add gpu pyserving part

python/paddle_serving_server/pyserver.py

@@ -1099,7 +1099,7 @@ class PyServer(object):
                         processed_op.add(other_op.name)
         output_channel = Channel(self._manager, name=channel_name_gen.next())
         channels.append(output_channel)
-        last_op = dag_views[-1][0]
+        last_op = dag_views[-1][0]  # TODO: fix it
         last_op.add_output_channel(output_channel)
 
         self._actual_ops = virtual_ops

python/paddle_serving_server_gpu/pyserver.py

+# 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.
+# pylint: disable=doc-string-missing
+import threading
+import multiprocessing
+import multiprocessing.queues
+import Queue
+import os
+import sys
+import paddle_serving_server
+from paddle_serving_client import MultiLangClient as Client
+from paddle_serving_client import MultiLangPredictFuture
+from concurrent import futures
+import numpy as np
+import grpc
+from .proto import general_model_config_pb2 as m_config
+from .proto import general_python_service_pb2 as pyservice_pb2
+from .proto import pyserving_channel_pb2 as channel_pb2
+from .proto import general_python_service_pb2_grpc
+import logging
+import random
+import time
+import func_timeout
+import enum
+import collections
+
+
+class _TimeProfiler(object):
+    def __init__(self):
+        self._pid = os.getpid()
+        self._print_head = 'PROFILE\tpid:{}\t'.format(self._pid)
+        self._time_record = Queue.Queue()
+        self._enable = False
+
+    def enable(self, enable):
+        self._enable = enable
+
+    def record(self, name_with_tag):
+        if self._enable is False:
+            return
+        name_with_tag = name_with_tag.split("_")
+        tag = name_with_tag[-1]
+        name = '_'.join(name_with_tag[:-1])
+        self._time_record.put((name, tag, int(round(time.time() * 1000000))))
+
+    def print_profile(self):
+        if self._enable is False:
+            return
+        sys.stderr.write(self._print_head)
+        tmp = {}
+        while not self._time_record.empty():
+            name, tag, timestamp = self._time_record.get()
+            if name in tmp:
+                ptag, ptimestamp = tmp.pop(name)
+                sys.stderr.write("{}_{}:{} ".format(name, ptag, ptimestamp))
+                sys.stderr.write("{}_{}:{} ".format(name, tag, timestamp))
+            else:
+                tmp[name] = (tag, timestamp)
+        sys.stderr.write('\n')
+        for name, item in tmp.items():
+            tag, timestamp = item
+            self._time_record.put((name, tag, timestamp))
+
+
+_profiler = _TimeProfiler()
+
+
+class ChannelDataEcode(enum.Enum):
+    OK = 0
+    TIMEOUT = 1
+    NOT_IMPLEMENTED = 2
+    TYPE_ERROR = 3
+    RPC_PACKAGE_ERROR = 4
+    UNKNOW = 5
+
+
+class ChannelDataType(enum.Enum):
+    CHANNEL_PBDATA = 0
+    CHANNEL_FUTURE = 1
+    CHANNEL_NPDATA = 2
+    ERROR = 3
+
+
+class ChannelData(object):
+    def __init__(self,
+                 datatype=None,
+                 future=None,
+                 pbdata=None,
+                 npdata=None,
+                 data_id=None,
+                 callback_func=None,
+                 ecode=None,
+                 error_info=None):
+        '''
+        There are several ways to use it:
+        
+        1. ChannelData(ChannelDataType.CHANNEL_FUTURE.value, future, pbdata[, callback_func])
+        2. ChannelData(ChannelDataType.CHANNEL_FUTURE.value, future, data_id[, callback_func])
+        3. ChannelData(ChannelDataType.CHANNEL_PBDATA.value, pbdata)
+        4. ChannelData(ChannelDataType.CHANNEL_PBDATA.value, npdata, data_id)
+        5. ChannelData(ChannelDataType.CHANNEL_NPDATA.value, npdata, data_id)
+        6. ChannelData(ecode, error_info, data_id)
+
+        Protobufs are not pickle-able:
+        https://stackoverflow.com/questions/55344376/how-to-import-protobuf-module
+        '''
+        if ecode is not None:
+            if data_id is None or error_info is None:
+                raise ValueError("data_id and error_info cannot be None")
+            datatype = ChannelDataType.ERROR.value
+        else:
+            if datatype == ChannelDataType.CHANNEL_FUTURE.value:
+                if data_id is None:
+                    raise ValueError("data_id cannot be None")
+                ecode = ChannelDataEcode.OK.value
+            elif datatype == ChannelDataType.CHANNEL_PBDATA.value:
+                if pbdata is None:
+                    if data_id is None:
+                        raise ValueError("data_id cannot be None")
+                    pbdata = channel_pb2.ChannelData()
+                    ecode, error_info = self._check_npdata(npdata)
+                    if ecode != ChannelDataEcode.OK.value:
+                        logging.error(error_info)
+                    else:
+                        for name, value in npdata.items():
+                            inst = channel_pb2.Inst()
+                            inst.data = value.tobytes()
+                            inst.name = name
+                            inst.shape = np.array(
+                                value.shape, dtype="int32").tobytes()
+                            inst.type = str(value.dtype)
+                            pbdata.insts.append(inst)
+            elif datatype == ChannelDataType.CHANNEL_NPDATA.value:
+                ecode, error_info = self._check_npdata(npdata)
+                if ecode != ChannelDataEcode.OK.value:
