（5）Spark on K8S 计算资源配置

简介

在Spark On K8s场景下, 涉及计算资源的设置, 除了driver.cores、 executor.cores、 driver.memory 和executor.memory这4个Spark自有的参数外, 还会受到K8s对资源控制的影响, 例如k8s request和limits。为此我们有必要了解在spark on ks的场景下，这些参数是否存在相互影响，以及在实际的应用中应该如何设置，才是最好的选择。

在spark on k8s场景下，可以通过两种方式设置，spark的资源参数，一个是在spark-submit提交程序的时候，通过–conf 设置资源参数，另外一种就是Pod Template中设置资源参数。

通过–conf配置

用例一：

driver.cores	executor.cores	driver.request.cores	executor.request.cores	driver.limit.cores	executor.limit.cores
1	1	1.5	1.5	2	2

driver.cores < driver.limit.cores driver可以正常启动运行

executor.cores < executor.limit.cores executor可以正常启动运行

用例二：

driver.cores	executor.cores	driver.request.cores	executor.request.cores	driver.limit.cores	executor.limit.cores
2	2	1.5	1.5	2	2

driver.cores = driver.limit.cores driver可以正常启动运行

executor.cores = executor.limit.cores executor可以正常启动运行

用例三：按理说是不能正常启动的因为需要的资源超过了限制资源

driver.cores	executor.cores	driver.request.cores	executor.request.cores	driver.limit.cores	executor.limit.cores
3	3	1.5	1.5	2	2

driver.cores > driver.limit.cores driver能启动运行

executor.cores > executor.limit.cores executor能启动运行

结论：根据实际情况来看，在–conf 配置方式下，K8S的request和limit资源参数实际上并不起作用，起作用的还是spark自有的driver和executor配置

提交命令如下：

./spark-3.2.3/bin/spark-submit \
 --name BasicWordCount \
 --verbose \
  --master k8s://https://172.16.252.105:6443 \
 --deploy-mode cluster \
 --conf spark.network.timeout=300 \
 --conf spark.executor.instances=1 \
 --conf spark.driver.cores=1 \
 --conf spark.executor.cores=1 \
 --conf spark.kubernetes.driver.request.cores=1.5 \
 --conf spark.kubernetes.driver.limit.cores=2 \
 --conf spark.kubernetes.executor.request.cores=1.5 \
 --conf spark.kubernetes.executor.limit.cores=2 \
 --conf spark.driver.memory=1024m \
 --conf spark.executor.memory=1024m \
 --conf spark.driver.extraJavaOptions="-Dio.netty.tryReflectionSetAccessible=true -Dlog.file=/opt/spark/logs/driver-res-1.log" \
 --conf spark.executor.extraJavaOptions="-Dio.netty.tryReflectionSetAccessible=true -Dlog.file=/opt/spark/logs/executor-res-1.log" \
 --conf spark.eventLog.enabled=true \
 --conf spark.eventLog.dir=file:///opt/spark/eventLog \
 --conf spark.history.fs.logDirectory=file:///opt/spark/eventLog \
 --conf spark.kubernetes.namespace=apache-spark \
 --conf spark.kubernetes.authenticate.driver.serviceAccountName=spark-service-account \
 --conf spark.kubernetes.authenticate.executor.serviceAccountName=spark-service-account \
 --conf spark.kubernetes.container.image.pullPolicy=IfNotPresent \
 --conf spark.kubernetes.container.image=apache/spark:v3.2.3 \
 --conf spark.kubernetes.driver.volumes.persistentVolumeClaim.spark-logs-pvc.mount.path=/opt/spark/logs \
 --conf spark.kubernetes.driver.volumes.persistentVolumeClaim.spark-logs-pvc.options.claimName=spark-logs-pvc \
 --conf spark.kubernetes.driver.volumes.persistentVolumeClaim.eventlog-pvc.mount.path=/opt/spark/eventLog \
 --conf spark.kubernetes.driver.volumes.persistentVolumeClaim.eventlog-pvc.options.claimName=spark-historyserver-pvc \
 --conf spark.kubernetes.executor.volumes.persistentVolumeClaim.spark-logs-pvc.mount.path=/opt/spark/logs \
 --conf spark.kubernetes.executor.volumes.persistentVolumeClaim.spark-logs-pvc.options.claimName=spark-logs-pvc \
 --conf spark.kubernetes.driverEnv.TZ=Asia/Shanghai \
 --conf spark.kubernetes.executorEnv.TZ=Asia/Shanghai \
 --class org.fblinux.StreamWordCount \
http://172.16.252.105:8080/spark-on-k8s-demo-1.0-SNAPSHOT.jar \
http://172.16.252.105:8080/wc_input.txt

执行完成之后，通过spark-history验证driver和executor的资源配置：

通过Pod Template配置

用例一：

driver.cores	executor.cores	driver.request.cores	executor.request.cores	driver.limit.cores	executor.limit.cores
1	3	1.5	1.5	2	2

driver.cores < driver.limit.cores driver可以启动但会报错

executor.cores > executor.limit.cores executor不可以启动

用例二：

driver.cores	executor.cores	driver.request.cores	executor.request.cores	driver.limit.cores	executor.limit.cores
3	3	1.5	1.5	2	2

driver.cores > driver.limit.cores driver不可以启动

executor.cores > executor.limit.cores executor不可以启动

用例3：

driver.memory	executor.memory	driver.request.memory	executor.request.memory	driver.limit.memory	executor.limit.memory
4g	4g	1.5g	1.5g	2g	2g

