failed get stage by index:" << si;
return ERR_INTERNAL_FAILURE;
}
ViewStage* vstage = butil::get_object<ViewStage>();
if (vstage == NULL) {
LOG(ERROR) << "Failed get vstage from object pool"
LOG(ERROR) << "(logid=" << log_id
<< ") Failed get vstage from object pool"
<< "at:" << si;
return ERR_MEM_ALLOC_FAILURE;
}
VLOG(2) << "stage[" << si << "] name: " << stage->full_name;
VLOG(2) << "stage[" << si << "] node size: " << stage->nodes.size();
VLOG(2) << "(logid=" << log_id << ") stage[" << si
<< "] name: " << stage->full_name;
VLOG(2) << "(logid=" << log_id << ") stage[" << si
<< "] node size: " << stage->nodes.size();
vstage->full_name = service_name + NAME_DELIMITER + stage->full_name;
uint32_t node_size = stage->nodes.size();
// create tls view node
......@@ -54,31 +59,39 @@ int DagView::init(Dag* dag, const std::string& service_name) {
DagNode* node = stage->nodes[ni];
ViewNode* vnode = butil::get_object<ViewNode>();
if (vnode == NULL) {
LOG(ERROR) << "Failed get vnode at:" << ni;
LOG(ERROR) << "(logid=" << log_id << ") Failed get vnode at:" << ni;
return ERR_MEM_ALLOC_FAILURE;
}
// factory type
Op* op = OpRepository::instance().get_op(node->type);
if (op == NULL) {
LOG(ERROR) << "Failed get op with type:" << node->type;
LOG(ERROR) << "(logid=" << log_id
<< ") Failed get op with type:" << node->type;
return ERR_INTERNAL_FAILURE;
}
// initialize a TLS op object
VLOG(2) << "dag view initialized: \n"
VLOG(2) << "(logid=" << log_id << ") dag view initialized: \n"
<< "node id: " << node->id << "\n"
<< "node name: " << node->name << "\n"
<< "node type: " << node->type;
if (op->init(_bus, dag, node->id, node->name, node->type, node->conf) !=
0) {
LOG(WARNING) << "Failed init op, type:" << node->type;
if (op->init(_bus,
dag,
node->id,
node->name,
node->type,
node->conf,
log_id) != 0) {
LOG(WARNING) << "(logid=" << log_id
<< ") Failed init op, type:" << node->type;
return ERR_INTERNAL_FAILURE;
}
op->set_full_name(service_name + NAME_DELIMITER + node->full_name);
// Set the name of the Op as the key of the matching engine.
VLOG(2) << "op->set_engine_name(" << node->name.c_str() << ")";
VLOG(2) << "(logid=" << log_id << ") op->set_engine_name("
<< node->name.c_str() << ")";
op->set_engine_name(node->name);
vnode->conf = node;
......@@ -88,7 +101,7 @@ int DagView::init(Dag* dag, const std::string& service_name) {
it != vnode->conf->depends.end();
++it) {
std::string pre_node_name = it->first;
VLOG(2) << "add op pre name: \n"
VLOG(2) << "(logid=" << log_id << ") add op pre name: \n"
<< "current op name: " << vnode->op->op_name()
<< ", previous op name: " << pre_node_name;
vnode->op->add_pre_node_name(pre_node_name);
......@@ -102,7 +115,7 @@ int DagView::init(Dag* dag, const std::string& service_name) {
//<< " previous op name: "
//<< _view[si - 1]->nodes.back()->op->op_name();
// vstage->nodes.back()->op->set_pre_node_name(
//_view[si - 1]->nodes.back()->op->op_name());
// _view[si - 1]->nodes.back()->op->op_name());
/*}*/
_view.push_back(vstage);
}
......@@ -133,14 +146,15 @@ int DagView::deinit() {
return ERR_OK;
}
int DagView::execute(butil::IOBufBuilder* debug_os) {
int DagView::execute(const uint64_t log_id, butil::IOBufBuilder* debug_os) {
uint32_t stage_size = _view.size();
for (uint32_t si = 0; si < stage_size; si++) {
TRACEPRINTF("start to execute stage[%u]", si);
int errcode = execute_one_stage(_view[si], debug_os);
TRACEPRINTF("finish to execute stage[%u]", si);
TRACEPRINTF("(logid=%" PRIu64 ") start to execute stage[%u]", log_id, si);
int errcode = execute_one_stage(_view[si], log_id, debug_os);
TRACEPRINTF("(logid=%" PRIu64 ") finish to execute stage[%u]", log_id, si);
if (errcode < 0) {
LOG(ERROR) << "failed execute stage[" << _view[si]->debug();
LOG(ERROR) << "(logid=" << log_id << ") Failed execute stage["
<< _view[si]->debug();
return errcode;
}
}
......@@ -151,29 +165,34 @@ int DagView::execute(butil::IOBufBuilder* debug_os) {
// You can derive a subclass to implement this func.
// ParallelDagView maybe the one you want.
int DagView::execute_one_stage(ViewStage* vstage,
const uint64_t log_id,
butil::IOBufBuilder* debug_os) {
butil::Timer stage_time(butil::Timer::STARTED);
uint32_t node_size = vstage->nodes.size();
VLOG(2) << "vstage->nodes.size(): " << node_size;
VLOG(2) << "(logid=" << log_id << ") vstage->nodes.size(): " << node_size;
for (uint32_t ni = 0; ni < node_size; ni++) {
ViewNode* vnode = vstage->nodes[ni];
DagNode* conf = vnode->conf;
Op* op = vnode->op;
TRACEPRINTF("start to execute op[%s]", op->name());
int errcode = op->process(debug_os != NULL);
TRACEPRINTF("finish to execute op[%s]", op->name());
TRACEPRINTF(
"(logid=%" PRIu64 ") start to execute op[%s]", log_id, op->name());
int errcode = op->process(log_id, debug_os != NULL);
TRACEPRINTF(
"(logid=%" PRIu64 ") finish to execute op[%s]", log_id, op->name());
if (errcode < 0) {
LOG(ERROR) << "Execute failed, Op:" << op->debug_string();
LOG(ERROR) << "(logid=" << log_id
<< ") Execute failed, Op:" << op->debug_string();
return errcode;
}
if (errcode > 0) {
LOG(INFO) << "Execute ignore, Op:" << op->debug_string();
LOG(INFO) << "(logid=" << log_id
<< ") Execute ignore, Op:" << op->debug_string();
continue;
}
if (debug_os) {
(*debug_os) << "{\"op_name\": \"" << op->name()
(*debug_os) << "(logid=" << log_id << ") {\"op_name\": \"" << op->name()
<< "\", \"debug_str:\": \"" << op->debug_string()
<< "\", \"time_info\": \"" << op->time_info() << "\"}";
}
......@@ -186,34 +205,34 @@ int DagView::execute_one_stage(ViewStage* vstage,
return ERR_OK;
}
int DagView::set_request_channel(Channel& request) {
int DagView::set_request_channel(Channel& request, const uint64_t log_id) {
// Each workflow should get the very beginning
// request (channel), and commit it to bus, for
// the first stage ops consuming.
request.share_to_bus(_bus);
request.share_to_bus(_bus, log_id);
return ERR_OK;
}
const Channel* DagView::get_response_channel() const {
const Channel* DagView::get_response_channel(const uint64_t log_id) const {
// Caller obtains response channel from bus, and
// writes it to rpc response(protbuf/json)
if (_view.size() < 1) {
LOG(ERROR) << "invalid empty view stage!";
LOG(ERROR) << "(logid=" << log_id << ") invalid empty view stage!";
return NULL;
}
ViewStage* last_stage = _view[_view.size() - 1];
if (last_stage->nodes.size() != 1 || last_stage->nodes[0] == NULL) {
LOG(ERROR) << "Invalid last stage, size[" << last_stage->nodes.size()
<< "] != 1";
LOG(ERROR) << "(logid=" << log_id << ") Invalid last stage, size["
<< last_stage->nodes.size() << "] != 1";
return NULL;
}
Op* last_op = last_stage->nodes[0]->op;
if (last_op == NULL) {
LOG(ERROR) << "Last op is NULL";
LOG(ERROR) << "(logid=" << log_id << ") Last op is NULL";
return NULL;
}
return last_op->mutable_channel();
......
core/predictor/framework/dag_view.h
浏览文件 @ b7ad237a
......@@ -47,21 +47,22 @@ class DagView {
~DagView() {}
int init(Dag* dag, const std::string& service_name);
int init(Dag* dag, const std::string& service_name, const uint64_t log_id);
int deinit();
int execute(butil::IOBufBuilder* debug_os);
int execute(const uint64_t log_id, butil::IOBufBuilder* debug_os);
// The default execution strategy is in sequencing
// You can derive a subclass to implement this func.