+                    logging.error(error_info)
+            else:
+                raise ValueError("datatype not match")
+        self.future = future
+        self.pbdata = pbdata
+        self.npdata = npdata
+        self.datatype = datatype
+        self.callback_func = callback_func
+        self.id = data_id
+        self.ecode = ecode
+        self.error_info = error_info
+
+    def _check_npdata(self, npdata):
+        ecode = ChannelDataEcode.OK.value
+        error_info = None
+        for name, value in npdata.items():
+            if not isinstance(name, (str, unicode)):
+                ecode = ChannelDataEcode.TYPE_ERROR.value
+                error_info = log("the key of postped_data must " \
+                        "be str, but get {}".format(type(name)))
+                break
+            if not isinstance(value, np.ndarray):
+                ecode = ChannelDataEcode.TYPE_ERROR.value
+                error_info = log("the value of postped_data must " \
+                        "be np.ndarray, but get {}".format(type(value)))
+                break
+        return ecode, error_info
+
+    def parse(self):
+        # return narray
+        feed = None
+        if self.datatype == ChannelDataType.CHANNEL_PBDATA.value:
+            feed = {}
+            for inst in self.pbdata.insts:
+                feed[inst.name] = np.frombuffer(inst.data, dtype=inst.type)
+                feed[inst.name].shape = np.frombuffer(inst.shape, dtype="int32")
+        elif self.datatype == ChannelDataType.CHANNEL_FUTURE.value:
+            feed = self.future.result()
+            if self.callback_func is not None:
+                feed = self.callback_func(feed)
+        elif self.datatype == ChannelDataType.CHANNEL_NPDATA.value:
+            feed = self.npdata
+        else:
+            raise TypeError("Error type({}) in datatype.".format(self.datatype))
+        return feed
+
+    def __str__(self):
+        return "type[{}], ecode[{}], id[{}]".format(
+            ChannelDataType(self.datatype).name, self.ecode, self.id)
+
+
+class Channel(multiprocessing.queues.Queue):
+    """ 
+    The channel used for communication between Ops.
+
+    1. Support multiple different Op feed data (multiple producer)
+        Different types of data will be packaged through the data ID
+    2. Support multiple different Op fetch data (multiple consumer)
+        Only when all types of Ops get the data of the same ID,
+        the data will be poped; The Op of the same type will not
+        get the data of the same ID.
+    3. (TODO) Timeout and BatchSize are not fully supported.
+
+    Note:
+    1. The ID of the data in the channel must be different.
+    2. The function add_producer() and add_consumer() are not thread safe,
+       and can only be called during initialization.
+    """
+
+    def __init__(self, manager, name=None, maxsize=0, timeout=None):
+        # https://stackoverflow.com/questions/39496554/cannot-subclass-multiprocessing-queue-in-python-3-5/
+        multiprocessing.queues.Queue.__init__(self, maxsize=maxsize)
+        self._maxsize = maxsize
+        self._timeout = timeout
+        self.name = name
+        self._stop = False
+
+        self._cv = multiprocessing.Condition()
+
+        self._producers = []
+        self._producer_res_count = manager.dict()  # {data_id: count}
+        # self._producer_res_count = {}  # {data_id: count}
+        self._push_res = manager.dict()  # {data_id: {op_name: data}}
+        # self._push_res = {}  # {data_id: {op_name: data}}
+
+        self._consumers = manager.dict()  # {op_name: idx}
+        # self._consumers = {}  # {op_name: idx}
+        self._idx_consumer_num = manager.dict()  # {idx: num}
+        # self._idx_consumer_num = {}  # {idx: num}
+        self._consumer_base_idx = manager.Value('i', 0)
+        # self._consumer_base_idx = 0
+        self._front_res = manager.list()
+        # self._front_res = []
+
+    def get_producers(self):
+        return self._producers
+
+    def get_consumers(self):
+        return self._consumers.keys()
+
+    def _log(self, info_str):
+        return "[{}] {}".format(self.name, info_str)
+
+    def debug(self):
+        return self._log("p: {}, c: {}".format(self.get_producers(),
+                                               self.get_consumers()))
+
+    def add_producer(self, op_name):
+        """ not thread safe, and can only be called during initialization. """
+        if op_name in self._producers:
+            raise ValueError(
+                self._log("producer({}) is already in channel".format(op_name)))
+        self._producers.append(op_name)
+
+    def add_consumer(self, op_name):
+        """ not thread safe, and can only be called during initialization. """
+        if op_name in self._consumers:
+            raise ValueError(
+                self._log("consumer({}) is already in channel".format(op_name)))
+        self._consumers[op_name] = 0
+
+        if self._idx_consumer_num.get(0) is None:
+            self._idx_consumer_num[0] = 0
+        self._idx_consumer_num[0] += 1
+
+    def push(self, channeldata, op_name=None):
+        logging.debug(
+            self._log("{} try to push data: {}".format(op_name,
+                                                       channeldata.__str__())))
+        if len(self._producers) == 0:
+            raise Exception(
+                self._log(
+                    "expected number of producers to be greater than 0, but the it is 0."
