老男孩的成长之路

在Kubernetes中用Alluxio加速Spark数据访问(

2020-08-07  本文已影响0人  路人甲java

简介: 本文档主要说明怎样在k8s上用alluxio加速spark的数据访问。文档将演示结合spark、alluxio和k8s完成一个对文件单词进行计数的任务。在实验中,我们将关闭alluxio的short-circuit的功能,验证spark executor与alluxio worker之间的通信是否通过网络栈完成。

相关文章:

1.背景信息

1.1 alluxio

Alluxio是一个开源的基于内存的分布式存储系统,适合作为云上大数据和AI / ML的数据编排方案。Alluxio可以同时管理多个底层文件系统,将不同的文件系统统一在同一个名称空间下,让上层客户端可以自由访问统一名称空间内的不同路径,不同存储系统的数据。

alluxio的short-circuit功能可以使alluxio客户端直接访问alluxio worker所在主机的工作存储,而不需要通过网络栈与alluxio worker完成通信,可以提高性能。

1.2 spark operator

Spark-operator用于管理k8s集群中spark job。通过spark-operator可以在k8s集群中创建、查看和删除spark job。

2.前提条件

本文档的操作依赖如下的一些条件:

image.png

3.实验步骤

实验步骤主要包括如下几步:

下面将对每个步骤进行说明:

3.1 部署alluxio

进入容器服务应用目录,在右上角的搜索框中搜索"alluxio",然后进入alluxio主界面,如图:

image.png

选择“参数”,修改配置中properties部分的"alluxio.user.short.circuit.enabled"值为"false",然后选择将alluxio安装到目标集群上(本次实验的集群为"ack-create-by-openapi-1"),最后点击创建,如图

image.png

点击创建后,使用kubectl给待安装的alluxio组件的节点打上标签"alluxio=true",首先查看该集群有哪些节点:

$ kubectl get nodes -o wide
NAME                      STATUS   ROLES    AGE   VERSION            INTERNAL-IP    EXTERNAL-IP   OS-IMAGE                               KERNEL-VERSION            CONTAINER-RUNTIME
cn-beijing.192.168.8.12   Ready    master   21d   v1.16.6-aliyun.1   192.168.8.12   <none>        Aliyun Linux 2.1903 (Hunting Beagle)   4.19.57-15.1.al7.x86_64   docker://19.3.5
cn-beijing.192.168.8.13   Ready    master   21d   v1.16.6-aliyun.1   192.168.8.13   <none>        Aliyun Linux 2.1903 (Hunting Beagle)   4.19.57-15.1.al7.x86_64   docker://19.3.5
cn-beijing.192.168.8.14   Ready    master   21d   v1.16.6-aliyun.1   192.168.8.14   <none>        Aliyun Linux 2.1903 (Hunting Beagle)   4.19.57-15.1.al7.x86_64   docker://19.3.5
cn-beijing.192.168.8.15   Ready    <none>   21d   v1.16.6-aliyun.1   192.168.8.15   <none>        Aliyun Linux 2.1903 (Hunting Beagle)   4.19.57-15.1.al7.x86_64   docker://19.3.5
cn-beijing.192.168.8.16   Ready    <none>   21d   v1.16.6-aliyun.1   192.168.8.16   <none>        Aliyun Linux 2.1903 (Hunting Beagle)   4.19.57-15.1.al7.x86_64   docker://19.3.5
cn-beijing.192.168.8.17   Ready    <none>   21d   v1.16.6-aliyun.1   192.168.8.17   <none>        Aliyun Linux 2.1903 (Hunting Beagle)   4.19.57-15.1.al7.x86_64   docker://19.3.5

可以看到有三个worker节点,分别为:

我们给是三个节点都打上标签"alluxio=true":

$ kubectl label nodes cn-beijing.192.168.8.15 \
  cn-beijing.192.168.8.16 \
  cn-beijing.192.168.8.17 \
  alluxio=true

使用kubectl查看各个pod是否都处于running状态:

$ kubectl get po -n alluxio
NAME                   READY   STATUS    RESTARTS   AGE
alluxio-master-0       2/2     Running   0          4h1m
alluxio-worker-5zg26   2/2     Running   0          4h1m
alluxio-worker-ckmr9   2/2     Running   0          4h1m
alluxio-worker-dvgvd   2/2     Running   0          4h1m

验证alluxio是否处于ready:

$ kubectl exec -ti alluxio-master-0 -n alluxio bash

//下面步骤alluxio-master-0 pod中执行
bash-4.4# alluxio fsadmin report capacity

Capacity information for all workers:
    Total Capacity: 3072.00MB
        Tier: MEM  Size: 3072.00MB
    Used Capacity: 0B
        Tier: MEM  Size: 0B
    Used Percentage: 0%
    Free Percentage: 100%

Worker Name      Last Heartbeat   Storage       MEM
192.168.8.15    0                capacity      1024.00MB
                                  used          0B (0%)
192.168.8.16    0                capacity      1024.00MB
                                  used          0B (0%)
192.168.8.17    0                capacity      1024.00MB
                                  used          0B (0%)

3.2 部署spark-operator

进入容器服务应用目录,在右上角的搜索框中搜索"ack-spark-operator",然后进入ack-spark-operator主界面,如图:
[图片上传失败...(image-5d3f28-1596765851121)]

选择将ack-spark-operator安装到目标集群上(本次实验的集群为"ack-create-by-openapi-1"),然后点击创建,如图:


image.png

本次实验将会使用sparkctl向k8s集群提交一个spark job,需要将sparkctl安装到我们在"2.前提条件"中所提到的实验环境"alluxio-test"中:

$ wget http://spark-on-k8s.oss-cn-beijing.aliyuncs.com/sparkctl/sparkctl-linux-amd64 -O /usr/local/bin/sparkctl
$ chmod +x /usr/local/bin/sparkctl

3.3 制作spark docker镜像

spark下载页面下载所需的spark版本,本次实验选择的saprk版本为2.4.6。运行如下命令下载spark:

$ cd /root
$ wget https://mirror.bit.edu.cn/apache/spark/spark-2.4.6/spark-2.4.6-bin-hadoop2.7.tgz
#

下载完成后,执行解压操作:

$ tar -xf spark-2.4.6-bin-hadoop2.7.tgz
$ export SPARK_HOME=/root/spark-2.4.6-bin-hadoop2.7

spark docker镜像是我们提交spark任务时使用到的镜像,这个镜像中需要包含alluxio client jar包。使用如下的命令获取alluxio client jar包:

$ id=$(docker create alluxio/alluxio-enterprise:2.2.1-1.4)
$ docker cp $id:/opt/alluxio/client/alluxio-enterprise-2.2.1-1.4-client.jar \
    $SPARK_HOME/jars/alluxio-enterprise-2.2.1-1.4-client.jar
$ docker rm -v $id 1>/dev/null

alluxio client jar包准备好以后,开始构建镜像:

$ docker build -t spark-alluxio:2.4.6 -f kubernetes/dockerfiles/spark/Dockerfile $SPARK_HOME

请记住镜像名称“spark-alluxio:2.4.6”,在向k8s提交spark job中会用到这个信息。

镜像构建完成以后,对镜像的处理有两种方式:

3.4 上传文件到alluxio

文章开头提到过:本次实验是提交一个spark job到k8s中,该spark job的目标是对某一个文件统计每一个单词出现的次数。现在需要把这个文件传到alluxio存储上,这里为了方便,直接把alluxio master中/opt/alluxio-2.3.0-SNAPSHOT/LICENSE(文件路径可能因alluxio版本有点差异)这个文件传到alluxio上。

使用"kubectl exec"进入alluxio master pod,并拷贝当前目录下的LICENSE文件到alluxio的根目录中:

$ kubectl exec -ti alluxio-master-0  -n alluxio bash
//下面步骤alluxio-master-0 pod中执行
bash-4.4# alluxio fs copyFromLocal LICENSE /

接着查看一下LICENSE这个文件分成的block被alluxio放到哪些worker上了。

$ kubectl exec -ti alluxio-master-0 -n alluxio bash
//下面步骤alluxio-master-0 pod中执行