// ParallelDagView maybe the one you want.
virtual int execute_one_stage(ViewStage* vstage,
const uint64_t log_id,
butil::IOBufBuilder* debug_os);
int set_request_channel(Channel& request); // NOLINT
int set_request_channel(Channel& request, const uint64_t log_id); // NOLINT
const Channel* get_response_channel() const;
const Channel* get_response_channel(const uint64_t log_id) const;
const std::string& name() const { return _name; }
......
core/predictor/framework/service.cpp
浏览文件 @ b7ad237a
......@@ -19,6 +19,7 @@
#include <butil/time.h> // butil::Timer
#endif
#include <inttypes.h>
#include <list>
#include <string>
#include <vector>
......@@ -135,50 +136,63 @@ const std::string& InferService::name() const { return _infer_service_format; }
// ´®ÐÐÖ´ÐÐÿ¸öworkflow
int InferService::inference(const google::protobuf::Message* request,
google::protobuf::Message* response,
const uint64_t log_id,
butil::IOBufBuilder* debug_os) {
TRACEPRINTF("start to inference");
TRACEPRINTF("(logid=%" PRIu64 ") start to inference", log_id);
// when funtion call begins, framework will reset
// thread local variables&resources automatically.
if (Resource::instance().thread_clear() != 0) {
LOG(ERROR) << "Failed thread clear whole resource";
LOG(ERROR) << "(logid=" << log_id << ") Failed thread clear whole resource";
return ERR_INTERNAL_FAILURE;
}
TRACEPRINTF("finish to thread clear");
TRACEPRINTF("(logid=%" PRIu64 ") finish to thread clear", log_id);
if (_enable_map_request_to_workflow) {
LOG(INFO) << "enable map request == True";
std::vector<Workflow*>* workflows = _map_request_to_workflow(request);
VLOG(2) << "(logid=" << log_id << ") enable map request == True";
std::vector<Workflow*>* workflows =
_map_request_to_workflow(request, log_id);
if (!workflows || workflows->size() == 0) {
LOG(ERROR) << "Failed to map request to workflow";
LOG(ERROR) << "(logid=" << log_id
<< ") Failed to map request to workflow";
return ERR_INTERNAL_FAILURE;
}
size_t fsize = workflows->size();
for (size_t fi = 0; fi < fsize; ++fi) {
Workflow* workflow = (*workflows)[fi];
if (workflow == NULL) {
LOG(ERROR) << "Failed to get valid workflow at: " << fi;
LOG(ERROR) << "(logid=" << log_id
<< ") Failed to get valid workflow at: " << fi;
return ERR_INTERNAL_FAILURE;
}
TRACEPRINTF("start to execute workflow[%s]", workflow->name().c_str());
int errcode = _execute_workflow(workflow, request, response, debug_os);
TRACEPRINTF("finish to execute workflow[%s]", workflow->name().c_str());
TRACEPRINTF("(logid=%" PRIu64 ") start to execute workflow[%s]",
log_id,
workflow->name().c_str());
int errcode =
_execute_workflow(workflow, request, response, log_id, debug_os);
TRACEPRINTF("(logid=%" PRIu64 ") finish to execute workflow[%s]",
log_id,
workflow->name().c_str());
if (errcode < 0) {
LOG(ERROR) << "Failed execute workflow[" << workflow->name()
<< "] in:" << name();
LOG(ERROR) << "(logid=" << log_id << ") Failed execute workflow["
<< workflow->name() << "] in:" << name();
return errcode;
}
}
} else {
LOG(INFO) << "enable map request == False";
TRACEPRINTF("start to execute one workflow");
VLOG(2) << "(logid=" << log_id << ") enable map request == False";
TRACEPRINTF("(logid=%" PRIu64 ") start to execute one workflow", log_id);
size_t fsize = _flows.size();
for (size_t fi = 0; fi < fsize; ++fi) {
TRACEPRINTF("start to execute one workflow-%lu", fi);
int errcode = execute_one_workflow(fi, request, response, debug_os);
TRACEPRINTF("finish to execute one workflow-%lu", fi);
TRACEPRINTF(
"(logid=%" PRIu64 ") start to execute one workflow-%lu", log_id, fi);
int errcode =
execute_one_workflow(fi, request, response, log_id, debug_os);
TRACEPRINTF(
"(logid=%" PRIu64 ") finish to execute one workflow-%lu", log_id, fi);
if (errcode < 0) {
LOG(ERROR) << "Failed execute 0-th workflow in:" << name();
LOG(ERROR) << "(logid=" << log_id
<< ") Failed execute 0-th workflow in:" << name();
return errcode;
}
}
......@@ -188,26 +202,30 @@ int InferService::inference(const google::protobuf::Message* request,
int InferService::debug(const google::protobuf::Message* request,
google::protobuf::Message* response,
const uint64_t log_id,
butil::IOBufBuilder* debug_os) {
return inference(request, response, debug_os);
return inference(request, response, log_id, debug_os);
}
int InferService::execute_one_workflow(uint32_t index,
const google::protobuf::Message* request,
google::protobuf::Message* response,
const uint64_t log_id,
butil::IOBufBuilder* debug_os) {
if (index >= _flows.size()) {
LOG(ERROR) << "Faield execute workflow, index: " << index
LOG(ERROR) << "(logid=" << log_id
<< ") Faield execute workflow, index: " << index
<< " >= max:" << _flows.size();
return ERR_OVERFLOW_FAILURE;
}
Workflow* workflow = _flows[index];
return _execute_workflow(workflow, request, response, debug_os);
return _execute_workflow(workflow, request, response, log_id, debug_os);
}
int InferService::_execute_workflow(Workflow* workflow,
const google::protobuf::Message* request,
google::protobuf::Message* response,
const uint64_t log_id,
butil::IOBufBuilder* debug_os) {
butil::Timer workflow_time(butil::Timer::STARTED);
// create and submit beginer channel
......@@ -215,54 +233,62 @@ int InferService::_execute_workflow(Workflow* workflow,
req_channel.init(0, START_OP_NAME);
req_channel = request;
DagView* dv = workflow->fetch_dag_view(full_name());
dv->set_request_channel(req_channel);
DagView* dv = workflow->fetch_dag_view(full_name(), log_id);
dv->set_request_channel(req_channel, log_id);
// call actual inference interface
int errcode = dv->execute(debug_os);
int errcode = dv->execute(log_id, debug_os);
if (errcode < 0) {
LOG(ERROR) << "Failed execute dag for workflow:" << workflow->name();
LOG(ERROR) << "(logid=" << log_id
<< ") Failed execute dag for workflow:" << workflow->name();
return errcode;
}
TRACEPRINTF("finish to dv execute");
TRACEPRINTF("(logid=%" PRIu64 ") finish to dv execute", log_id);
// create ender channel and copy
const Channel* res_channel = dv->get_response_channel();
const Channel* res_channel = dv->get_response_channel(log_id);
if (res_channel == NULL) {
LOG(ERROR) << "(logid=" << log_id << ") Failed get response channel";
return ERR_INTERNAL_FAILURE;
}
if (!_merger || !_merger->merge(res_channel->message(), response)) {
LOG(ERROR) << "Failed merge channel res to response";
LOG(ERROR) << "(logid=" << log_id
<< ") Failed merge channel res to response";
return ERR_INTERNAL_FAILURE;
}
TRACEPRINTF("finish to copy from");
TRACEPRINTF("(logid=%" PRIu64 ") finish to copy from", log_id);
workflow_time.stop();
LOG(INFO) << "workflow total time: " << workflow_time.u_elapsed();
LOG(INFO) << "(logid=" << log_id
<< ") workflow total time: " << workflow_time.u_elapsed();
PredictorMetric::GetInstance()->update_latency_metric(
WORKFLOW_METRIC_PREFIX + dv->full_name(), workflow_time.u_elapsed());
// return tls data to object pool
workflow->return_dag_view(dv);
TRACEPRINTF("finish to return dag view");
TRACEPRINTF("(logid=%" PRIu64 ") finish to return dag view", log_id);
return ERR_OK;
}
std::vector<Workflow*>* InferService::_map_request_to_workflow(
const google::protobuf::Message* request) {
const google::protobuf::Message* request, const uint64_t log_id) {
const google::protobuf::Descriptor* desc = request->GetDescriptor();
const google::protobuf::FieldDescriptor* field =
desc->FindFieldByName(_request_field_key);
if (field == NULL) {
LOG(ERROR) << "No field[" << _request_field_key << "] in ["
<< desc->full_name() << "].";
LOG(ERROR) << "(logid=" << log_id << ") No field[" << _request_field_key
<< "] in [" << desc->full_name() << "].";
return NULL;
}
if (field->is_repeated()) {
LOG(ERROR) << "field[" << desc->full_name() << "." << _request_field_key
<< "] is repeated.";
LOG(ERROR) << "(logid=" << log_id << ") field[" << desc->full_name() << "."
<< _request_field_key << "] is repeated.";
return NULL;
}
if (field->cpp_type() != google::protobuf::FieldDescriptor::CPPTYPE_STRING) {
LOG(ERROR) << "field[" << desc->full_name() << "." << _request_field_key
<< "] should be string";
LOG(ERROR) << "(logid=" << log_id << ") field[" << desc->full_name() << "."