+                ))
+        elif len(self._producers) == 1:
+            with self._cv:
+                while self._stop is False:
+                    try:
+                        self.put(channeldata, timeout=0)
+                        break
+                    except Queue.Full:
+                        self._cv.wait()
+                logging.debug(
+                    self._log("{} channel size: {}".format(op_name,
+                                                           self.qsize())))
+                self._cv.notify_all()
+                logging.debug(self._log("{} notify all".format(op_name)))
+            logging.debug(self._log("{} push data succ!".format(op_name)))
+            return True
+        elif op_name is None:
+            raise Exception(
+                self._log(
+                    "There are multiple producers, so op_name cannot be None."))
+
+        producer_num = len(self._producers)
+        data_id = channeldata.id
+        put_data = None
+        with self._cv:
+            logging.debug(self._log("{} get lock".format(op_name)))
+            if data_id not in self._push_res:
+                self._push_res[data_id] = {
+                    name: None
+                    for name in self._producers
+                }
+                self._producer_res_count[data_id] = 0
+            # see: https://docs.python.org/3.6/library/multiprocessing.html?highlight=multiprocess#proxy-objects
+            # self._push_res[data_id][op_name] = channeldata
+            tmp_push_res = self._push_res[data_id]
+            tmp_push_res[op_name] = channeldata
+            self._push_res[data_id] = tmp_push_res
+
+            if self._producer_res_count[data_id] + 1 == producer_num:
+                put_data = self._push_res[data_id]
+                self._push_res.pop(data_id)
+                self._producer_res_count.pop(data_id)
+            else:
+                self._producer_res_count[data_id] += 1
+
+            if put_data is None:
+                logging.debug(
+                    self._log("{} push data succ, but not push to queue.".
+                              format(op_name)))
+            else:
+                while self._stop is False:
+                    try:
+                        logging.debug(
+                            self._log("{} push data succ: {}".format(
+                                op_name, put_data.__str__())))
+                        self.put(put_data, timeout=0)
+                        break
+                    except Queue.Empty:
+                        self._cv.wait()
+
+                logging.debug(
+                    self._log("multi | {} push data succ!".format(op_name)))
+            self._cv.notify_all()
+        return True
+
+    def front(self, op_name=None):
+        logging.debug(self._log("{} try to get data...".format(op_name)))
+        if len(self._consumers) == 0:
+            raise Exception(
+                self._log(
+                    "expected number of consumers to be greater than 0, but the it is 0."
+                ))
+        elif len(self._consumers) == 1:
+            resp = None
+            with self._cv:
+                while self._stop is False and resp is None:
+                    try:
+                        logging.debug(
+                            self._log("{} try to get(with channel size: {})".
+                                      format(op_name, self.qsize())))
+                        #TODO: bug to fix
+                        # (multiple processes) the queue is not empty, but it raise Queue.Empty                   
+                        resp = self.get(timeout=1e-3)
+                        break
+                    except Queue.Empty:
+                        logging.debug(
+                            self._log(
+                                "{} wait for empty queue(with channel size: {})".
+                                format(op_name, self.qsize())))
+                        self._cv.wait()
+            logging.debug(
+                self._log("{} get data succ: {}".format(op_name, resp.__str__(
+                ))))
+            return resp
+        elif op_name is None:
+            raise Exception(
+                self._log(
+                    "There are multiple consumers, so op_name cannot be None."))
+
+        with self._cv:
+            # data_idx = consumer_idx - base_idx
+            while self._stop is False and self._consumers[
+                    op_name] - self._consumer_base_idx.value >= len(
+                        self._front_res):
+                logging.debug(
+                    self._log(
+                        "({}) B self._consumers: {}, self._consumer_base_idx: {}, len(self._front_res): {}".
+                        format(op_name, self._consumers, self.
+                               _consumer_base_idx.value, len(self._front_res))))
+                try:
+                    logging.debug(
+                        self._log("{} try to get(with channel size: {})".format(
+                            op_name, self.qsize())))
+                    #TODO: bug to fix
+                    # (multiple processes) the queue is not empty, but it raise Queue.Empty
+                    channeldata = self.get(timeout=1e-3)
+                    self._front_res.append(channeldata)
+                    break
+                except Queue.Empty:
+                    logging.debug(
+                        self._log(
+                            "{} wait for empty queue(with channel size: {})".
+                            format(op_name, self.qsize())))
+                    self._cv.wait()
+
+            consumer_idx = self._consumers[op_name]
+            base_idx = self._consumer_base_idx.value
+            data_idx = consumer_idx - base_idx
+            resp = self._front_res[data_idx]
+            logging.debug(self._log("{} get data: {}".format(op_name, resp)))
+
+            self._idx_consumer_num[consumer_idx] -= 1
+            if consumer_idx == base_idx and self._idx_consumer_num[
+                    consumer_idx] == 0:
+                self._idx_consumer_num.pop(consumer_idx)
+                self._front_res.pop(0)
+                self._consumer_base_idx.value += 1
+
+            self._consumers[op_name] += 1
+            new_consumer_idx = self._consumers[op_name]
+            if self._idx_consumer_num.get(new_consumer_idx) is None:
+                self._idx_consumer_num[new_consumer_idx] = 0
+            self._idx_consumer_num[new_consumer_idx] += 1
+            logging.debug(
+                self._log(
+                    "({}) A self._consumers: {}, self._consumer_base_idx: {}, len(self._front_res): {}".