driver.memory > driver.limit.memory driver可以启动

executor.memory < executor.limit.memory executor可以启动

结论：使用pod template可以限制cpu核数，但不能限制内存。

driver template文件：

[root@k8s-demo001 nginx_down]# cat driver-res.yaml 
apiVersion: v1
kind: Pod
metadata:
  labels:
    app.kubernetes.io/name: apache-spark
    app.kubernetes.io/instance: apache-spark
    app.kubernetes.io/version: v3.2.3
  namespace: apache-spark
  name: driver
spec:
  serviceAccountName: spark-service-account
  hostAliases:
    - ip: "172.16.252.105"
      hostnames:
        - "k8s-demo001"
    - ip: "172.16.252.134"
      hostnames:
        - "k8s-demo002"
    - ip: "172.16.252.135"
      hostnames:
        - "k8s-demo003"
    - ip: "172.16.252.136"
      hostnames:
        - "k8s-demo004"
  containers:
    - image: apache/spark:v3.2.3
      name: driver
      imagePullPolicy: IfNotPresent
      env:
        - name: TZ
          value: Asia/Shanghai
      resources:
        requests:
          cpu: 1.5
          memory: 1.5Gi
        limits:
          cpu: 2
          memory: 2Gi
      volumeMounts:
        - name: spark-historyserver   # 挂载eventLog归档目录
          mountPath: /opt/spark/eventLog
        - name: spark-logs   # 挂载日志
          mountPath: /opt/spark/logs
  volumes:
    - name: spark-historyserver
      persistentVolumeClaim:
        claimName: spark-historyserver-pvc
    - name: spark-logs
      persistentVolumeClaim:
        claimName: spark-logs-pvc

executor template文件：

[root@k8s-demo001 nginx_down]# cat executor-res.yaml 
apiVersion: v1
kind: Pod
metadata:
  labels:
    app.kubernetes.io/name: apache-spark
    app.kubernetes.io/instance: apache-spark
    app.kubernetes.io/version: v3.2.3
  namespace: apache-spark
  name: executor
spec:
  serviceAccountName: spark-service-account
  hostAliases:
    - ip: "172.16.252.105"
      hostnames:
        - "k8s-demo001"
    - ip: "172.16.252.134"
      hostnames:
        - "k8s-demo002"
    - ip: "172.16.252.135"
      hostnames:
        - "k8s-demo003"
    - ip: "172.16.252.136"
      hostnames:
        - "k8s-demo004"
  containers:
    - image: apache/spark:v3.2.3
      name: executor
      imagePullPolicy: IfNotPresent
      env:
        - name: TZ
          value: Asia/Shanghai
      resources:
        requests:
          cpu: 1.5
          memory: 1.5Gi
        limits:
          cpu: 2
          memory: 2Gi
      volumeMounts:
        - name: spark-historyserver   # 挂载eventLog归档目录
          mountPath: /opt/spark/eventLog
        - name: spark-logs   # 挂载日志
          mountPath: /opt/spark/logs
  volumes:
    - name: spark-historyserver
      persistentVolumeClaim:
        claimName: spark-historyserver-pvc
    - name: spark-logs
      persistentVolumeClaim:
        claimName: spark-logs-pvc

提交命令：

./spark-3.2.3/bin/spark-submit \
 --name BasicWordCount \
 --verbose \
 --master k8s://https://172.16.252.105:6443 \
 --deploy-mode cluster \
 --conf spark.network.timeout=300 \
 --conf spark.executor.instances=1 \
 --conf spark.driver.cores=1 \
 --conf spark.executor.cores=1 \
 --conf spark.driver.memory=4g \
 --conf spark.executor.memory=4g \
 --conf spark.driver.extraJavaOptions="-Dio.netty.tryReflectionSetAccessible=true -Dlog.file=/opt/spark/logs/driver-res-6.log" \
 --conf spark.executor.extraJavaOptions="-Dio.netty.tryReflectionSetAccessible=true -Dlog.file=/opt/spark/logs/executor-res-6.log" \
 --conf spark.eventLog.enabled=true \
 --conf spark.eventLog.dir=file:///opt/spark/eventLog \
 --conf spark.history.fs.logDirectory=file:///opt/spark/eventLog \
 --conf spark.kubernetes.namespace=apache-spark \
 --conf spark.kubernetes.authenticate.driver.serviceAccountName=spark-service-account \
 --conf spark.kubernetes.authenticate.executor.serviceAccountName=spark-service-account \
 --conf spark.kubernetes.container.image.pullPolicy=IfNotPresent \
 --conf spark.kubernetes.container.image=apache/spark:v3.2.3 \
 --conf spark.kubernetes.driver.podTemplateFile=http://172.16.252.105:8080/driver-res.yaml \
 --conf spark.kubernetes.executor.podTemplateFile=http://172.16.252.105:8080/executor-res.yaml \
 --class org.fblinux.BasicWordCount \
http://172.16.252.105:8080/spark-on-k8s-demo-1.0-SNAPSHOT.jar \
http://172.16.252.105:8080/wc_input.txt

结论

在实际应用中，CPU配置方面，建议将spark.driver.cores和spark.executor.cores与pod template的limits.cores配置一致，如果不使用pod template，只需要配置spark.driver.cores和spark.executor.cores即可，不需要配置其他参数，不然不光不起作用，放进去还会引起误解内存方面：无论是方式一还是方式二，对内存的限制都不起作用，实际使用中，将spark.driver.memory和spark.executor.memory与pod template的limits.memory配置一致即可，或者只配置spark.driver.memory和spark.executor.memory

转载请注明：西门飞冰的博客 » （5）Spark on K8S 计算资源配置

一	二	三	四	五	六	日
« 六
				1	2	3
4	5	6	7	8	9	10
11	12	13	14	15	16	17
18	19	20	21	22	23	24
25	26	27	28	29	30