<< _request_field_key << "] should be string";
return NULL;
}
const std::string& field_value =
......@@ -270,7 +296,7 @@ std::vector<Workflow*>* InferService::_map_request_to_workflow(
std::vector<Workflow*>* p_workflow =
_request_to_workflow_map.seek(field_value);
if (p_workflow == NULL) {
LOG(ERROR) << "cannot find key[" << field_value
LOG(ERROR) << "(logid=" << log_id << ") cannot find key[" << field_value
<< "] in _request_to_workflow_map";
return NULL;
}
......
core/predictor/framework/service.h
浏览文件 @ b7ad237a
......@@ -52,25 +52,29 @@ class InferService {
// Execute each workflow serially
virtual int inference(const google::protobuf::Message* request,
google::protobuf::Message* response,
const uint64_t log_id,
butil::IOBufBuilder* debug_os = NULL);
int debug(const google::protobuf::Message* request,
google::protobuf::Message* response,
const uint64_t log_id,
butil::IOBufBuilder* debug_os);
int execute_one_workflow(uint32_t index,
const google::protobuf::Message* request,
google::protobuf::Message* response,
const uint64_t log_id,
butil::IOBufBuilder* debug_os);
private:
int _execute_workflow(Workflow* workflow,
const google::protobuf::Message* request,
google::protobuf::Message* response,
const uint64_t log_id,
butil::IOBufBuilder* debug_os);
std::vector<Workflow*>* _map_request_to_workflow(
const google::protobuf::Message* request);
const google::protobuf::Message* request, const uint64_t log_id);
private:
std::vector<Workflow*> _flows;
......@@ -88,6 +92,7 @@ class ParallelInferService : public InferService {
// Execute workflows in parallel
int inference(const google::protobuf::Message* request,
google::protobuf::Message* response,
const uint64_t log_id,
butil::IOBufBuilder* debug_os) {
return 0;
}
......
core/predictor/framework/workflow.cpp
浏览文件 @ b7ad237a
......@@ -32,21 +32,22 @@ int Workflow::init(const configure::Workflow& conf) {
return 0;
}
DagView* Workflow::fetch_dag_view(const std::string& service_name) {
DagView* Workflow::fetch_dag_view(const std::string& service_name,
const uint64_t log_id) {
DagView* view = NULL;
if (_type == "Sequence") {
view = butil::get_object<DagView>();
} else if (_type == "Parallel") {
view = butil::get_object<ParallelDagView>();
} else {
LOG(ERROR) << "Unknown dag type:" << _type << "!";
LOG(ERROR) << "(logid=" << log_id << ") Unknown dag type:" << _type << "!";
return NULL;
}
if (view == NULL) {
LOG(ERROR) << "create dag view from pool failed!";
LOG(ERROR) << "(logid=" << log_id << ") create dag view from pool failed!";
return NULL;
}
view->init(&_dag, service_name);
view->init(&_dag, service_name, log_id);
return view;
}
......
core/predictor/framework/workflow.h
浏览文件 @ b7ad237a
......@@ -36,7 +36,8 @@ class Workflow {
// different apps.
int init(const configure::Workflow& conf);
DagView* fetch_dag_view(const std::string& service_name);
DagView* fetch_dag_view(const std::string& service_name,
const uint64_t log_id);
int deinit() { return 0; }
......
core/predictor/op/op.cpp
浏览文件 @ b7ad237a
......@@ -35,7 +35,8 @@ int Op::init(Bus* bus,
uint32_t id,
const std::string& name,
const std::string& type,
void* conf) {
void* conf,
const uint64_t log_id) {
_bus = bus;
_dag = dag;
_id = id;
......@@ -45,7 +46,8 @@ int Op::init(Bus* bus,
_timer = butil::get_object<TimerFlow>();
if (!_timer) {
LOG(ERROR) << "Invalid timerflow in op:" << this->name();
LOG(ERROR) << "(logid=" << log_id
<< ") Invalid timerflow in op:" << this->name();
return -1;
}
......@@ -55,7 +57,8 @@ int Op::init(Bus* bus,
Channel* channel = mutable_channel();
if (channel == NULL) {
LOG(ERROR) << "Failed mutable channel in op: " << this->id() << ", "
LOG(ERROR) << "(logid=" << log_id
<< ") Failed mutable channel in op: " << this->id() << ", "
<< this->name() << "!";
return -1;
}
......@@ -96,18 +99,20 @@ int Op::check_time(const char* tag) {
return 0;
}
int Op::process(bool debug) {
int Op::process(const uint64_t log_id, bool debug) {
butil::Timer op_time(butil::Timer::STARTED);
if (debug && _timer) {
_timer->start();
}
if (!_has_init) {
LOG(ERROR) << "Make sure op has been init before inference";
LOG(ERROR) << "(logid=" << log_id
<< ") Make sure op has been init before inference";
return ERR_INTERNAL_FAILURE;
}
if (_has_calc) {
LOG(INFO) << "Op: " << _name << " already processed before";
LOG(INFO) << "(logid=" << log_id << ") Op: " << _name
<< " already processed before";
return ERR_OK;
}
......@@ -143,7 +148,7 @@ int Op::process(bool debug) {
// 3. share output to bus
Channel* channel = mutable_channel();
channel->share_to_bus(_bus);
channel->share_to_bus(_bus, log_id);
// 4. mark has calculated
_has_calc = true;
......@@ -156,7 +161,8 @@ int Op::process(bool debug) {
op_time.stop();
PredictorMetric::GetInstance()->update_latency_metric(
OP_METRIC_PREFIX + full_name(), op_time.u_elapsed());
LOG(INFO) << " " << name() << "_time=[" << op_time.u_elapsed() << "]";
LOG(INFO) << "(logid=" << log_id << ") " << name() << "_time=["
<< op_time.u_elapsed() << "]";
return ERR_OK;
}
......
core/predictor/op/op.h
浏览文件 @ b7ad237a
......@@ -113,13 +113,14 @@ class Op {
uint32_t id,
const std::string& name,
const std::string& type,
void* conf);
void* conf,
const uint64_t log_id);
int deinit();
int check_time(const char* tag);
int process(bool debug);
int process(const uint64_t log_id, bool debug);
std::string time_info();
......
core/sdk-cpp/proto/general_model_service.proto
浏览文件 @ b7ad237a
......@@ -37,6 +37,7 @@ message Request {
repeated FeedInst insts = 1;
repeated string fetch_var_names = 2;
optional bool profile_server = 3 [ default = false ];
required uint64 log_id = 4 [ default = 0 ];
};
message Response {
......
doc/PIPELINE_SERVING.md
浏览文件 @ b7ad237a
......@@ -33,6 +33,7 @@ The graph execution engine consists of OPs and Channels, and the connected OPs s
- The default function of a single OP is to access a single Paddle Serving Service based on the input Channel data and put the result into the output Channel.
- OP supports user customization, including preprocess, process, postprocess functions that can be inherited and implemented by the user.
- OP can set the number of concurrencies to increase the number of concurrencies processed.
- OP can obtain data from multiple different RPC requests for Auto-Batching.
- OP can be started by a thread or process.
### Channel Design
......@@ -46,6 +47,7 @@ The graph execution engine consists of OPs and Channels, and the connected OPs s
</center>
### Extreme Case Consideration
- Request timeout
......@@ -59,9 +61,9 @@ The graph execution engine consists of OPs and Channels, and the connected OPs s
- Whether input buffers and output buffers in Channel will increase indefinitely
- It will not increase indefinitely. The input to the entire graph execution engine is placed inside a Channel's internal queue, directly acting as a traffic control buffer queue for the entire service.
- For input buffer, adjust the number of concurrencies of OP1 and OP2 according to the amount of computation, so that the number of input buffers from each input OP is relatively balanced.
- For output buffer, you can use a similar process as input buffer, which adjusts the concurrency of OP3 and OP4 to control the buffer length of output buffer.
- Note: The length of the input buffer depends on the speed at which each item in the internal queue is ready, and the length of the output buffer depends on the speed at which downstream OPs obtain data from the output buffer.
- For input buffer, adjust the number of concurrencies of OP1 and OP2 according to the amount of computation, so that the number of input buffers from each input OP is relatively balanced. (The length of the input buffer depends on the speed at which each item in the internal queue is ready)
- For output buffer, you can use a similar process as input buffer, which adjusts the concurrency of OP3 and OP4 to control the buffer length of output buffer. (The length of the output buffer depends on the speed at which downstream OPs obtain data from the output buffer)
- The amount of data in the Channel will not exceed `worker_num` of gRPC, that is, it will not exceed the thread pool size.