+                    format(op_name, self._consumers, self._consumer_base_idx.
+                           value, len(self._front_res))))
+            logging.debug(self._log("{} notify all".format(op_name)))
+            self._cv.notify_all()
+
+        logging.debug(self._log("multi | {} get data succ!".format(op_name)))
+        return resp  # reference, read only
+
+    def stop(self):
+        #TODO
+        self.close()
+        self._stop = True
+        self._cv.notify_all()
+
+
+class Op(object):
+    def __init__(self,
+                 name,
+                 inputs,
+                 server_model=None,
+                 server_port=None,
+                 device=None,
+                 client_config=None,
+                 server_name=None,
+                 fetch_names=None,
+                 concurrency=1,
+                 timeout=-1,
+                 retry=2):
+        self._is_run = False
+        self.name = name  # to identify the type of OP, it must be globally unique
+        self._concurrency = concurrency  # amount of concurrency
+        self.set_input_ops(inputs)
+        self._timeout = timeout
+        self._retry = max(1, retry)
+        self._input = None
+        self._outputs = []
+
+        self.with_serving = False
+        self._client_config = client_config
+        self._server_name = server_name
+        self._fetch_names = fetch_names
+        self._server_model = server_model
+        self._server_port = server_port
+        self._device = device
+        if self._client_config is not None and \
+                self._server_name is not None and \
+                self._fetch_names is not None and \
+                self._server_model is not None and \
+                self._server_port is not None and \
+                self._device is not None:
+            self.with_serving = True
+
+    def init_client(self, client_config, server_name, fetch_names):
+        if self.with_serving == False:
+            logging.debug("{} no client".format(self.name))
+            return
+        logging.debug("{} client_config: {}".format(self.name, client_config))
+        logging.debug("{} server_name: {}".format(self.name, server_name))
+        logging.debug("{} fetch_names: {}".format(self.name, fetch_names))
+        self._client = Client()
+        self._client.load_client_config(client_config)
+        self._client.connect([server_name])
+        self._fetch_names = fetch_names
+
+    def get_input_channel(self):
+        return self._input
+
+    def get_input_ops(self):
+        return self._input_ops
+
+    def set_input_ops(self, ops):
+        if not isinstance(ops, list):
+            ops = [] if ops is None else [ops]
+        self._input_ops = []
+        for op in ops:
+            if not isinstance(op, Op):
+                raise TypeError(
+                    self._log('input op must be Op type, not {}'.format(
+                        type(op))))
+            self._input_ops.append(op)
+
+    def add_input_channel(self, channel):
+        if not isinstance(channel, Channel):
+            raise TypeError(
+                self._log('input channel must be Channel type, not {}'.format(
+                    type(channel))))
+        channel.add_consumer(self.name)
+        self._input = channel
+
+    def get_output_channels(self):
+        return self._outputs
+
+    def add_output_channel(self, channel):
+        if not isinstance(channel, Channel):
+            raise TypeError(
+                self._log('output channel must be Channel type, not {}'.format(
+                    type(channel))))
+        channel.add_producer(self.name)
+        self._outputs.append(channel)
+
+    def preprocess(self, channeldata):
+        if isinstance(channeldata, dict):
+            raise NotImplementedError(
+                'this Op has multiple previous inputs. Please override this method'
+            )
+        feed = channeldata.parse()
+        return feed
+
+    def midprocess(self, data):
+        if not isinstance(data, dict):
+            raise Exception(
+                self._log(
+                    'data must be dict type(the output of preprocess()), but get {}'.
+                    format(type(data))))
+        logging.debug(self._log('data: {}'.format(data)))
+        logging.debug(self._log('fetch: {}'.format(self._fetch_names)))
+        call_future = self._client.predict(
+            feed=data, fetch=self._fetch_names, asyn=True)
+        logging.debug(self._log("get call_future"))
+        return call_future
+
+    def postprocess(self, output_data):
+        return output_data
+
+    def stop(self):
+        self._input.stop()
+        for channel in self._outputs:
+            channel.stop()
+        self._is_run = False
+
+    def _parse_channeldata(self, channeldata):
+        data_id, error_channeldata = None, None
+        if isinstance(channeldata, dict):
+            parsed_data = {}
+            key = channeldata.keys()[0]
+            data_id = channeldata[key].id
+            for _, data in channeldata.items():
+                if data.ecode != ChannelDataEcode.OK.value:
+                    error_channeldata = data
+                    break
+        else:
+            data_id = channeldata.id
+            if channeldata.ecode != ChannelDataEcode.OK.value:
+                error_channeldata = channeldata
+        return data_id, error_channeldata
+
+    def _push_to_output_channels(self, data, channels, name=None):
+        if name is None:
+            name = self.name
+        for channel in channels:
+            channel.push(data, name)
+
+    def start(self):
+        proces = []
+        for concurrency_idx in range(self._concurrency):
