Tensorflow多處理; UnknownError：無法啟動gRPC服務器

Question

我正在研究大型數據集上的粗麻布矩陣。 我正在嘗試在多個CPU上並行執行這些計算。 我的設置當前有1個節點和10個CPU。 我正在使用Python 2.7

我為代碼編寫了一個小抽象，以更好地理解分布式張量流。 下面是錯誤

2017-07-23 16:16:17.281414: I tensorflow/core/distributed_runtime/rpc/grpc_server_lib.cc:316] Started server with target: grpc://localhost:2225
Process Process-3:
Traceback (most recent call last):
  File "/home/skay/anaconda2/lib/python2.7/multiprocessing/process.py", line 258, in _bootstrap
    self.run()
  File "/home/skay/anaconda2/lib/python2.7/multiprocessing/process.py", line 114, in run
    self._target(*self._args, **self._kwargs)
  File "/home/skay/.PyCharmCE2017.1/config/scratches/scratch_6.py", line 32, in cifar10
    serv = tf.train.Server(cluster, job_name= params.job_name,task_index=params.task_index)
  File "/home/skay/anaconda2/lib/python2.7/site-packages/tensorflow/python/training/server_lib.py", line 145, in __init__
    self._server_def.SerializeToString(), status)
  File "/home/skay/anaconda2/lib/python2.7/contextlib.py", line 24, in __exit__
    self.gen.next()
  File "/home/skay/anaconda2/lib/python2.7/site-packages/tensorflow/python/framework/errors_impl.py", line 466, in raise_exception_on_not_ok_status
    pywrap_tensorflow.TF_GetCode(status)) UnknownError: Could not start gRPC server

每次運行代碼時，我都會收到此錯誤。 但是，它會進一步為我如下設置的兩個過程之一生成輸出

> `2017-07-23 16:27:48.605617: I tensorflow/core/distributed_runtime/master_session.cc:999] Start master session fe9fd6a338e2c9a7 with config: 

2017-07-23 16:27:48.607126: I tensorflow/core/distributed_runtime/master_session.cc:999] Start master session 3560417f98b00dea with config: 

[  1.   2.   3.   4.   5.   6.   7.   8.   9.  10.]
Process-3
[  1.   2.   3.   4.   5.   6.   7.   8.   9.  10.]
Process-3
[  1.   2.   3.   4.   5.   6.   7.   8.   9.  10.]
Process-3

在此它繼續等待下一個

ERROR:tensorflow:==================================
Object was never used (type <class 'tensorflow.python.framework.ops.Operation'>):
<tf.Operation 'worker_0/init' type=NoOp>
If you want to mark it as used call its "mark_used()" method.
It was originally created here:
['File "/home/skay/.PyCharmCE2017.1/config/scratches/scratch_6.py", line 83, in <module>\n    proc.start()', 'File "/home/skay/anaconda2/lib/python2.7/multiprocessing/process.py", line 130, in start\n    self._popen = Popen(self)', 'File "/home/skay/anaconda2/lib/python2.7/multiprocessing/forking.py", line 126, in __init__\n    code = process_obj._bootstrap()', 'File "/home/skay/anaconda2/lib/python2.7/multiprocessing/process.py", line 258, in _bootstrap\n    self.run()', 'File "/home/skay/anaconda2/lib/python2.7/multiprocessing/process.py", line 114, in run\n    self._target(*self._args, **self._kwargs)', 'File "/home/skay/.PyCharmCE2017.1/config/scratches/scratch_6.py", line 49, in cifar10\n    init_op=tf.initialize_all_variables(),logdir=\'/tmp/mydir\')', 'File "/home/skay/anaconda2/lib/python2.7/site-packages/tensorflow/python/util/tf_should_use.py", line 170, in wrapped\n    return _add_should_use_warning(fn(*args, **kwargs))', 'File "/home/skay/anaconda2/lib/python2.7/site-packages/tensorflow/python/util/tf_should_use.py", line 139, in _add_should_use_warning\n    wrapped = TFShouldUseWarningWrapper(x)', 'File "/home/skay/anaconda2/lib/python2.7/site-packages/tensorflow/python/util/tf_should_use.py", line 96, in __init__\n    stack = [s.strip() for s in traceback.format_stack()]']
==================================
2017-07-23 16:28:28.646871: I tensorflow/core/distributed_runtime/master.cc:209] CreateSession still waiting for response from worker: /job:worker/replica:0/task:0
2017-07-23 16:28:38.647276: I tensorflow/core/distributed_runtime/master.cc:209] CreateSession still waiting for response from worker: /job:worker/replica:0/task:0
2017-07-23 16:28:48.647526: I tensorflow/core/distributed_runtime/master.cc:209] CreateSession still waiting for response from worker: /job:worker/replica: 