## Detailed Design
......@@ -79,31 +81,36 @@ def __init__(name=None,
client_config=None,
concurrency=1,
timeout=-1,
retry=1)
retry=1,
batch_size=1,
auto_batching_timeout=None)
```
The meaning of each parameter is as follows:
| Parameter | Meaning |
| :--------------: | :----------------------------------------------------------: |
| :-------------------: | :----------------------------------------------------------: |
| name | (str) String used to identify the OP type, which must be globally unique. |
| input_ops | (list) A list of all previous OPs of the current Op. |
| server_endpoints | (list) List of endpoints for remote Paddle Serving Service. If this parameter is not set, the OP will not access the remote Paddle Serving Service, that is, the process operation will not be performed. |
| fetch_list | (list) List of fetch variable names for remote Paddle Serving Service. |
| client_config | (str) The path of the client configuration file corresponding to the Paddle Serving Service. |
| concurrency | (int) The number of concurrent OPs. |
| timeout | (int) The timeout time of the process operation, in seconds. If the value is less than zero, no timeout is considered. |
| timeout | (int) The timeout time of the process operation, in ms. If the value is less than zero, no timeout is considered. |
| retry | (int) Timeout number of retries. When the value is 1, no retries are made. |
| batch_size | (int) The expected batch_size of Auto-Batching, since building batches may time out, the actual batch_size may be less than the set value. |
| auto_batching_timeout | (float) Timeout for building batches of Auto-Batching (the unit is ms). |
#### 2. General OP Secondary Development Interface
| Interface or Variable | Explain |
| :--------------------------------------------: | :----------------------------------------------------------: |
| def preprocess(self, input_dicts) | Process the data obtained from the channel, and the processed data will be used as the input of the **process** function. |
| def process(self, feed_dict) | The RPC prediction process is based on the Paddle Serving Client, and the processed data will be used as the input of the **postprocess** function. |
| def postprocess(self, input_dicts, fetch_dict) | After processing the prediction results, the processed data will be put into the subsequent Channel to be obtained by the subsequent OP. |
| :----------------------------------------------: | :----------------------------------------------------------: |
| def preprocess(self, input_dicts) | Process the data obtained from the channel, and the processed data will be used as the input of the **process** function. (This function handles a **sample**) |
| def process(self, feed_dict_list, typical_logid) | The RPC prediction process is based on the Paddle Serving Client, and the processed data will be used as the input of the **postprocess** function. (This function handles a **batch**) |
| def postprocess(self, input_dicts, fetch_dict) | After processing the prediction results, the processed data will be put into the subsequent Channel to be obtained by the subsequent OP. (This function handles a **sample**) |
| def init_op(self) | Used to load resources (such as word dictionary). |
| self.concurrency_idx | Concurrency index of current thread / process (different kinds of OP are calculated separately). |
| self.concurrency_idx | Concurrency index of current process(not thread) (different kinds of OP are calculated separately). |
In a running cycle, OP will execute three operations: preprocess, process, and postprocess (when the `server_endpoints` parameter is not set, the process operation is not executed). Users can rewrite these three functions. The default implementation is as follows:
......@@ -117,24 +124,28 @@ def preprocess(self, input_dicts):
(_, input_dict), = input_dicts.items()
return input_dict
def process(self, feed_dict):
err, err_info = ChannelData.check_npdata(feed_dict)
def process(self, feed_dict_list, typical_logid):
err, err_info = ChannelData.check_batch_npdata(feed_dict_list)
if err != 0:
raise NotImplementedError(
"{} Please override preprocess func.".format(err_info))
call_result = self.client.predict(
feed=feed_dict, fetch=self._fetch_names)
feed=feed_dict_list, fetch=self._fetch_names, log_id=typical_logid)
if isinstance(self.client, MultiLangClient):
if call_result is None or call_result["serving_status_code"] != 0:
return None
call_result.pop("serving_status_code")
return call_result
def postprocess(self, input_dicts, fetch_dict):
return fetch_dict
```
The parameter of **preprocess** is the data `input_dicts` in the previous Channel. This variable is a dictionary with the name of the previous OP as key and the output of the corresponding OP as value.
The parameter of **preprocess** is the data `input_dicts` in the previous Channel. This variable (as a **sample**) is a dictionary with the name of the previous OP as key and the output of the corresponding OP as value.
The parameter of **process** is the input variable `fetch_dict` (the return value of the preprocess function) of the Paddle Serving Client prediction interface. This variable is a dictionary with feed_name as the key and the data in the ndarray format as the value.
The parameter of **process** is the input variable `fetch_dict_list` (a list of the return value of the preprocess function) of the Paddle Serving Client prediction interface. This variable (as a **batch**) is a list of dictionaries with feed_name as the key and the data in the ndarray format as the value. `typical_logid` is used as the logid that penetrates to PaddleServingService.
The parameters of **postprocess** are `input_dicts` and `fetch_dict`. `input_dicts` is consistent with the parameter of preprocess, and `fetch_dict` is the return value of the process function (if process is not executed, this value is the return value of preprocess).
The parameters of **postprocess** are `input_dicts` and `fetch_dict`. `input_dicts` is consistent with the parameter of preprocess, and `fetch_dict` (as a **sample**) is a sample of the return batch of the process function (if process is not executed, this value is the return value of preprocess).
Users can also rewrite the **init_op** function to load some custom resources (such as word dictionary). The default implementation is as follows:
......@@ -143,7 +154,7 @@ def init_op(self):
pass
```
It should be noted that in the threaded version of OP, each OP will only call this function once, so the loaded resources must be thread safe.
It should be **noted** that in the threaded version of OP, each OP will only call this function once, so the loaded resources must be thread safe.
#### 3. RequestOp Definition
......@@ -248,6 +259,8 @@ dag:
client_type: brpc # Use brpc or grpc client. The default is brpc
retry: 1 # The number of times DAG executor retries after failure. The default value is 1, that is, no retrying
use_profile: false # Whether to print the log on the server side. The default is false
tracer:
interval_s: 600 # Monitoring time interval of Tracer (in seconds). Do not start monitoring when the value is less than 1. The default value is -1
```
......@@ -282,14 +295,8 @@ from paddle_serving_server.pipeline import PipelineServer
from paddle_serving_server.pipeline.proto import pipeline_service_pb2
from paddle_serving_server.pipeline.channel import ChannelDataEcode
import numpy as np
import logging
from paddle_serving_app.reader import IMDBDataset
logging.basicConfig(level=logging.DEBUG)
_LOGGER = logging.getLogger()
class ImdbRequestOp(RequestOp):
def init_op(self):
self.imdb_dataset = IMDBDataset()
......@@ -390,15 +397,23 @@ dag:
use_profile: true
```
After the function is enabled, the server will print the corresponding log information to the standard output in the process of prediction. In order to show the time consumption of each stage more intuitively, scripts are provided for further analysis and processing of log files.
After the function is enabled, the server will print the corresponding log information to the standard output in the process of prediction. In order to show the time consumption of each stage more intuitively, Analyst module is provided for further analysis and processing of log files.
The output of the server is first saved to a file. Taking profile as an example, the script converts the time monitoring information in the log into JSON format and saves it to the trace file. The trace file can be visualized through the tracing function of Chrome browser.
