Tensorflow on Spark爬坑指南

北京 上海巡回站 | NVIDIA DLI深度学习培训

2018年1月26/1月12日

NVIDIA 深度学习学院 带你快速进入火热的DL领域

正文共10718个字 3张图，预计阅读时间27分钟。

由于机器学习和深度学习不断被炒热，Tensorflow作为Google家(Jeff Dean大神)推出的开源深度学习框架，也获得了很多关注。Tensorflow的灵活性很强，允许用户使用多台机器的多个设备(如不同的CPU和GPU)。但是由于Tensorflow 分布式的方式需要用户在客户端显示指定集群信息，另外需要手动拉起ps， worker等task. 对资源管理和使用上有诸多不便。因此，Yahoo开源了基于Spark的Tensorflow，使用executor执行worker和ps task. 项目地址为：https://github.com/yahoo/TensorFlowOnSpark。

写在前面，前方高能，请注意！

虽然yahoo提供了如何在Spark集群中运行Tensorflow的步骤，但是由于这个guideline过于简单，一般情况下，根据这个guideline是跑不起来的。

Tensorflow on spark介绍

TensorflowOnSpark 支持使用Spark/Hadoop集群分布式的运行Tensorflow，号称支持所有的Tensorflow操作。需要注意的是用户需要对原有的TF程序进行简单的改造，就能够运行在Spark集群之上。

如何跑起来Tensorflow on spark

虽然Yahoo在github上说明了安装部署TFS (https://github.com/yahoo/TensorFlowOnSpark/wiki/GetStarted_YARN), 但是根据实际实践，根据这个文档如果能跑起来，那真的要谢天谢地。因为在实际过程中，会因为环境问题遇到一些unexpected error。以下就是我将自己在实践过程中遇到的一些问题总结列举。

1、编译python和pip

yahoo提供的编译步骤为:

# download and extract Python 2.7export PYTHON_ROOT=~/Pythoncurl -O https://www.python.org/ftp/python/2.7.12/Python-2.7.12.tgztar -xvf Python-2.7.12.

tgzrm Python-2.7.12.tgz

# compile into local PYTHON_ROOT

pushd Python-2.7.12./configure --prefix="$" --enable-unicode=ucs4makemake installpopdrm -rf Python-2.7.12 # install pip

pushd "$"

curl -O https://bootstrap.pypa.io/get-pip.pybin/python get-pip.pyrm get-pip.py

# install tensorflow (and any custom dependencies)

$/bin/pip install pydoop

# Note: add any extra dependencies here

popd

在实际编译过程中，采用的Centos7.2操作系统，可能出现以下问题：

安装pip报错

bin/python get-pip.pyERROR:root:code for hash sha224 was not found.Traceback (most recent call last):

报这个错一般是因为python中缺少_ssl.so 和 _hashlib.so库造成，可以从系统python库中找对应版本的拷贝到相应的python文件夹下(例如：lib/python2.7/lib-dynload)。

缺少zlib

bin/python get-pip.pyTraceback (most recent call last): File "get-pip.py", line 20061, in main() File "get-pip.py", line 194, in main bootstrap(tmpdir=tmpdir) File "get-pip.py", line 82, in bootstrap import pipzipimport.ZipImportError: can't decompress data; zlib not available

解决这个问题的方法是使用yum安装zlib*后，重新编译python后，即可解决。

ssl 报错

bin/python get-pip.pypip is configured with locations that require TLS/SSL, however the ssl module in Python is not available.Collecting pip Could not fetch URL https://pypi.python.org/simple/pip/: There was a problem confirming the ssl certificate: Can't connect to HTTPS URL because the SSL module is not available. - skipping Could not find a version that satisfies the requirement pip (from versions: )No matching distribution found for pip

解决方法： 在Python安装目录下打开文件lib/python2.7/ssl.py，注释掉 , HAS_ALPN

from _ssl import HAS_SNI, HAS_ECDH, HAS_NPN#, HAS_ALPN

pip install pydoop报错

gcc: error trying to exec 'cc1plus': execvp:

解决办法：需要在机器上安装g++编译器

2、安装编译 tensorflow w/RDMA support

注意编译过程需要google的bazel和protoc, 这两个工具需要提前装好。

3、接下来的步骤按照

https://github.com/yahoo/TensorFlowOnSpark/wiki/GetStarted_YARN 指导的步骤完成。

4、在HDP2.5部署的spark on Yarn环境上运行tensorflow

在yarn-env.sh中设置环境变量，增加 * export HADOOP_HDFS_HOME=/usr/hdp/2.5.0.0-1245/hadoop-hdfs/*

因为这个环境变量需要在执行tensorflow任务时被用到，如果没有export，会报错。

重启YARN，使上述改动生效。

按照Yahoo github上的步骤，执行训练mnist任务时，按下面命令提交作业:

export PYTHON_ROOT=/data2/Python/export LD_LIBRARY_PATH=$export PYSPARK_PYTHON=$/bin/pythonexport SPARK_YARN_USER_ENV="PYSPARK_PYTHON=Python/bin/python"export PATH=$/bin/:$PATHexport QUEUE=default spark-submit \--master yarn \--deploy-mode cluster \--queue $ \--num-executors 4 \--executor-memory 1G \--py-files /data2/tesorflowonSpark/TensorFlowOnSpark/tfspark.zip,/data2/tesorflowonSpark/TensorFlowOnSpark/examples/mnist/spark/mnist_dist.py \--conf spark.dynamicAllocation.enabled=false \--conf spark.yarn.maxAppAttempts=1 \--archives hdfs:///user/$/Python.zip#Python \--conf spark.executorEnv.LD_LIBRARY_PATH="/usr/jdk64/jdk1.8.0_77/jre/lib/amd64/server/" \/data2/tesorflowonSpark/TensorFlowOnSpark/examples/mnist/spark/mnist_spark.py \--images mnist/csv/test/images \--labels mnist/csv/test/labels \--mode inference \--model mnist_model \--output predictions

此时，通过Spark界面可以观察到worker0处于阻塞状态。

17/03/21 18:17:18 INFO MemoryStore: Block broadcast_1_piece0 stored as bytes in memory (estimated size 28.4 KB, free 542.6 KB)17/03/21 18:17:18 INFO TorrentBroadcast: Reading broadcast variable 1 took 17 ms17/03/21 18:17:18 INFO MemoryStore: Block broadcast_1 stored as values in memory (estimated size 440.6 KB, free 983.3 KB)2017-03-21 18:17:18,404 INFO (MainThread-14872) Connected to TFSparkNode.mgr on ochadoop03, ppid=14685, state='running'2017-03-21 18:17:18,411 INFO (MainThread-14872) mgr.state='running'2017-03-21 18:17:18,411 INFO (MainThread-14872) Feeding partition into input queue 17/03/21 18:17:20 INFO PythonRunner: Times: total = 2288, boot = -5387, init = 5510, finish = 216517/03/21 18:17:20 INFO PythonRunner: Times: total = 101, boot = 3, init = 21, finish = 772017-03-21 18:17:20.587060: I tensorflow/core/distributed_runtime/master_session.cc:1011] Start master session b5d9a21a16799e0b with config:

通过分析原因发现，在mnist例子中，logdir设置的是hdfs的路径，可能是由于tf对hdfs的支持有限或者存在bug（惭愧，并没有深究 :)）。将logdir改为本地目录，就可以正常运行。但是由此又带来了另一个问题，因为Spark每次启动时worker0的位置并不确定，有可能每次启动的机器都不同，这就导致在inference的时候没有办法获得训练的模型。

一个解决办法是：在worker 0训练完模型后，将模型同步到hdfs中，在inference的之前，再将hdfs的checkpoints文件夹拉取到本地执行。以下为我对yahoo提供的mnist example做的类似的修改.

def writeFileToHDFS():

rootdir = '/tmp/mnist_model'

client = HdfsClient(hosts='localhost:50070')

client.mkdirs('/user/root/mnist_model')

for parent,dirnames,filenames in os.walk(rootdir):

for dirname in dirnames:

print("parent is:".format(parent))

for filename in filenames:

client.copy_from_local(os.path.join(parent,filename), os.path.join('/user/root/mnist_model',filename), overwrite=True)