+            p = multiprocessing.Process(
+                target=self._run,
+                args=(concurrency_idx, self.get_input_channel(),
+                      self.get_output_channels()))
+            p.start()
+            proces.append(p)
+        return proces
+
+    def _run(self, concurrency_idx, input_channel, output_channels):
+        self.init_client(self._client_config, self._server_name,
+                         self._fetch_names)
+        op_info_prefix = "[{}|{}]".format(self.name, concurrency_idx)
+        log = self._get_log_func(op_info_prefix)
+        self._is_run = True
+        while self._is_run:
+            _profiler.record("{}-get_0".format(op_info_prefix))
+            channeldata = input_channel.front(self.name)
+            _profiler.record("{}-get_1".format(op_info_prefix))
+            logging.debug(log("input_data: {}".format(channeldata)))
+
+            data_id, error_channeldata = self._parse_channeldata(channeldata)
+
+            # error data in predecessor Op
+            if error_channeldata is not None:
+                self._push_to_output_channels(error_channeldata,
+                                              output_channels)
+                continue
+
+            # preprecess
+            try:
+                _profiler.record("{}-prep_0".format(op_info_prefix))
+                preped_data = self.preprocess(channeldata)
+                _profiler.record("{}-prep_1".format(op_info_prefix))
+            except NotImplementedError as e:
+                # preprocess function not implemented
+                error_info = log(e)
+                logging.error(error_info)
+                self._push_to_output_channels(
+                    ChannelData(
+                        ecode=ChannelDataEcode.NOT_IMPLEMENTED.value,
+                        error_info=error_info,
+                        data_id=data_id),
+                    output_channels)
+                continue
+            except TypeError as e:
+                # Error type in channeldata.datatype
+                error_info = log(e)
+                logging.error(error_info)
+                self._push_to_output_channels(
+                    ChannelData(
+                        ecode=ChannelDataEcode.TYPE_ERROR.value,
+                        error_info=error_info,
+                        data_id=data_id),
+                    output_channels)
+                continue
+            except Exception as e:
+                error_info = log(e)
+                logging.error(error_info)
+                self._push_to_output_channels(
+                    ChannelData(
+                        ecode=ChannelDataEcode.UNKNOW.value,
+                        error_info=error_info,
+                        data_id=data_id),
+                    output_channels)
+                continue
+
+            # midprocess
+            call_future = None
+            if self.with_serving:
+                ecode = ChannelDataEcode.OK.value
+                _profiler.record("{}-midp_0".format(op_info_prefix))
+                if self._timeout <= 0:
+                    try:
+                        call_future = self.midprocess(preped_data)
+                    except Exception as e:
+                        ecode = ChannelDataEcode.UNKNOW.value
+                        error_info = log(e)
+                        logging.error(error_info)
+                else:
+                    for i in range(self._retry):
+                        try:
+                            call_future = func_timeout.func_timeout(
+                                self._timeout,
+                                self.midprocess,
+                                args=(preped_data, ))
+                        except func_timeout.FunctionTimedOut as e:
+                            if i + 1 >= self._retry:
+                                ecode = ChannelDataEcode.TIMEOUT.value
+                                error_info = log(e)
+                                logging.error(error_info)
+                            else:
+                                logging.warn(
+                                    log("timeout, retry({})".format(i + 1)))
+                        except Exception as e:
+                            ecode = ChannelDataEcode.UNKNOW.value
+                            error_info = log(e)
+                            logging.error(error_info)
+                            break
+                        else:
+                            break
+                if ecode != ChannelDataEcode.OK.value:
+                    self._push_to_output_channels(
+                        ChannelData(
+                            ecode=ecode, error_info=error_info,
+                            data_id=data_id),
+                        output_channels)
+                    continue
+                _profiler.record("{}-midp_1".format(op_info_prefix))
+
+            # postprocess
+            output_data = None
+            _profiler.record("{}-postp_0".format(op_info_prefix))
+            if self.with_serving:
+                # use call_future
+                output_data = ChannelData(
+                    datatype=ChannelDataType.CHANNEL_FUTURE.value,
+                    future=call_future,
+                    data_id=data_id,
+                    callback_func=self.postprocess)
+                #TODO: for future are not picklable
+                npdata = self.postprocess(call_future.result())
+                self._push_to_output_channels(
+                    ChannelData(
+                        ChannelDataType.CHANNEL_NPDATA.value,
+                        npdata=npdata,
+                        data_id=data_id),