The output of the server is first saved to a file. Taking `profile.txt` as an example, the script converts the time monitoring information in the log into JSON format and saves it to the `trace` file. The `trace` file can be visualized through the tracing function of Chrome browser.
```shell
python timeline_trace.py profile trace
from paddle_serving_server.pipeline import Analyst
import json
import sys
if __name__ == "__main__":
log_filename = "profile.txt"
trace_filename = "trace"
analyst = Analyst(log_filename)
analyst.save_trace(trace_filename)
```
Specific operation: open Chrome browser, input in the address bar `chrome://tracing/` , jump to the tracing page, click the load button, open the saved trace file, and then visualize the time information of each stage of the prediction service.
Specific operation: open Chrome browser, input in the address bar `chrome://tracing/` , jump to the tracing page, click the load button, open the saved `trace` file, and then visualize the time information of each stage of the prediction service.
### Output profile information on client side
......
doc/PIPELINE_SERVING_CN.md
浏览文件 @ b7ad237a
......@@ -6,6 +6,7 @@ Paddle Serving 通常用于单模型的一键部署，但端到端的深度学
Paddle Serving 提供了用户友好的多模型组合服务编程框架，Pipeline Serving，旨在降低编程门槛，提高资源使用率（尤其是GPU设备），提升整体的预估效率。
## 整体架构设计
Server端基于 gRPC 和图执行引擎构建，两者的关系如下图所示。
......@@ -30,9 +31,10 @@ Server端基于 gRPC 和图执行引擎构建，两者的关系如下图所示
### OP的设计
- 单个OP默认的功能是根据输入的 Channel 数据，访问一个 Paddle Serving 的单模型服务，并将结果存在输出的 Channel
- 单个 OP 默认的功能是根据输入的 Channel 数据，访问一个 Paddle Serving 的单模型服务，并将结果存在输出的 Channel
- 单个 OP 可以支持用户自定义，包括 preprocess，process，postprocess 三个函数都可以由用户继承和实现
- 单个 OP 可以控制并发数，从而增加处理并发数
- 单个 OP 可以获取多个不同 RPC 请求的数据，以实现 Auto-Batching
- OP 可以由线程或进程启动
### Channel的设计
......@@ -59,11 +61,9 @@ Server端基于 gRPC 和图执行引擎构建，两者的关系如下图所示
- Channel 设计中的 input buffer 和 output buffer 是否会无限增加
- 不会。整个图执行引擎的输入会放到一个 Channel 的 internal queue 里面，直接作为整个服务的流量控制缓冲队列
- 对于 input buffer，根据计算量的情况调整 OP1 和 OP2 的并发数，使得 input buffer 来自各个输入 OP 的数量相对平衡
- 对于 output buffer，可以采用和 input buffer 类似的处理方法，即调整 OP3 和 OP4 的并发数，使得 output buffer 的缓冲长度得到控制
- 注：input buffer 的长度取决于 internal queue 中每个 item 完全 ready 的速度，output buffer 的长度取决于下游 OP 从 output buffer 获取数据的速度
## 详细设计
- 对于 input buffer，根据计算量的情况调整 OP1 和 OP2 的并发数，使得 input buffer 来自各个输入 OP 的数量相对平衡（input buffer 的长度取决于 internal queue 中每个 item 完全 ready 的速度）
- 对于 output buffer，可以采用和 input buffer 类似的处理方法，即调整 OP3 和 OP4 的并发数，使得 output buffer 的缓冲长度得到控制（output buffer 的长度取决于下游 OP 从 output buffer 获取数据的速度）
- 同时 Channel 中数据量不会超过 gRPC 的 `worker_num`，即线程池大小
### 用户接口设计
......@@ -79,31 +79,36 @@ def __init__(name=None,
client_config=None,
concurrency=1,
timeout=-1,
retry=1)
retry=1,
batch_size=1,
auto_batching_timeout=None)
```
各参数含义如下
| 参数名 | 含义 |
| :--------------: | :----------------------------------------------------------: |
| :-------------------: | :----------------------------------------------------------: |
| name | （str）用于标识 OP 类型的字符串，该字段必须全局唯一。 |
| input_ops | （list）当前 OP 的所有前继 OP 的列表。 |
| server_endpoints | （list）远程 Paddle Serving Service 的 endpoints 列表。如果不设置该参数，则不访问远程 Paddle Serving Service，即 不会执行 process 操作。 |
| fetch_list | （list）远程 Paddle Serving Service 的 fetch 列表。 |
| client_config | （str）Paddle Serving Service 对应的 Client 端配置文件路径。 |
| concurrency | （int）OP 的并发数。 |
| timeout | （int）process 操作的超时时间，单位为秒。若该值小于零，则视作不超时。 |
| timeout | （int）process 操作的超时时间，单位为毫秒。若该值小于零，则视作不超时。 |
| retry | （int）超时重试次数。当该值为 1 时，不进行重试。 |
| batch_size | （int）进行 Auto-Batching 的期望 batch_size 大小，由于构建 batch 可能超时，实际 batch_size 可能小于设定值。 |
| auto_batching_timeout | （float）进行 Auto-Batching 构建 batch 的超时时间，单位为毫秒。 |
#### 2. 普通 OP二次开发接口
| 变量或接口 | 说明 |
| :--------------------------------------------: | :----------------------------------------------------------: |
| def preprocess(self, input_dicts) | 对从 Channel 中获取的数据进行处理，处理完的数据将作为 **process** 函数的输入。 |
| def process(self, feed_dict) | 基于 Paddle Serving Client 进行 RPC 预测，处理完的数据将作为 **postprocess** 函数的输入。 |
| def postprocess(self, input_dicts, fetch_dict) | 处理预测结果，处理完的数据将被放入后继 Channel 中，以被后继 OP 获取。 |
| :----------------------------------------------: | :----------------------------------------------------------: |
| def preprocess(self, input_dicts) | 对从 Channel 中获取的数据进行处理，处理完的数据将作为 **process** 函数的输入。（该函数对一个 **sample** 进行处理） |
| def process(self, feed_dict_list, typical_logid) | 基于 Paddle Serving Client 进行 RPC 预测，处理完的数据将作为 **postprocess** 函数的输入。（该函数对一个 **batch** 进行处理） |
| def postprocess(self, input_dicts, fetch_dict) | 处理预测结果，处理完的数据将被放入后继 Channel 中，以被后继 OP 获取。（该函数对一个 **sample** 进行处理） |
| def init_op(self) | 用于加载资源（如字典等）。 |
| self.concurrency_idx | 当前线程（进程）的并发数索引（不同种类的 OP 单独计算）。 |
| self.concurrency_idx | 当前进程（非线程）的并发数索引（不同种类的 OP 单独计算）。 |
OP 在一个运行周期中会依次执行 preprocess，process，postprocess 三个操作（当不设置 `server_endpoints` 参数时，不执行 process 操作），用户可以对这三个函数进行重写，默认实现如下：
......@@ -117,25 +122,28 @@ def preprocess(self, input_dicts):
(_, input_dict), = input_dicts.items()
return input_dict
def process(self, feed_dict):
err, err_info = ChannelData.check_npdata(feed_dict)
def process(self, feed_dict_list, typical_logid):
err, err_info = ChannelData.check_batch_npdata(feed_dict_list)
if err != 0:
raise NotImplementedError(
"{} Please override preprocess func.".format(err_info))
call_result = self.client.predict(
feed=feed_dict, fetch=self._fetch_names)
feed=feed_dict_list, fetch=self._fetch_names, log_id=typical_logid)
if isinstance(self.client, MultiLangClient):
if call_result is None or call_result["serving_status_code"] != 0:
return None
call_result.pop("serving_status_code")
return call_result
def postprocess(self, input_dicts, fetch_dict):
return fetch_dict
```
**preprocess** 的参数是前继 Channel 中的数据 `input_dicts`，该变量（作为一个 **sample**）是一个以前继 OP 的 name 为 Key，对应 OP 的输出为 Value 的字典。
**preprocess** 的参数是前继 Channel 中的数据 `input_dicts`，该变量是一个以前继 OP 的 name 为 Key，对应 OP 的输出为 Value 的字典。
**process** 的参数是 Paddle Serving Client 预测接口的输入变量 `fetch_dict`（preprocess 函数的返回值），该变量是一个以 feed_name 为 Key，对应 ndarray 格式的数据为 Value 的字典。
**process** 的参数是 Paddle Serving Client 预测接口的输入变量 `fetch_dict_list`（preprocess 函数的返回值的列表），该变量（作为一个 **batch**）是一个列表，列表中的元素为以 feed_name 为 Key，对应 ndarray 格式的数据为 Value 的字典。`typical_logid` 作为向 PaddleServingService 穿透的 logid。
**postprocess** 的参数是 `input_dicts` 和 `fetch_dict`，`input_dicts` 与 preprocess 的参数一致，`fetch_dict` 是 process 函数的返回值（如果没有执行 process ，则该值为 preprocess 的返回值）。
**postprocess** 的参数是 `input_dicts` 和 `fetch_dict`，`input_dicts` 与 preprocess 的参数一致，`fetch_dict` （作为一个 **sample**）是 process 函数的返回 batch 中的一个 sample（如果没有执行 process ，则该值为 preprocess 的返回值）。
用户还可以对 **init_op** 函数进行重写，已加载自定义的一些资源（比如字典等），默认实现如下：
......@@ -144,7 +152,7 @@ def init_op(self):
pass
```
需要注意的是，在线程版 OP 中，每个 OP 只会调用一次该函数，故加载的资源必须要求是线程安全的。
需要**注意**的是，在线程版 OP 中，每个 OP 只会调用一次该函数，故加载的资源必须要求是线程安全的。
#### 3. RequestOp 定义
......@@ -249,6 +257,8 @@ dag:
client_type: brpc # 使用 brpc 或 grpc client，默认为 brpc
retry: 1 # DAG Executor 在失败后重试次数，默认为 1，即不重试
use_profile: false # 是否在 Server 端打印日志，默认为 false
tracer:
interval_s: 600 # Tracer 监控的时间间隔，单位为秒。当该值小于 1 时不启动监控，默认为 -1
```
......@@ -283,14 +293,8 @@ from paddle_serving_server.pipeline import PipelineServer
from paddle_serving_server.pipeline.proto import pipeline_service_pb2
from paddle_serving_server.pipeline.channel import ChannelDataEcode
import numpy as np
import logging
from paddle_serving_app.reader import IMDBDataset
logging.basicConfig(level=logging.DEBUG)
_LOGGER = logging.getLogger()
class ImdbRequestOp(RequestOp):
def init_op(self):
self.imdb_dataset = IMDBDataset()
......@@ -311,7 +315,6 @@ class CombineOp(Op):
def preprocess(self, input_data):
combined_prediction = 0
for op_name, data in input_data.items():
_LOGGER.info("{}: {}".format(op_name, data["prediction"]))
combined_prediction += data["prediction"]
data = {"prediction": combined_prediction / 2}
return data
......@@ -391,15 +394,23 @@ dag:
use_profile: true
```
开启该功能后，Server 端在预测的过程中会将对应的日志信息打印到标准输出，为了更直观地展现各阶段的耗时，提供脚本对日志文件做进一步的分析处理。
开启该功能后，Server 端在预测的过程中会将对应的日志信息打印到标准输出，为了更直观地展现各阶段的耗时，提供 Analyst 模块对日志文件做进一步的分析处理。
使用时先将 Server 的输出保存到文件，以 profile 为例，脚本将日志中的时间打点信息转换成 json 格式保存到trace 文件，trace 文件可以通过 chrome 浏览器的 tracing 功能进行可视化。
使用时先将 Server 的输出保存到文件，以 `profile.txt` 为例，脚本将日志中的时间打点信息转换成 json 格式保存到 `trace` 文件，`trace` 文件可以通过 chrome 浏览器的 tracing 功能进行可视化。
```shell
python timeline_trace.py profile trace
```python
from paddle_serving_server.pipeline import Analyst
import json
import sys
if __name__ == "__main__":
log_filename = "profile.txt"
trace_filename = "trace"
analyst = Analyst(log_filename)
analyst.save_trace(trace_filename)
```
具体操作：打开 chrome 浏览器，在地址栏输入 chrome://tracing/ ，跳转至 tracing 页面，点击 load 按钮，打开保存的 trace 文件，即可将预测服务的各阶段时间信息可视化。
具体操作：打开 chrome 浏览器，在地址栏输入 `chrome://tracing/` ，跳转至 tracing 页面，点击 load 按钮，打开保存的 `trace` 文件，即可将预测服务的各阶段时间信息可视化。
### 在 Client 端输出 Profile 信息
......