#logdir = TFNode.hdfs_path(ctx, args.model)

logdir = "/tmp/" + args.model

while not sv.should_stop() and step

# Run a training step asynchronously.

# See `tf.train.SyncReplicasOptimizer` for additional details on how to

# perform *synchronous* training.

# using feed_dict

batch_xs, batch_ys = feed_dict()

feed =

if len(batch_xs) != batch_size:

print("done feeding")

break

else:

if args.mode == "train":

_, step = sess.run([train_op, global_step], feed_dict=feed)

# print accuracy and save model checkpoint to HDFS every 100 steps

if (step % 100 == 0):

print(" step: accuracy: ".format(datetime.now().isoformat(), step, sess.run(accuracy,)))

else:

# args.mode == "inference"

labels, preds, acc = sess.run([label, prediction, accuracy], feed_dict=feed)

results = [" Label: , Prediction: ".format(datetime.now().isoformat(), l, p) for l,p in zip(labels,preds)]

TFNode.batch_results(ctx.mgr, results)

print("acc: ".format(acc))

if task_index == 0:

writeFileToHDFS()

当然这段代码只是为了进行说明，并不是很严谨，在上传hdfs的时候，是需要对文件夹是否存在等要做一系列的判断。

5、train& inference

向Spark集群提交训练任务.

spark-submit \--master yarn \--deploy-mode cluster \--queue $ \--num-executors 3 \--executor-memory 7G \--py-files /data2/tesorflowonSpark/TensorFlowOnSpark/tfspark.zip,/data2/tesorflowonSpark/TensorFlowOnSpark/examples/mnist/spark/mnist_dist.py \--conf spark.dynamicAllocation.enabled=false \--conf spark.yarn.maxAppAttempts=1 \--archives hdfs:///user/$/Python.zip#Python \--conf spark.executorEnv.LD_LIBRARY_PATH="/usr/jdk64/jdk1.8.0_77/jre/lib/amd64/server/" \/data2/tesorflowonSpark/TensorFlowOnSpark/examples/mnist/spark/mnist_spark.py \--images mnist/csv/train/images \--labels mnist/csv/train/labels \--mode train \--model mnist_model

执行起来后，查看Spark UI，可以看到当前训练过程中的作业执行情况。

6.46.43.png

执行完后，检查hdsf，checkpoint目录, 可以看到模型的checkpoints已经上传到hdfs中。

hadoop fs -ls /user/root/mnist_modelFound 8 items-rwxr-xr-x 3 root hdfs 179 2017-03-21 18:53 /user/root/mnist_model/checkpoint-rwxr-xr-x 3 root hdfs 117453 2017-03-21 18:53 /user/root/mnist_model/graph.pbtxt-rwxr-xr-x 3 root hdfs 814164 2017-03-21 18:53 /user/root/mnist_model/model.ckpt-0.data-00000-of-00001-rwxr-xr-x 3 root hdfs 372 2017-03-21 18:53 /user/root/mnist_model/model.ckpt-0.index-rwxr-xr-x 3 root hdfs 45557 2017-03-21 18:53 /user/root/mnist_model/model.ckpt-0.meta-rwxr-xr-x 3 root hdfs 814164 2017-03-21 18:53 /user/root/mnist_model/model.ckpt-338.data-00000-of-00001-rwxr-xr-x 3 root hdfs 372 2017-03-21 18:53 /user/root/mnist_model/model.ckpt-338.index-rwxr-xr-x 3 root hdfs 45557 2017-03-21 18:53 /user/root/mnist_model/model.ckpt-338.meta