+                    output_channels)
+                continue
+            else:
+                try:
+                    postped_data = self.postprocess(preped_data)
+                except Exception as e:
+                    ecode = ChannelDataEcode.UNKNOW.value
+                    error_info = log(e)
+                    logging.error(error_info)
+                    self._push_to_output_channels(
+                        ChannelData(
+                            ecode=ecode, error_info=error_info,
+                            data_id=data_id),
+                        output_channels)
+                    continue
+                if not isinstance(postped_data, dict):
+                    ecode = ChannelDataEcode.TYPE_ERROR.value
+                    error_info = log("output of postprocess funticon must be " \
+                            "dict type, but get {}".format(type(postped_data)))
+                    logging.error(error_info)
+                    self._push_to_output_channels(
+                        ChannelData(
+                            ecode=ecode, error_info=error_info,
+                            data_id=data_id),
+                        output_channels)
+                    continue
+
+                output_data = ChannelData(
+                    ChannelDataType.CHANNEL_NPDATA.value,
+                    npdata=postped_data,
+                    data_id=data_id)
+            _profiler.record("{}-postp_1".format(op_info_prefix))
+
+            # push data to channel (if run succ)
+            _profiler.record("{}-push_0".format(op_info_prefix))
+            self._push_to_output_channels(output_data, output_channels)
+            _profiler.record("{}-push_1".format(op_info_prefix))
+
+    def _log(self, info):
+        return "{} {}".format(self.name, info)
+
+    def _get_log_func(self, op_info_prefix):
+        def log_func(info_str):
+            return "{} {}".format(op_info_prefix, info_str)
+
+        return log_func
+
+    def get_concurrency(self):
+        return self._concurrency
+
+
+class VirtualOp(Op):
+    ''' For connecting two channels. '''
+
+    def __init__(self, name, concurrency=1):
+        super(VirtualOp, self).__init__(
+            name=name, inputs=None, concurrency=concurrency)
+        self._virtual_pred_ops = []
+
+    def add_virtual_pred_op(self, op):
+        self._virtual_pred_ops.append(op)
+
+    def add_output_channel(self, channel):
+        if not isinstance(channel, Channel):
+            raise TypeError(
+                self._log('output channel must be Channel type, not {}'.format(
+                    type(channel))))
+        for op in self._virtual_pred_ops:
+            channel.add_producer(op.name)
+        self._outputs.append(channel)
+
+    def _run(self, input_channel, output_channels):
+        op_info_prefix = "[{}|{}]".format(self.name, concurrency_idx)
+        log = self._get_log_func(op_info_prefix)
+        self._is_run = True
+        while self._is_run:
+            _profiler.record("{}-get_0".format(op_info_prefix))
+            channeldata = input_channel.front(self.name)
+            _profiler.record("{}-get_1".format(op_info_prefix))
+
+            _profiler.record("{}-push_0".format(op_info_prefix))
+            if isinstance(channeldata, dict):
+                for name, data in channeldata.items():
+                    self._push_to_output_channels(
+                        data, channels=output_channels, name=name)
+            else:
+                self._push_to_output_channels(
+                    channeldata,
+                    channels=output_channels,
+                    name=self._virtual_pred_ops[0].name)
+            _profiler.record("{}-push_1".format(op_info_prefix))
+
+
+class GeneralPythonService(
+        general_python_service_pb2_grpc.GeneralPythonService):
+    def __init__(self, in_channel, out_channel, retry=2):
+        super(GeneralPythonService, self).__init__()
+        self.name = "#G"
+        self.set_in_channel(in_channel)
+        self.set_out_channel(out_channel)
+        logging.debug(self._log(in_channel.debug()))
+        logging.debug(self._log(out_channel.debug()))
+        #TODO: 
+        #  multi-lock for different clients
+        #  diffenert lock for server and client
+        self._id_lock = threading.Lock()
+        self._cv = threading.Condition()
+        self._globel_resp_dict = {}
+        self._id_counter = 0
+        self._retry = retry
+        self._recive_func = threading.Thread(
+            target=GeneralPythonService._recive_out_channel_func, args=(self, ))
+        self._recive_func.start()
+
+    def _log(self, info_str):
+        return "[{}] {}".format(self.name, info_str)
+
+    def set_in_channel(self, in_channel):
+        if not isinstance(in_channel, Channel):
+            raise TypeError(
+                self._log('in_channel must be Channel type, but get {}'.format(
+                    type(in_channel))))
+        in_channel.add_producer(self.name)
+        self._in_channel = in_channel
+
+    def set_out_channel(self, out_channel):
+        if not isinstance(out_channel, Channel):
+            raise TypeError(
+                self._log('out_channel must be Channel type, but get {}'.format(
+                    type(out_channel))))
+        out_channel.add_consumer(self.name)
+        self._out_channel = out_channel
+
+    def _recive_out_channel_func(self):
+        while True:
+            channeldata = self._out_channel.front(self.name)
+            if not isinstance(channeldata, ChannelData):
+                raise TypeError(
+                    self._log('data must be ChannelData type, but get {}'.