bash-4.4# alluxio fs stat /LICENSE
/LICENSE is a file path.
FileInfo{fileId=33554431, fileIdentifier=null, name=LICENSE, path=/LICENSE, ufsPath=/opt/alluxio-2.3.0-SNAPSHOT/underFSStorage/LICENSE, length=27040, blockSizeBytes=67108864, creationTimeMs=1592381889733, completed=true, folder=false, pinned=false, pinnedlocation=[], cacheable=true, persisted=false, blockIds=[16777216], inMemoryPercentage=100, lastModificationTimesMs=1592381890390, ttl=-1, lastAccessTimesMs=1592381890390, ttlAction=DELETE, owner=root, group=root, mode=420, persistenceState=TO_BE_PERSISTED, mountPoint=false, replicationMax=-1, replicationMin=0, fileBlockInfos=[FileBlockInfo{blockInfo=BlockInfo{id=16777216, length=27040, locations=[BlockLocation{workerId=8217561227881498090, address=WorkerNetAddress{host=192.168.8.17, containerHost=, rpcPort=29999, dataPort=29999, webPort=30000, domainSocketPath=, tieredIdentity=TieredIdentity(node=192.168.8.17, rack=null)}, tierAlias=MEM, mediumType=MEM}]}, offset=0, ufsLocations=[]}], mountId=1, inAlluxioPercentage=100, ufsFingerprint=, acl=user::rw-,group::r--,other::r--, defaultAcl=}
Containing the following blocks:
BlockInfo{id=16777216, length=27040, locations=[BlockLocation{workerId=8217561227881498090, address=WorkerNetAddress{host=192.168.8.17, containerHost=, rpcPort=29999, dataPort=29999, webPort=30000, domainSocketPath=, tieredIdentity=TieredIdentity(node=192.168.8.17, rack=null)}, tierAlias=MEM, mediumType=MEM}]}

可以看到LICENSE这个文件只有一个block(id为16777216),被放在了ip为192.168.8.17的k8s节点上。我们使用kubectl查看该节点名称为cn-beijing.192.168.8.17

$ kubectl get nodes -o wide
NAME                      STATUS   ROLES    AGE   VERSION            INTERNAL-IP    EXTERNAL-IP   OS-IMAGE                               KERNEL-VERSION            CONTAINER-RUNTIME
cn-beijing.192.168.8.12   Ready    master   21d   v1.16.6-aliyun.1   192.168.8.12   <none>        Aliyun Linux 2.1903 (Hunting Beagle)   4.19.57-15.1.al7.x86_64   docker://19.3.5
cn-beijing.192.168.8.13   Ready    master   21d   v1.16.6-aliyun.1   192.168.8.13   <none>        Aliyun Linux 2.1903 (Hunting Beagle)   4.19.57-15.1.al7.x86_64   docker://19.3.5
cn-beijing.192.168.8.14   Ready    master   21d   v1.16.6-aliyun.1   192.168.8.14   <none>        Aliyun Linux 2.1903 (Hunting Beagle)   4.19.57-15.1.al7.x86_64   docker://19.3.5
cn-beijing.192.168.8.15   Ready    <none>   21d   v1.16.6-aliyun.1   192.168.8.15   <none>        Aliyun Linux 2.1903 (Hunting Beagle)   4.19.57-15.1.al7.x86_64   docker://19.3.5
cn-beijing.192.168.8.16   Ready    <none>   21d   v1.16.6-aliyun.1   192.168.8.16   <none>        Aliyun Linux 2.1903 (Hunting Beagle)   4.19.57-15.1.al7.x86_64   docker://19.3.5
cn-beijing.192.168.8.17   Ready    <none>   21d   v1.16.6-aliyun.1   192.168.8.17   <none>        Aliyun Linux 2.1903 (Hunting Beagle)   4.19.57-15.1.al7.x86_64   docker://19.3.5

3.5 提交spark job

下面的步骤将提交一个spark job到k8s集群中,该job主要是计算alluxio中/LICENSE文件的每个单词出现的次数。

在步骤3.4中我们获取到LICENSE这个文件所包含的block都在节点cn-beijing.192.168.8.17上,此次实验中,我们通过指定node selector让spark driver和spark executor都运行在节点cn-beijing.192.168.8.17,验证在关闭alluxio的short-circuit功能的情况下,spark executor和alluxio worker之间的通信是否通过网络栈完成。

首先生成提交spark job的yaml文件:

$ export SPARK_ALLUXIO_IMAGE=<步骤3.3中制作的image,即spark-alluxio:2.4.6>
$ export ALLUXIO_MASTER="alluxio-master-0"
$ export TARGET_NODE=<步骤3.4获取到的LICENSE文件的block存储的节点,即cn-beijing.192.168.8.17>
$ cat > /tmp/spark-example.yaml <<- EOF
apiVersion: "sparkoperator.k8s.io/v1beta2"
kind: SparkApplication
metadata:
  name: spark-count-words
  namespace: default
spec:
  type: Scala
  mode: cluster
  image: "$SPARK_ALLUXIO_IMAGE"
  imagePullPolicy: Always
  mainClass: org.apache.spark.examples.JavaWordCount
  mainApplicationFile: "local:///opt/spark/examples/jars/spark-examples_2.11-2.4.6.jar"
  arguments:
    - alluxio://${ALLUXIO_MASTER}.alluxio:19998/LICENSE
  sparkVersion: "2.4.6"
  restartPolicy:
    type: Never
  volumes:
    - name: "test-volume"
      hostPath:
        path: "/tmp"
        type: Directory
  driver:
    cores: 1
    coreLimit: "1200m"
    memory: "512m"
    labels:
      version: 2.4.6
    serviceAccount: spark
    volumeMounts:
      - name: "test-volume"
        mountPath: "/tmp"
    nodeSelector:
      kubernetes.io/hostname: "$TARGET_NODE"
  executor:
    cores: 1
    instances: 1
    memory: "512m"
    labels:
      version: 2.4.6
    nodeSelector:
      kubernetes.io/hostname: "$TARGET_NODE"
    volumeMounts:
      - name: "test-volume"
        mountPath: "/tmp"
EOF

然后,使用sparkctl提交spark job:

$ sparkctl create /tmp/spark-example.yaml

4.实验结果

当提交任务后,使用kubectl查看spark driver的日志:

$ kubectl get po -l spark-role=driver
NAME                                 READY   STATUS      RESTARTS   AGE
spark-alluxio-1592296972094-driver   0/1     Completed   0          4h33m

$ kubectl logs spark-alluxio-1592296972094-driver --tail 20

USE,: 3
Patents: 2
d): 1
comment: 1
executed: 1
replaced: 1
mechanical: 1
20/06/16 13:14:28 INFO SparkUI: Stopped Spark web UI at http://spark-alluxio-1592313250782-driver-svc.default.svc:4040
20/06/16 13:14:28 INFO KubernetesClusterSchedulerBackend: Shutting down all executors
20/06/16 13:14:28 INFO KubernetesClusterSchedulerBackend$KubernetesDriverEndpoint: Asking each executor to shut down
20/06/16 13:14:28 WARN ExecutorPodsWatchSnapshotSource: Kubernetes client has been closed (this is expected if the application is shutting down.)
20/06/16 13:14:28 INFO MapOutputTrackerMasterEndpoint: MapOutputTrackerMasterEndpoint stopped!
20/06/16 13:14:28 INFO MemoryStore: MemoryStore cleared
20/06/16 13:14:28 INFO BlockManager: BlockManager stopped
20/06/16 13:14:28 INFO BlockManagerMaster: BlockManagerMaster stopped
20/06/16 13:14:28 INFO OutputCommitCoordinator$OutputCommitCoordinatorEndpoint: OutputCommitCoordinator stopped!
20/06/16 13:14:28 INFO SparkContext: Successfully stopped SparkContext
20/06/16 13:14:28 INFO ShutdownHookManager: Shutdown hook called
20/06/16 13:14:28 INFO ShutdownHookManager: Deleting directory /var/data/spark-2f619243-59b2-4258-ba5e-69b8491123a6/spark-3d70294a-291a-423a-b034-8fc779244f40
20/06/16 13:14:28 INFO ShutdownHookManager: Deleting directory /tmp/spark-054883b4-15d3-43ee-94c3-5810a8a6cdc7

最后我们登陆到alluxio master上,查看相关指标统计到的值:

$ kubectl exec -ti alluxio-master-0 -n alluxio bash
//下面步骤alluxio-master-0 pod中执行
bash-4.4# alluxio fsadmin report metrics
Cluster.BytesReadAlluxio  (Type: COUNTER, Value: 290.47KB)
Cluster.BytesReadAlluxioThroughput  (Type: GAUGE, Value: 22.34KB/MIN)
Cluster.BytesReadDomain  (Type: COUNTER, Value: 0B)
Cluster.BytesReadDomainThroughput  (Type: GAUGE, Value: 0B/MIN)

BytesReadAlluxio和BytesReadAlluxioThroughput代表数据从网络栈传输;BytesReadDomain和BytesReadDomainThroughput代表数据从domain socket传输。可以看到所有数据都是从网络栈传输的(即使spark executor和LICENSE文件的block在同一k8s节点上)。

上一篇下一篇

猜你喜欢

热点阅读