根据训练的结果，执行模型inference

spark-submit \--master yarn \--deploy-mode cluster \--queue $ \--num-executors 4 \--executor-memory 1G \--py-files /data2/tesorflowonSpark/TensorFlowOnSpark/tfspark.zip,/data2/tesorflowonSpark/TensorFlowOnSpark/examples/mnist/spark/mnist_dist.py \--conf spark.dynamicAllocation.enabled=false \--conf spark.yarn.maxAppAttempts=1 \--archives hdfs:///user/$/Python.zip#Python \--conf spark.executorEnv.LD_LIBRARY_PATH="/usr/jdk64/jdk1.8.0_77/jre/lib/amd64/server/" \/data2/tesorflowonSpark/TensorFlowOnSpark/examples/mnist/spark/mnist_spark.py \--images mnist/csv/test/images \--labels mnist/csv/test/labels \--mode inference \--model mnist_model \--output predictions

等任务执行完成后，会发现，模型判断的结果已经输出到hdfs相关目录下了。

hadoop fs -ls /user/root/predictionsFound 11 items-rw-r--r-- 3 root hdfs 0 2017-03-21 19:16 /user/root/predictions/_SUCCESS-rw-r--r-- 3 root hdfs 51000 2017-03-21 19:16 /user/root/predictions/part-00000-rw-r--r-- 3 root hdfs 51000 2017-03-21 19:16 /user/root/predictions/part-00001-rw-r--r-- 3 root hdfs 51000 2017-03-21 19:16 /user/root/predictions/part-00002-rw-r--r-- 3 root hdfs 51000 2017-03-21 19:16 /user/root/predictions/part-00003-rw-r--r-- 3 root hdfs 51000 2017-03-21 19:16 /user/root/predictions/part-00004-rw-r--r-- 3 root hdfs 51000 2017-03-21 19:16 /user/root/predictions/part-00005-rw-r--r-- 3 root hdfs 51000 2017-03-21 19:16 /user/root/predictions/part-00006-rw-r--r-- 3 root hdfs 51000 2017-03-21 19:16 /user/root/predictions/part-00007-rw-r--r-- 3 root hdfs 51000 2017-03-21 19:16 /user/root/predictions/part-00008-rw-r--r-- 3 root hdfs 51000 2017-03-21 19:16 /user/root/predictions/part-00009

查看其中的某一个文件，可看到里面保存的是测试集的标签和根据模型预测的结果。

# hadoop fs -cat /user/root/predictions/part-000002017-03-21T19:16:40.795694 Label: 7, Prediction: 72017-03-21T19:16:40.795729 Label: 2, Prediction: 22017-03-21T19:16:40.795741 Label: 1, Prediction: 12017-03-21T19:16:40.795750 Label: 0, Prediction: 02017-03-21T19:16:40.795759 Label: 4, Prediction: 42017-03-21T19:16:40.795769 Label: 1, Prediction: 12017-03-21T19:16:40.795778 Label: 4, Prediction: 42017-03-21T19:16:40.795787 Label: 9, Prediction: 92017-03-21T19:16:40.795796 Label: 5, Prediction: 62017-03-21T19:16:40.795805 Label: 9, Prediction: 92017-03-21T19:16:40.795814 Label: 0, Prediction: 02017-03-21T19:16:40.795822 Label: 6, Prediction: 62017-03-21T19:16:40.795831 Label: 9, Prediction: 92017-03-21T19:16:40.795840 Label: 0, Prediction: 02017-03-21T19:16:40.795848 Label: 1, Prediction: 12017-03-21T19:16:40.795857 Label: 5, Prediction: 52017-03-21T19:16:40.795866 Label: 9, Prediction: 92017-03-21T19:16:40.795875 Label: 7, Prediction: 72017-03-21T19:16:40.795883 Label: 3, Prediction: 32017-03-21T19:16:40.795892 Label: 4, Prediction: 42017-03-21T19:16:40.795901 Label: 9, Prediction: 92017-03-21T19:16:40.795909 Label: 6, Prediction: 62017-03-21T19:16:40.795918 Label: 6, Prediction: 6

Spark集群和tensorflow job task的对应关系，如下图，spark集群起了4个executor，其中一个作为PS, 另外3个作为worker，而谁做ps谁做worker是由Yarn和spark调度的。

7.22.23.png

Cluster spec: {'ps': ['ochadoop02:50060'], 'worker': ['ochadoop04:52150', 'ochadoop03:52733', 'ochad