scripts/staging/google-cloud - systemds

tree: c95a875d783c5e7db55ffc571d9b23262a8316d3

README.md

scripts/staging/google-cloud/README.md

Create a dataproc cluster

Create a cluster name

CLUSTERNAME=dp-systemds

Set Dataproc cluster region

gcloud config set dataproc/region us-central1

Now, create a new cluster

gcloud dataproc clusters create reference

gcloud dataproc clusters create ${CLUSTERNAME} \
  --scopes=cloud-platform \
  --tags systemds \
  --zone=us-central1-c \
  --worker-machine-type n1-standard-2 \
  --worker-boot-disk-size 500 \
  --master-machine-type n1-standard-2 \
  --master-boot-disk-size 500 \
  --image-version 2.0

Submit a Spark job to the cluster

Jobs can be submitted via a Cloud Dataproc API jobs.submit request

Submit an example job using gcloud tool from the Cloud Shell command line

Test that the cluster is setup properly:

gcloud dataproc jobs submit spark --cluster ${CLUSTERNAME} \
  --class org.apache.spark.examples.SparkPi \
  --jars file:///usr/lib/spark/examples/jars/spark-examples.jar -- 1000

Add SystemDS library to the cluster

SSH into the cluster, download the artifacts from https://dlcdn.apache.org/systemds/ and copy jar file in the lib folder.

gcloud compute ssh ${CLUSTERNAME}-m --zone=us-central1-c
wget https://dlcdn.apache.org/systemds/2.2.0/systemds-2.2.0-bin.zip
unzip -q systemds-2.2.0-bin.zip
mkdir /usr/lib/systemds
cp systemds-2.2.0-bin/systemds-2.2.0.jar /usr/lib/systemds

Run SystemDS as a Spark job

gcloud dataproc jobs submit spark --cluster ${CLUSTERNAME} \
  --class org.apache.sysds.api.DMLScript \
  --jars file:///usr/lib/systemds/systemds-2.2.0.jar -- 1000

Job info and connect

List all the jobs:

gcloud dataproc jobs list --cluster ${CLUSTERNAME}

To get output of a specific job note jobID and in the below command replace jobID.

gcloud dataproc jobs wait jobID

Resizing the cluster

For intensive computations, to add more nodes to the cluster either to speed up.

Existing cluster configuration

gcloud dataproc clusters describe ${CLUSTERNAME}

Add preemptible nodes to increase cluster size:

gcloud dataproc clusters update ${CLUSTERNAME} --num-preemptible-workers=1

Note: workerConfig and secondaryWorkerConfig will be present.

SSH into the cluster

SSH into the cluster (primary node) would provide fine grained control of the cluster.

gcloud compute ssh ${CLUSTERNAME}-m --zone=us-central1-c

Note: For the first time, we run ssh command on Cloud Shell, it will generate SSH keys for your account.

The --scopes=cloud-platform would allow us to run gcloud inside the cluster too. For example,

gcloud dataproc clusters list --region=us-central1

to exit the cluster primary instance

logout

Deleting the cluster

gcloud dataproc clusters delete ${CLUSTERNAME}