我在這里有2個問題

1. 如何解決有關Grpc的錯誤
2. 我已經使用Manager（）設置了一個多處理隊列“結果”，並在設置流程時將其傳遞給兩個工作人員。 我希望一旦達到條件，每個進程就會將其Job ID寫入隊列，但是似乎隊列始終包含最后完成的進程。 這是否意味着隊列中的某個位置被另一個進程覆蓋

[{'worker'：0}，{'worker'：0}]

我可以使用多處理隊列在tensorflow上兩個不同進程上運行的兩個會話之間共享字典嗎？

下面是我的代碼

# build a python mutliprocess.py
import multiprocessing
import time
import tensorflow as tf
from tensorflow.contrib.training import HParams
import os
import psutil
import numpy as np
from tensorflow.python.client import device_lib
from resnet import *
import Queue

cluster_spec ={"ps": ["localhost:2226"
                      ],
    "worker": [
        "localhost:2227",
        "localhost:2228"]}

cluster = tf.train.ClusterSpec(cluster_spec)
im_Test = np.linspace(1,10,10)

def model_fun(input):
    print multiprocessing.current_process().name
    return input

def cifar10(device,return_dict,result_t):
    params = HParams(cluster=cluster,
                     job_name = device[0],
                     task_index = device[1])

    serv = tf.train.Server(cluster, job_name= params.job_name,task_index=params.task_index)
    input_img=[]
    true_lab=[]

    if params.job_name == "ps":
        ##try and wait for all the wokers t
        serv.join()
    elif params.job_name == "worker":
        with tf.device(tf.train.replica_device_setter(worker_device="/job:worker/replica:0/task:%d" % params.task_index,
                                                      cluster=cluster)):
            # with tf.Graph().as_default(), tf.device('/cpu:%d' % params.task_index):
            # with tf.container('%s %d' % ('batchname', params.task_index)) as scope:
            input_img = tf.placeholder(dtype=tf.float32, shape=[10,])
            with tf.name_scope('%s_%d' % (params.job_name, params.task_index)) as scope:
                hess_op = model_fun(input_img)
                global_step = tf.contrib.framework.get_or_create_global_step()
                sv = tf.train.Supervisor(is_chief=(params.task_index == 0),
                                         global_step=global_step,
                                         init_op=tf.initialize_all_variables(),logdir='/tmp/mydir')
                with sv.prepare_or_wait_for_session(serv.target) as sess:
                    step = 0
                    while not sv.should_stop() :
                        hess = sess.run(hess_op, feed_dict={input_img:im_Test })
                        print(np.array(hess))
                        print multiprocessing.current_process().name
                        step += 1
                        if(step==3):
                            return_dict[params.job_name] = params.task_index
                            result_t.put(return_dict)
                            break
                    sv.stop()
                    sess.close()


    return

if __name__ == '__main__':

    logger = multiprocessing.log_to_stderr()
    manager = multiprocessing.Manager()
    result = manager.Queue()
    return_dict = manager.dict()
    processes = []
    devices = [['ps', 0],
               ['worker', 0],
               ['worker', 1]
               ]

    for i in (devices):
        start_time = time.time()
        proc = multiprocessing.Process(target=cifar10,args=(i,return_dict,result))
        processes.append(proc)
        proc.start()

    for p in processes:
        p.join()

    # print return_dict.values()
    kill = []
    while True:
        if result.empty() == True:
                break
        kill.append(result.get())
        print kill


    print("time taken = %d" % (start_time - time.time()))