+                              format(type(channeldata))))
+            with self._cv:
+                data_id = channeldata.id
+                self._globel_resp_dict[data_id] = channeldata
+                self._cv.notify_all()
+
+    def _get_next_id(self):
+        with self._id_lock:
+            self._id_counter += 1
+            return self._id_counter - 1
+
+    def _get_data_in_globel_resp_dict(self, data_id):
+        resp = None
+        with self._cv:
+            while data_id not in self._globel_resp_dict:
+                self._cv.wait()
+            resp = self._globel_resp_dict.pop(data_id)
+            self._cv.notify_all()
+        return resp
+
+    def _pack_data_for_infer(self, request):
+        logging.debug(self._log('start inferce'))
+        data_id = self._get_next_id()
+        npdata = {}
+        try:
+            for idx, name in enumerate(request.feed_var_names):
+                logging.debug(
+                    self._log('name: {}'.format(request.feed_var_names[idx])))
+                logging.debug(
+                    self._log('data: {}'.format(request.feed_insts[idx])))
+                npdata[name] = np.frombuffer(
+                    request.feed_insts[idx], dtype=request.type[idx])
+                npdata[name].shape = np.frombuffer(
+                    request.shape[idx], dtype="int32")
+        except Exception as e:
+            return ChannelData(
+                ecode=ChannelDataEcode.RPC_PACKAGE_ERROR.value,
+                error_info="rpc package error",
+                data_id=data_id), data_id
+        else:
+            return ChannelData(
+                datatype=ChannelDataType.CHANNEL_NPDATA.value,
+                npdata=npdata,
+                data_id=data_id), data_id
+
+    def _pack_data_for_resp(self, channeldata):
+        logging.debug(self._log('get channeldata'))
+        resp = pyservice_pb2.Response()
+        resp.ecode = channeldata.ecode
+        if resp.ecode == ChannelDataEcode.OK.value:
+            if channeldata.datatype == ChannelDataType.CHANNEL_PBDATA.value:
+                for inst in channeldata.pbdata.insts:
+                    resp.fetch_insts.append(inst.data)
+                    resp.fetch_var_names.append(inst.name)
+                    resp.shape.append(inst.shape)
+                    resp.type.append(inst.type)
+            elif channeldata.datatype in (ChannelDataType.CHANNEL_FUTURE.value,
+                                          ChannelDataType.CHANNEL_NPDATA.value):
+                feed = channeldata.parse()
+                for name, var in feed.items():
+                    resp.fetch_insts.append(var.tobytes())
+                    resp.fetch_var_names.append(name)
+                    resp.shape.append(
+                        np.array(
+                            var.shape, dtype="int32").tobytes())
+                    resp.type.append(str(var.dtype))
+            else:
+                raise TypeError(
+                    self._log("Error type({}) in datatype.".format(
+                        channeldata.datatype)))
+        else:
+            resp.error_info = channeldata.error_info
+        return resp
+
+    def inference(self, request, context):
+        _profiler.record("{}-prepack_0".format(self.name))
+        data, data_id = self._pack_data_for_infer(request)
+        _profiler.record("{}-prepack_1".format(self.name))
+
+        resp_channeldata = None
+        for i in range(self._retry):
+            logging.debug(self._log('push data'))
+            _profiler.record("{}-push_0".format(self.name))
+            self._in_channel.push(data, self.name)
+            _profiler.record("{}-push_1".format(self.name))
+
+            logging.debug(self._log('wait for infer'))
+            _profiler.record("{}-fetch_0".format(self.name))
+            resp_channeldata = self._get_data_in_globel_resp_dict(data_id)
+            _profiler.record("{}-fetch_1".format(self.name))
+
+            if resp_channeldata.ecode == ChannelDataEcode.OK.value:
+                break
+            if i + 1 < self._retry:
+                logging.warn("retry({}): {}".format(
+                    i + 1, resp_channeldata.error_info))
+
+        _profiler.record("{}-postpack_0".format(self.name))
+        resp = self._pack_data_for_resp(resp_channeldata)
+        _profiler.record("{}-postpack_1".format(self.name))
+        _profiler.print_profile()
+        return resp
+
+
+class PyServer(object):
+    def __init__(self, retry=2, profile=False):
+        self._channels = []
+        self._user_ops = []
+        self._actual_ops = []
+        self._port = None
+        self._worker_num = None
+        self._in_channel = None
+        self._out_channel = None
+        self._retry = retry
+        self._manager = multiprocessing.Manager()
+        _profiler.enable(profile)
+
+    def add_channel(self, channel):
+        self._channels.append(channel)
+
+    def add_op(self, op):
+        self._user_ops.append(op)
+
+    def add_ops(self, ops):
+        self._user_ops.extend(ops)
+
+    def gen_desc(self):
+        logging.info('here will generate desc for PAAS')
+        pass
+
+    def _topo_sort(self):
+        indeg_num = {}
+        que_idx = 0  # scroll queue 
+        ques = [Queue.Queue() for _ in range(2)]
+        for op in self._user_ops:
+            if len(op.get_input_ops()) == 0:
+                op.name = "#G"  # update read_op.name
+                break
+        outdegs = {op.name: [] for op in self._user_ops}
+        for idx, op in enumerate(self._user_ops):
+            # check the name of op is globally unique
+            if op.name in indeg_num:
+                raise Exception("the name of Op must be unique")
+            indeg_num[op.name] = len(op.get_input_ops())
+            if indeg_num[op.name] == 0:
+                ques[que_idx].put(op)
+            for pred_op in op.get_input_ops():
+                outdegs[pred_op.name].append(op)
+
+        # topo sort to get dag_views
+        dag_views = []
+        sorted_op_num = 0
+        while True:
+            que = ques[que_idx]
+            next_que = ques[(que_idx + 1) % 2]
+            dag_view = []
+            while que.qsize() != 0:
+                op = que.get()
+                dag_view.append(op)
+                sorted_op_num += 1
+                for succ_op in outdegs[op.name]:
+                    indeg_num[succ_op.name] -= 1
+                    if indeg_num[succ_op.name] == 0:
+                        next_que.put(succ_op)
+            dag_views.append(dag_view)
+            if next_que.qsize() == 0:
+                break