doc/pipeline_serving-image1.png
查看替换文件 @ 826a7c38 浏览文件 @ b7ad237a
doc/pipeline_serving-image1.png

96.0 KB | W: | H:

doc/pipeline_serving-image1.png

107.7 KB | W: | H:

doc/pipeline_serving-image1.png
doc/pipeline_serving-image1.png
doc/pipeline_serving-image1.png
doc/pipeline_serving-image1.png
  • 2-up
  • Swipe
  • Onion skin
java/src/main/java/io/paddle/serving/client/Client.java
浏览文件 @ b7ad237a
......@@ -192,14 +192,16 @@ public class Client {
private InferenceRequest _packInferenceRequest(
List<HashMap<String, INDArray>> feed_batch,
Iterable<String> fetch) throws IllegalArgumentException {
Iterable<String> fetch,
long log_id) throws IllegalArgumentException {
List<String> feed_var_names = new ArrayList<String>();
feed_var_names.addAll(feed_batch.get(0).keySet());
InferenceRequest.Builder req_builder = InferenceRequest.newBuilder()
.addAllFeedVarNames(feed_var_names)
.addAllFetchVarNames(fetch)
.setIsPython(false);
.setIsPython(false)
.setLogId(log_id);
for (HashMap<String, INDArray> feed_data: feed_batch) {
FeedInst.Builder inst_builder = FeedInst.newBuilder();
for (String name: feed_var_names) {
......@@ -332,76 +334,151 @@ public class Client {
public Map<String, INDArray> predict(
HashMap<String, INDArray> feed,
Iterable<String> fetch) {
return predict(feed, fetch, false);
return predict(feed, fetch, false, 0);
}
public Map<String, INDArray> predict(
HashMap<String, INDArray> feed,
Iterable<String> fetch,
long log_id) {
return predict(feed, fetch, false, log_id);
}
public Map<String, HashMap<String, INDArray>> ensemble_predict(
HashMap<String, INDArray> feed,
Iterable<String> fetch) {
return ensemble_predict(feed, fetch, false);
return ensemble_predict(feed, fetch, false, 0);
}
public Map<String, HashMap<String, INDArray>> ensemble_predict(
HashMap<String, INDArray> feed,
Iterable<String> fetch,
long log_id) {
return ensemble_predict(feed, fetch, false, log_id);
}
public PredictFuture asyn_predict(
HashMap<String, INDArray> feed,
Iterable<String> fetch) {
return asyn_predict(feed, fetch, false);
return asyn_predict(feed, fetch, false, 0);
}
public PredictFuture asyn_predict(
HashMap<String, INDArray> feed,
Iterable<String> fetch,
long log_id) {
return asyn_predict(feed, fetch, false, log_id);
}
public Map<String, INDArray> predict(
HashMap<String, INDArray> feed,
Iterable<String> fetch,
Boolean need_variant_tag) {
return predict(feed, fetch, need_variant_tag, 0);
}
public Map<String, INDArray> predict(
HashMap<String, INDArray> feed,
Iterable<String> fetch,
Boolean need_variant_tag,
long log_id) {
List<HashMap<String, INDArray>> feed_batch
= new ArrayList<HashMap<String, INDArray>>();
feed_batch.add(feed);
return predict(feed_batch, fetch, need_variant_tag);
return predict(feed_batch, fetch, need_variant_tag, log_id);
}
public Map<String, HashMap<String, INDArray>> ensemble_predict(
HashMap<String, INDArray> feed,
Iterable<String> fetch,
Boolean need_variant_tag) {
return ensemble_predict(feed, fetch, need_variant_tag, 0);
}
public Map<String, HashMap<String, INDArray>> ensemble_predict(
HashMap<String, INDArray> feed,
Iterable<String> fetch,
Boolean need_variant_tag,
long log_id) {
List<HashMap<String, INDArray>> feed_batch
= new ArrayList<HashMap<String, INDArray>>();
feed_batch.add(feed);
return ensemble_predict(feed_batch, fetch, need_variant_tag);
return ensemble_predict(feed_batch, fetch, need_variant_tag, log_id);
}
public PredictFuture asyn_predict(
HashMap<String, INDArray> feed,
Iterable<String> fetch,
Boolean need_variant_tag) {
return asyn_predict(feed, fetch, need_variant_tag, 0);
}
public PredictFuture asyn_predict(
HashMap<String, INDArray> feed,
Iterable<String> fetch,
Boolean need_variant_tag,
long log_id) {
List<HashMap<String, INDArray>> feed_batch
= new ArrayList<HashMap<String, INDArray>>();
feed_batch.add(feed);
return asyn_predict(feed_batch, fetch, need_variant_tag);
return asyn_predict(feed_batch, fetch, need_variant_tag, log_id);
}
public Map<String, INDArray> predict(
List<HashMap<String, INDArray>> feed_batch,
Iterable<String> fetch) {
return predict(feed_batch, fetch, false);
return predict(feed_batch, fetch, false, 0);
}
public Map<String, INDArray> predict(
List<HashMap<String, INDArray>> feed_batch,
Iterable<String> fetch,
long log_id) {
return predict(feed_batch, fetch, false, log_id);
}
public Map<String, HashMap<String, INDArray>> ensemble_predict(
List<HashMap<String, INDArray>> feed_batch,
Iterable<String> fetch) {
return ensemble_predict(feed_batch, fetch, false);
return ensemble_predict(feed_batch, fetch, false, 0);
}
public Map<String, HashMap<String, INDArray>> ensemble_predict(
List<HashMap<String, INDArray>> feed_batch,
Iterable<String> fetch,
long log_id) {
return ensemble_predict(feed_batch, fetch, false, log_id);
}
public PredictFuture asyn_predict(
List<HashMap<String, INDArray>> feed_batch,
Iterable<String> fetch) {
return asyn_predict(feed_batch, fetch, false);
return asyn_predict(feed_batch, fetch, false, 0);
}
public PredictFuture asyn_predict(
List<HashMap<String, INDArray>> feed_batch,
Iterable<String> fetch,
long log_id) {
return asyn_predict(feed_batch, fetch, false, log_id);
}
public Map<String, INDArray> predict(
List<HashMap<String, INDArray>> feed_batch,
Iterable<String> fetch,
Boolean need_variant_tag) {
return predict(feed_batch, fetch, need_variant_tag, 0);
}
public Map<String, INDArray> predict(
List<HashMap<String, INDArray>> feed_batch,
Iterable<String> fetch,
Boolean need_variant_tag,
long log_id) {
try {
profiler_.record("java_prepro_0");
InferenceRequest req = _packInferenceRequest(feed_batch, fetch);
InferenceRequest req = _packInferenceRequest(
feed_batch, fetch, log_id);
profiler_.record("java_prepro_1");
profiler_.record("java_client_infer_0");
......@@ -415,7 +492,7 @@ public class Client {
= new ArrayList<Map.Entry<String, HashMap<String, INDArray>>>(
ensemble_result.entrySet());
if (list.size() != 1) {
System.out.format("predict failed: please use ensemble_predict impl.\n");
System.out.format("Failed to predict: please use ensemble_predict impl.\n");
return null;
}
profiler_.record("java_postpro_1");
......@@ -423,7 +500,7 @@ public class Client {
return list.get(0).getValue();
} catch (StatusRuntimeException e) {
System.out.format("predict failed: %s\n", e.toString());
System.out.format("Failed to predict: %s\n", e.toString());
return null;
}
}
......@@ -432,9 +509,18 @@ public class Client {
List<HashMap<String, INDArray>> feed_batch,
Iterable<String> fetch,
Boolean need_variant_tag) {
return ensemble_predict(feed_batch, fetch, need_variant_tag, 0);
}
public Map<String, HashMap<String, INDArray>> ensemble_predict(
List<HashMap<String, INDArray>> feed_batch,
Iterable<String> fetch,
Boolean need_variant_tag,
long log_id) {
try {
profiler_.record("java_prepro_0");
InferenceRequest req = _packInferenceRequest(feed_batch, fetch);
InferenceRequest req = _packInferenceRequest(
feed_batch, fetch, log_id);
profiler_.record("java_prepro_1");
profiler_.record("java_client_infer_0");
......@@ -449,7 +535,7 @@ public class Client {
return ensemble_result;
} catch (StatusRuntimeException e) {
System.out.format("predict failed: %s\n", e.toString());
System.out.format("Failed to predict: %s\n", e.toString());
return null;
}
}
......@@ -458,7 +544,16 @@ public class Client {
List<HashMap<String, INDArray>> feed_batch,
Iterable<String> fetch,
Boolean need_variant_tag) {
InferenceRequest req = _packInferenceRequest(feed_batch, fetch);
return asyn_predict(feed_batch, fetch, need_variant_tag, 0);
}
public PredictFuture asyn_predict(
List<HashMap<String, INDArray>> feed_batch,
Iterable<String> fetch,
Boolean need_variant_tag,
long log_id) {
InferenceRequest req = _packInferenceRequest(
feed_batch, fetch, log_id);
ListenableFuture<InferenceResponse> future = futureStub_.inference(req);
PredictFuture predict_future = new PredictFuture(future,
(InferenceResponse resp) -> {
......
java/src/main/proto/multi_lang_general_model_service.proto
浏览文件 @ b7ad237a
......@@ -37,6 +37,7 @@ message InferenceRequest {
repeated string feed_var_names = 2;
repeated string fetch_var_names = 3;
required bool is_python = 4 [ default = false ];
required uint64 log_id = 5 [ default = 0 ];
};
message InferenceResponse {
......