Answer 1

就我而言，我發現ps會引發此錯誤，並且當我提交tensorflowonspark工作紗線簇模式時，woker等待響應。

ps錯誤如下

2018-01-17 11：08：46,366 INFO（MainThread-7305）在后台進程上的群集節點0上啟動TensorFlow ps：0 2018-01-17 11：08：56,085 INFO（MainThread-7395）0：==== ==== ps：0 ========= 2018年1月17日11：08：56,086信息（MainThread-7395）0：群集規格：{'ps'：['172.16.5.30:33088'] ，'worker'：['172.16.5.22:41428'，'172.16.5.30:33595']} 2018-01-17 11：08：56,086 INFO（MainThread-7395）0：使用CPU 2018-01-17 11： 08：56.087452：I tensorflow / core / platform / cpu_feature_guard.cc：137]您的CPU支持該TensorFlow二進制文件未編譯為使用的指令：SSE4.1 SSE4.2 AVX AVX2 FMA E0117 11：08：56.088501182 7395 ev_epoll1_linux.c ：1051] grpc epoll fd：10 E0117 11：08：56.088860707 7395 server_chttp2.c：38] {“ created”：“ @ 1516158536.088783549”，“ description”：“未在總共1個解析的地址中添加地址”，“ file”： “ external / grpc / src / core / ext / transport / chttp2 / server / chttp2_server.c”，“ file_line”：245，“ referenced_errors”：[{“ created”：“ @ 1516158536.088779164”，“ description”：“失敗添加任何通配符偵聽器”， “ file”：“ external / grpc / src / core / lib / iomgr / tcp_server_posix.c”，“ file_line”：338，“ referenced_errors”：[{“ created”：“ @ 1516158536.088771177”，“ description”：“無法配置套接字“，” fd“：12，”文件“：” external / grpc / src / core / lib / iomgr / tcp_server_utils_posix_common.c“，” file_line“：200，” referenced_errors“：[{” created“：” @ 1516158536.088767669“，”描述“：” OS錯誤“，” errno“：98，”文件“：” external / grpc / src / core / lib / iomgr / tcp_server_utils_posix_common.c“，” file_line“：173，” os_error“： “地址已在使用中”，“ syscall”：“綁定”}]}}，{“創建”：“ @ 1516158536.088778651”，“描述”：“無法配置套接字”，“ fd”：12，“文件”：“ external / grpc / src / core / lib / iomgr / tcp_server_utils_posix_common.c“，” file_line“：200，” referenced_errors“：[{” created“：” @ 1516158536.088776541“，” description“：” OS Error“，” errno“ ：98，“文件”：“外部/ grpc / src /核心/lib/iomgr/tcp_server_utils_posix_common.c”，“文件行”：173，“ os_error”：“地址已在使用中”，“ syscall”：“綁定”} ]}]}]}處理流程2：追溯（最近一次呼叫最近）：文件“ / data / yarn /nm/usercache/hdfs/appcache/application_1515984940590_0270/container_e13_1515984940590_0270_01_000002/Python/lib/python2.7/multiprocessing/process.py“，第258行，在_bootstrap self.run（）文件中，“ / data / yarn / nm / usercache / hdfs /appcache/application_1515984940590_0270/container_e13_1515984940590_0270_01_000002/Python/lib/python2.7/multiprocessing/process.py“，行114，運行self._target（* self._args，** self._kwargs）文件” / data / yarn / nm /usercache/hdfs/appcache/application_1515984940590_0270/container_e13_1515984940590_0270_01_000001/tfspark.zip/tensorflowonspark/TFSparkNode.py “線路269，在wrapper_fn文件”/數據/紗線/納米/ usercache / HDFS /應用程序緩存/ application_1515984940590_0270 / container_e13_1515984940590_0270_01_000002 / pyfiles / mnist_dist。 py“，第38行，在map_fun集群中，服務器= ctx.start_cluster_server（1，args.rdma）文件” /data/yarn/nm/usercache/hdfs/appcache/application_1515984940590_0270/container_e13_1515984940590_0270_01_000002/tfspark.py/Tensorparkons.py “，第56行，在 start_cluster_server返回TFNode.start_cluster_server（自身，num_gpus，rdma）文件“ /data/yarn/nm/usercache/hdfs/appcache/application_1515984940590_0270/container_e13_1515984940590_0270_01_000002/tfspark.zip/tensorflowons_park,ster_server,lu/tc_server_lu，server_lu，服務器端。 .train.Server（群集，ctx.job_name，ctx.task_index）文件“ /data/yarn/nm/usercache/hdfs/appcache/application_1515984940590_0270/container_e13_1515984940590_0270_01_000002/Python/lib/python2.7/site-packages/tensorflow/python/ init self._server_def.SerializeToString（），狀態）第145行中的“ training / server_lib.py”，狀態）文件“ /data/yarn/nm/usercache/hdfs/appcache/application_1515984940590_0270/container_e13_1515984940590_0270_01_000002/Python/lib/python2.7/site -packages / tensorflow / python / framework / errors_impl.py“，行473，在出口 c_api.TF_GetCode（self.status.status））中，UnknownError：無法啟動gRPC服務器

woker:1 log

2018-01-17 11：09：14.614244：I tensorflow / core / distributed_runtime / master.cc：221] CreateSession仍在等待工作者的響應：/ job：ps /副本：0 / task：0

然后，我檢查ps服務器中的端口。 是的，該端口已被使用。

因此，重新提交工作即可解決問題。

但是，如果您每次運行代碼都收到此錯誤，則應檢查更多日志以查找原因。