+            que_idx = (que_idx + 1) % 2
+        if sorted_op_num < len(self._user_ops):
+            raise Exception("not legal DAG")
+        if len(dag_views[0]) != 1:
+            raise Exception("DAG contains multiple input Ops")
+        if len(dag_views[-1]) != 1:
+            raise Exception("DAG contains multiple output Ops")
+
+        # create channels and virtual ops
+        def name_generator(prefix):
+            def number_generator():
+                idx = 0
+                while True:
+                    yield "{}{}".format(prefix, idx)
+                    idx += 1
+
+            return number_generator()
+
+        virtual_op_name_gen = name_generator("vir")
+        channel_name_gen = name_generator("chl")
+        virtual_ops = []
+        channels = []
+        input_channel = None
+        actual_view = None
+        for v_idx, view in enumerate(dag_views):
+            if v_idx + 1 >= len(dag_views):
+                break
+            next_view = dag_views[v_idx + 1]
+            if actual_view is None:
+                actual_view = view
+            actual_next_view = []
+            pred_op_of_next_view_op = {}
+            for op in actual_view:
+                # find actual succ op in next view and create virtual op
+                for succ_op in outdegs[op.name]:
+                    if succ_op in next_view:
+                        if succ_op not in actual_next_view:
+                            actual_next_view.append(succ_op)
+                        if succ_op.name not in pred_op_of_next_view_op:
+                            pred_op_of_next_view_op[succ_op.name] = []
+                        pred_op_of_next_view_op[succ_op.name].append(op)
+                    else:
+                        # create virtual op
+                        virtual_op = None
+                        virtual_op = VirtualOp(name=virtual_op_name_gen.next())
+                        virtual_ops.append(virtual_op)
+                        outdegs[virtual_op.name] = [succ_op]
+                        actual_next_view.append(virtual_op)
+                        pred_op_of_next_view_op[virtual_op.name] = [op]
+                        virtual_op.add_virtual_pred_op(op)
+            actual_view = actual_next_view
+            # create channel
+            processed_op = set()
+            for o_idx, op in enumerate(actual_next_view):
+                if op.name in processed_op:
+                    continue
+                channel = Channel(self._manager, name=channel_name_gen.next())
+                channels.append(channel)
+                logging.debug("{} => {}".format(channel.name, op.name))
+                op.add_input_channel(channel)
+                pred_ops = pred_op_of_next_view_op[op.name]
+                if v_idx == 0:
+                    input_channel = channel
+                else:
+                    # if pred_op is virtual op, it will use ancestors as producers to channel
+                    for pred_op in pred_ops:
+                        logging.debug("{} => {}".format(pred_op.name,
+                                                        channel.name))
+                        pred_op.add_output_channel(channel)
+                processed_op.add(op.name)
+                # find same input op to combine channel
+                for other_op in actual_next_view[o_idx + 1:]:
+                    if other_op.name in processed_op:
+                        continue
+                    other_pred_ops = pred_op_of_next_view_op[other_op.name]
+                    if len(other_pred_ops) != len(pred_ops):
+                        continue
+                    same_flag = True
+                    for pred_op in pred_ops:
+                        if pred_op not in other_pred_ops:
+                            same_flag = False
+                            break
+                    if same_flag:
+                        logging.debug("{} => {}".format(channel.name,
+                                                        other_op.name))
+                        other_op.add_input_channel(channel)
+                        processed_op.add(other_op.name)
+        output_channel = Channel(self._manager, name=channel_name_gen.next())
+        channels.append(output_channel)
+        last_op = dag_views[-1][0]  # TODO: fix it
+        last_op.add_output_channel(output_channel)
+
+        self._actual_ops = virtual_ops
+        for op in self._user_ops:
+            if len(op.get_input_ops()) == 0:
+                # pass read op
+                continue
+            self._actual_ops.append(op)
+        self._channels = channels
+        for c in channels:
+            logging.debug(c.debug())
+        return input_channel, output_channel
+
+    def prepare_server(self, port, worker_num):
+        self._port = port
+        self._worker_num = worker_num
+
+        input_channel, output_channel = self._topo_sort()
+        self._in_channel = input_channel
+        self._out_channel = output_channel
+        for op in self._actual_ops:
+            if op.with_serving:
+                self.prepare_serving(op)
+        self.gen_desc()
+
+    def _run_ops(self):
+        proces = []
+        for op in self._actual_ops:
+            proces.extend(op.start())
+        return proces
+
+    def _stop_ops(self):
+        for op in self._actual_ops:
+            op.stop()
+
+    def run_server(self):
+        op_proces = self._run_ops()
+        server = grpc.server(
+            futures.ThreadPoolExecutor(max_workers=self._worker_num))
+        general_python_service_pb2_grpc.add_GeneralPythonServiceServicer_to_server(
+            GeneralPythonService(self._in_channel, self._out_channel,
+                                 self._retry), server)
+        server.add_insecure_port('[::]:{}'.format(self._port))
+        server.start()
+        server.wait_for_termination()
+        self._stop_ops()  # TODO
+        for p in op_proces:
+            p.join()
+
+    def prepare_serving(self, op):
+        model_path = op._server_model
+        port = op._server_port
+        device = op._device
+
+        if device == "cpu":
+            cmd = "(Use MultiLangServer) python -m paddle_serving_server.serve" \
+                  " --model {} --thread 4 --port {} --use_multilang &>/dev/null &".format(model_path, port)
+        else:
+            cmd = "(Use MultiLangServer) python -m paddle_serving_server_gpu.serve" \
+                  " --model {} --thread 4 --port {} --use_multilang &>/dev/null &".format(model_path, port)
+        # run a server (not in PyServing)
+        logging.info("run a server (not in PyServing): {}".format(cmd))