python/examples/pipeline/imdb_model_ensemble/config.yml
浏览文件 @ b7ad237a
port: 18080
worker_num: 1
worker_num: 4
build_dag_each_worker: false
dag:
is_thread_op: true
is_thread_op: false
client_type: brpc
retry: 1
use_profile: false
tracer:
interval_s: 10
python/examples/pipeline/imdb_model_ensemble/test_pipeline_client.py
浏览文件 @ b7ad237a
......@@ -11,7 +11,7 @@
# 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.
from paddle_serving_client.pipeline import PipelineClient
from paddle_serving_server.pipeline import PipelineClient
import numpy as np
client = PipelineClient()
......@@ -20,12 +20,16 @@ client.connect(['127.0.0.1:18080'])
words = 'i am very sad | 0'
futures = []
for i in range(100):
for i in range(4):
futures.append(
client.predict(
feed_dict={"words": words}, fetch=["prediction"], asyn=True))
feed_dict={"words": words},
fetch=["prediction"],
asyn=True,
profile=False))
for f in futures:
res = f.result()
if res["ecode"] != 0:
print("predict failed: {}".format(res))
print(res)
python/examples/pipeline/imdb_model_ensemble/test_pipeline_server.py
浏览文件 @ b7ad237a
......@@ -12,18 +12,21 @@
# See the License for the specific language governing permissions and
# limitations under the License.
# pylint: disable=doc-string-missing
from paddle_serving_server.pipeline import Op, RequestOp, ResponseOp
from paddle_serving_server.pipeline import PipelineServer
from paddle_serving_server.pipeline.proto import pipeline_service_pb2
from paddle_serving_server.pipeline.channel import ChannelDataEcode
import numpy as np
import logging
from paddle_serving_app.reader import IMDBDataset
logging.basicConfig(level=logging.DEBUG)
import logging
_LOGGER = logging.getLogger()
user_handler = logging.StreamHandler()
user_handler.setLevel(logging.INFO)
user_handler.setFormatter(
logging.Formatter(
"%(levelname)s %(asctime)s [%(filename)s:%(lineno)d] %(message)s"))
_LOGGER.addHandler(user_handler)
class ImdbRequestOp(RequestOp):
......@@ -76,7 +79,9 @@ bow_op = Op(name="bow",
client_config="imdb_bow_client_conf/serving_client_conf.prototxt",
concurrency=1,
timeout=-1,
retry=1)
retry=1,
batch_size=3,
auto_batching_timeout=1000)
cnn_op = Op(name="cnn",
input_ops=[read_op],
server_endpoints=["127.0.0.1:9292"],
......@@ -84,13 +89,17 @@ cnn_op = Op(name="cnn",
client_config="imdb_cnn_client_conf/serving_client_conf.prototxt",
concurrency=1,
timeout=-1,
retry=1)
retry=1,
batch_size=1,
auto_batching_timeout=None)
combine_op = CombineOp(
name="combine",
input_ops=[bow_op, cnn_op],
concurrency=5,
concurrency=1,
timeout=-1,
retry=1)
retry=1,
batch_size=2,
auto_batching_timeout=None)
# fetch output of bow_op
# response_op = ImdbResponseOp(input_ops=[bow_op])
......
python/paddle_serving_client/__init__.py
浏览文件 @ b7ad237a
......@@ -233,7 +233,7 @@ class Client(object):
# key))
pass
def predict(self, feed=None, fetch=None, need_variant_tag=False):
def predict(self, feed=None, fetch=None, need_variant_tag=False, log_id=0):
self.profile_.record('py_prepro_0')
if feed is None or fetch is None:
......@@ -319,12 +319,12 @@ class Client(object):
res = self.client_handle_.numpy_predict(
float_slot_batch, float_feed_names, float_shape, int_slot_batch,
int_feed_names, int_shape, fetch_names, result_batch_handle,
self.pid)
self.pid, log_id)
elif self.has_numpy_input == False:
res = self.client_handle_.batch_predict(
float_slot_batch, float_feed_names, float_shape, int_slot_batch,
int_feed_names, int_shape, fetch_names, result_batch_handle,
self.pid)
self.pid, log_id)
else:
raise ValueError(
"Please make sure the inputs are all in list type or all in numpy.array type"
......@@ -347,6 +347,11 @@ class Client(object):
result_map[name] = result_batch_handle.get_int64_by_name(
mi, name)
shape = result_batch_handle.get_shape(mi, name)
if result_map[name].size == 0:
raise ValueError(
"Failed to fetch, maybe the type of [{}]"
" is wrong, please check the model file".format(
name))
result_map[name].shape = shape
if name in self.lod_tensor_set:
result_map["{}.lod".format(
......@@ -354,6 +359,11 @@ class Client(object):
elif self.fetch_names_to_type_[name] == float32_type:
result_map[name] = result_batch_handle.get_float_by_name(
mi, name)
if result_map[name].size == 0:
raise ValueError(
"Failed to fetch, maybe the type of [{}]"
" is wrong, please check the model file".format(
name))
shape = result_batch_handle.get_shape(mi, name)
result_map[name].shape = shape
if name in self.lod_tensor_set:
......@@ -364,6 +374,11 @@ class Client(object):
# result_map[name] will be py::array(numpy array)
result_map[name] = result_batch_handle.get_int32_by_name(
mi, name)
if result_map[name].size == 0:
raise ValueError(
"Failed to fetch, maybe the type of [{}]"
" is wrong, please check the model file".format(
name))
shape = result_batch_handle.get_shape(mi, name)
result_map[name].shape = shape
if name in self.lod_tensor_set:
......@@ -466,10 +481,11 @@ class MultiLangClient(object):
if var.is_lod_tensor:
self.lod_tensor_set_.add(var.alias_name)
def _pack_inference_request(self, feed, fetch, is_python):
def _pack_inference_request(self, feed, fetch, is_python, log_id):
req = multi_lang_general_model_service_pb2.InferenceRequest()
req.fetch_var_names.extend(fetch)
req.is_python = is_python
req.log_id = log_id
feed_batch = None
if isinstance(feed, dict):
feed_batch = [feed]
......@@ -602,12 +618,13 @@ class MultiLangClient(object):
fetch,
need_variant_tag=False,
asyn=False,
is_python=True):
is_python=True,
log_id=0):
if not asyn:
try:
self.profile_.record('py_prepro_0')
req = self._pack_inference_request(
feed, fetch, is_python=is_python)
feed, fetch, is_python=is_python, log_id=log_id)
self.profile_.record('py_prepro_1')
self.profile_.record('py_client_infer_0')
......@@ -626,7 +643,8 @@ class MultiLangClient(object):
except grpc.RpcError as e:
return {"serving_status_code": e.code()}
else:
req = self._pack_inference_request(feed, fetch, is_python=is_python)
req = self._pack_inference_request(
feed, fetch, is_python=is_python, log_id=log_id)
call_future = self.stub_.Inference.future(
req, timeout=self.rpc_timeout_s_)
return MultiLangPredictFuture(
......
python/paddle_serving_server/__init__.py
浏览文件 @ b7ad237a
......@@ -502,6 +502,7 @@ class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
feed_names = list(request.feed_var_names)
fetch_names = list(request.fetch_var_names)
is_python = request.is_python
log_id = request.log_id
feed_batch = []
for feed_inst in request.insts:
feed_dict = {}
......@@ -530,7 +531,7 @@ class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
data.shape = list(feed_inst.tensor_array[idx].shape)
feed_dict[name] = data
feed_batch.append(feed_dict)
return feed_batch, fetch_names, is_python
return feed_batch, fetch_names, is_python, log_id
def _pack_inference_response(self, ret, fetch_names, is_python):
resp = multi_lang_general_model_service_pb2.InferenceResponse()
......@@ -583,10 +584,13 @@ class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
return resp
def Inference(self, request, context):
feed_dict, fetch_names, is_python = self._unpack_inference_request(
request)
feed_dict, fetch_names, is_python, log_id = \
self._unpack_inference_request(request)
ret = self.bclient_.predict(
feed=feed_dict, fetch=fetch_names, need_variant_tag=True)
feed=feed_dict,
fetch=fetch_names,
need_variant_tag=True,
log_id=log_id)
return self._pack_inference_response(ret, fetch_names, is_python)
def GetClientConfig(self, request, context):
......
python/paddle_serving_server_gpu/__init__.py
浏览文件 @ b7ad237a
......@@ -552,6 +552,7 @@ class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
feed_names = list(request.feed_var_names)
fetch_names = list(request.fetch_var_names)
is_python = request.is_python
log_id = request.log_id
feed_batch = []
for feed_inst in request.insts:
feed_dict = {}
......@@ -580,7 +581,7 @@ class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
data.shape = list(feed_inst.tensor_array[idx].shape)
feed_dict[name] = data
feed_batch.append(feed_dict)
return feed_batch, fetch_names, is_python
return feed_batch, fetch_names, is_python, log_id
def _pack_inference_response(self, ret, fetch_names, is_python):
resp = multi_lang_general_model_service_pb2.InferenceResponse()
......@@ -633,10 +634,13 @@ class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
return resp
def Inference(self, request, context):
feed_dict, fetch_names, is_python = self._unpack_inference_request(
request)
feed_dict, fetch_names, is_python, log_id \
= self._unpack_inference_request(request)
ret = self.bclient_.predict(
feed=feed_dict, fetch=fetch_names, need_variant_tag=True)
feed=feed_dict,
fetch=fetch_names,
need_variant_tag=True,
log_id=log_id)
return self._pack_inference_response(ret, fetch_names, is_python)
def GetClientConfig(self, request, context):
......
python/pipeline/__init__.py
浏览文件 @ b7ad237a
......@@ -11,7 +11,7 @@
# 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.
import logger # this module must be the first to import
from operator import Op, RequestOp, ResponseOp
from pipeline_server import PipelineServer
from pipeline_client import PipelineClient
......
python/pipeline/analyse.py
浏览文件 @ b7ad237a
......@@ -17,7 +17,7 @@ import copy
import re
import logging
_LOGGER = logging.getLogger()
_LOGGER = logging.getLogger(__name__)
class Analyst(object):
......@@ -69,7 +69,7 @@ class Analyst(object):
with open(self._profile_file) as f:
for line in f.readlines():
line = line.strip().split("\t")
if line[0] == "PROFILE":
if line[0] == "PROFILE" and len(line) >= 3:
trace_list = self._prase_line(line[1], line[2], counter)
counter += 1
for trace in trace_list:
......@@ -164,7 +164,7 @@ class OpAnalyst(object):
def add(self, name_str, ts_list):
if self._close:
_LOGGER.error("OpAnalyst is closed.")
_LOGGER.error("Failed to add item: OpAnalyst is closed.")
return
op_name, curr_idx, step = self._parse(name_str)
if op_name not in self.op_time_list_dict:
......
python/pipeline/channel.py
浏览文件 @ b7ad237a
此差异已折叠。
python/pipeline/dag.py
浏览文件 @ b7ad237a
此差异已折叠。
python/pipeline/logger.py 0 → 100644
浏览文件 @ b7ad237a
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# 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.
import logging
import logging.config
import os
class SectionLevelFilter(object):
def __init__(self, levels):
self._levels = levels
def filter(self, logRecord):
return logRecord.levelno in self._levels
log_dir = "PipelineServingLogs"
if not os.path.exists(log_dir):
os.makedirs(log_dir)
logger_config = {
"version": 1,
"formatters": {
"normal_fmt": {
"format":
"%(levelname)s %(asctime)s [%(filename)s:%(lineno)d] %(message)s",
},
"tracer_fmt": {
"format": "%(asctime)s %(message)s",
},
},
"handlers": {
"f_pipeline.log": {
"class": "logging.FileHandler",
"level": "INFO",
"formatter": "normal_fmt",
"filename": os.path.join(log_dir, "pipeline.log"),
},
"f_pipeline.log.wf": {
"class": "logging.FileHandler",
"level": "WARNING",
"formatter": "normal_fmt",
"filename": os.path.join(log_dir, "pipeline.log.wf"),
},
"f_tracer.log": {
"class": "logging.FileHandler",
"level": "INFO",
"formatter": "tracer_fmt",
"filename": os.path.join(log_dir, "pipeline.tracer"),
},
},
"loggers": {
# propagate = True
".".join(__name__.split(".")[:-1] + ["profiler"]): {
"level": "INFO",
"handlers": ["f_tracer.log"],
},
},
"root": {
"level": "DEBUG",
"handlers": ["f_pipeline.log", "f_pipeline.log.wf"],
},
}
logging.config.dictConfig(logger_config)
python/pipeline/operator.py
浏览文件 @ b7ad237a
此差异已折叠。
python/pipeline/pipeline_client.py
浏览文件 @ b7ad237a
......@@ -18,10 +18,11 @@ import numpy as np
from numpy import *
import logging
import functools
from .channel import ChannelDataEcode
from .proto import pipeline_service_pb2
from .proto import pipeline_service_pb2_grpc
_LOGGER = logging.getLogger()
_LOGGER = logging.getLogger(__name__)
class PipelineClient(object):
......@@ -59,7 +60,11 @@ class PipelineClient(object):
def _unpack_response_package(self, resp, fetch):
if resp.ecode != 0:
return {"ecode": resp.ecode, "error_info": resp.error_info}
return {
"ecode": resp.ecode,
"ecode_desc": ChannelDataEcode(resp.ecode),
"error_info": resp.error_info,
}
fetch_map = {"ecode": resp.ecode}
for idx, key in enumerate(resp.key):
if key == self._profile_key:
......
python/pipeline/pipeline_server.py
浏览文件 @ b7ad237a
此差异已折叠。
python/pipeline/profiler.py
浏览文件 @ b7ad237a
......@@ -22,10 +22,165 @@ elif sys.version_info.major == 3:
import queue as Queue
else:
raise Exception("Error Python version")
from time import time as _time
import time
import threading
import multiprocessing
_LOGGER = logging.getLogger()
_LOGGER = logging.getLogger(__name__)
_LOGGER.propagate = False
class PerformanceTracer(object):
def __init__(self, is_thread_mode, interval_s, server_worker_num):
self._is_thread_mode = is_thread_mode
if is_thread_mode:
# Because the Channel in the thread mode cannot be
# accessed across processes, when using thread mode,
# the PerformanceTracer is also the thread mode.
# However, performance may be affected by GIL.
self._data_buffer = Queue.Queue()
else:
self._data_buffer = multiprocessing.Manager().Queue()
self._interval_s = interval_s
self._thrd = None
self._proc = None
self._channels = []
# The size of data in Channel will not exceed server_worker_num
self._server_worker_num = server_worker_num
def data_buffer(self):
return self._data_buffer
def start(self):
if self._is_thread_mode:
self._thrd = threading.Thread(
target=self._trace_func, args=(self._channels, ))
self._thrd.daemon = True
self._thrd.start()
else:
self._proc = multiprocessing.Process(
target=self._trace_func, args=(self._channels, ))
self._proc.daemon = True
self._proc.start()
def set_channels(self, channels):
self._channels = channels
def _trace_func(self, channels):
all_actions = ["in", "prep", "midp", "postp", "out"]
calcu_actions = ["prep", "midp", "postp"]
while True:
op_cost = {}
err_request = []
err_count = 0
_LOGGER.info("==================== TRACER ======================")
# op
while True:
try:
item = self._data_buffer.get_nowait()
name = item["name"]
actions = item["actions"]
if name == "DAG":
succ = item["succ"]
req_id = item["id"]
if not succ:
err_count += 1
err_request.append(req_id)
if name not in op_cost:
op_cost[name] = {}
for action, cost in actions.items():
if action not in op_cost[name]:
op_cost[name][action] = []
op_cost[name][action].append(cost)
except Queue.Empty:
break
if len(op_cost) != 0:
for name in op_cost:
tot_cost, calcu_cost = 0.0, 0.0
for action, costs in op_cost[name].items():
op_cost[name][action] = sum(costs) / (1e3 * len(costs))
tot_cost += op_cost[name][action]
if name != "DAG":
_LOGGER.info("Op({}):".format(name))
for action in all_actions:
if action in op_cost[name]:
_LOGGER.info("\t{}[{} ms]".format(
action, op_cost[name][action]))
for action in calcu_actions:
if action in op_cost[name]:
calcu_cost += op_cost[name][action]
_LOGGER.info("\tidle[{}]".format(1 - 1.0 * calcu_cost /
tot_cost))
if "DAG" in op_cost:
calls = op_cost["DAG"].values()
calls.sort()
tot = len(calls)
qps = 1.0 * tot / self._interval_s
ave_cost = sum(calls) / tot
latencys = [50, 60, 70, 80, 90, 95, 99]
_LOGGER.info("DAGExecutor:")
_LOGGER.info("\tQuery count[{}]".format(tot))
_LOGGER.info("\tQPS[{} q/s]".format(qps))
_LOGGER.info("\tSucc[{}]".format(1 - 1.0 * err_count / tot))
_LOGGER.info("\tError req[{}]".format(", ".join(
[str(x) for x in err_request])))
_LOGGER.info("\tLatency:")
_LOGGER.info("\t\tave[{} ms]".format(ave_cost))
for latency in latencys:
_LOGGER.info("\t\t.{}[{} ms]".format(latency, calls[int(
tot * latency / 100.0)]))
# channel
_LOGGER.info("Channel (server worker num[{}]):".format(
self._server_worker_num))
for channel in channels:
_LOGGER.info("\t{}(In: {}, Out: {}) size[{}/{}]".format(
channel.name,
channel.get_producers(),
channel.get_consumers(),
channel.size(), channel.get_maxsize()))
time.sleep(self._interval_s)
class UnsafeTimeProfiler(object):
""" thread unsafe profiler """
def __init__(self):
self.pid = os.getpid()
self.print_head = 'PROFILE\tpid:{}\t'.format(self.pid)
self.time_record = [self.print_head]
self._enable = False
def enable(self, enable):
self._enable = enable
def record(self, name):
if self._enable is False:
return
timestamp = int(round(_time() * 1000000))
self.time_record.append('{}:{} '.format(name, timestamp))
return timestamp
def print_profile(self):
if self._enable is False:
return
sys.stderr.write(self.gen_profile_str())
def gen_profile_str(self):
if self._enable is False:
return
self.time_record.append('\n')
profile_str = ''.join(self.time_record)
self.time_record = [self.print_head]
return profile_str
class TimeProfiler(object):
......@@ -42,12 +197,13 @@ class TimeProfiler(object):
def record(self, name_with_tag):
if self._enable is False:
return
timestamp = int(round(time.time() * 1000000))
timestamp = int(round(_time() * 1000000))
name_with_tag = name_with_tag.split("_")
tag = name_with_tag[-1]
name = '_'.join(name_with_tag[:-1])
with self._lock:
self._time_record.put((name, tag, timestamp))
return timestamp
def print_profile(self):
if self._enable is False:
......
python/pipeline/util.py
浏览文件 @ b7ad237a
此差异已折叠。
tools/serving_build.sh
浏览文件 @ b7ad237a
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或