1. Start MesosClusterDispatcher on Marathon

MesosClusterDispatcher accepts client’s submit of spark-work, and negotiate with Mesos to obtain runtime resources. (Ref.)

The only parameter required for MesosClusterDispatcher to start is the --master, we can localte the leader of 3 masters by mesos-resolve. Here is the script to start MesosClusterDispatcher in cluster mode:

# find leader of 3 masters
$ mesos-resolve zk://mesos-master-1:2181,mesos-master-2:2181,mesos-master-3:2181/mesos
10.240.0.5:5050

# bring up MesosClusterDispatcher in foreground.
$ /opt/spark/current/bin/spark-class org.apache.spark.deploy.mesos.MesosClusterDispatcher \
  --master mesos://10.240.0.5:5050
Using Spark's default log4j profile: org/apache/spark/log4j-defaults.properties
16/01/22 03:01:01 INFO MesosClusterDispatcher: Registered signal handlers for [TERM, HUP, INT]
16/01/22 03:01:01 INFO MesosClusterDispatcher: Recovery mode in Mesos dispatcher set to: NONE
...
16/01/22 03:01:01 INFO MesosClusterUI: Started MesosClusterUI at http://10.240.0.4:8081
I0122 03:01:02.112972  6348 sched.cpp:166] Version: 0.26.0
I0122 03:01:02.114768  6342 sched.cpp:264] New master detected at master@10.240.0.5:5050
I0122 03:01:02.114918  6342 sched.cpp:274] No credentials provided. Attempting to register without authentication
I0122 03:01:02.117938  6342 sched.cpp:643] Framework registered with 00c2bae0-a02f-40ac-a5e1-3028a35810b0-0001
16/01/22 03:01:02 INFO MesosClusterScheduler: Registered as framework ID 00c2bae0-a02f-40ac-a5e1-3028a35810b0-0001
16/01/22 03:01:02 INFO Utils: Successfully started service on port 7077.
16/01/22 03:01:02 INFO MesosRestServer: Started REST server for submitting applications on port 7077
$

We will run MesosClusterDispatcher in Marathon, and locate the dispatcher through mesos-dns: ping spark-dispatcher.marathon.mesos. Here is the marathon config file:

$ cat def-tasks/marathon-spark-dispatcher.json 
{
    "id": "spark-dispatcher",
    "cmd": "MY_LEADER=`mesos-resolve zk://mesos-master-1:2181,mesos-master-2:2181,mesos-master-3:2181/mesos`; /opt/spark/current/bin/spark-class org.apache.spark.deploy.mesos.MesosClusterDispatcher --master \"mesos://$MY_LEADER\"",
    "mem": 512,
    "cpus": 0.5,
    "instances": 1,
    "disk": 0.0,
    "ports": [0]
}

# bring up 
$ curl -i -H 'Content-Type: application/json' \
  -d@./def-tasks/marathon-spark-dispatcher.json mesos-master-1:8080/v2/apps

A MesosClusterDispatcher is running as an application in Marathon: alt text

2. Run a spark test: SparkPi

# ssh into mesos-slave-X
$ /opt/spark/current/bin/spark-submit \
  --name SparkPiTestApp \
  --class org.apache.spark.examples.SparkPi \
  --master mesos://spark-dispatcher.marathon.mesos:7077 \
  --deploy-mode cluster \
  --executor-memory 5G \
  --total-executor-cores 30 \
  /opt/spark/current/lib/spark-examples-1.5.2-hadoop2.6.0.jar \
  100
  
# kill
$ /opt/spark/current/bin/spark-submit \
  --master mesos://spark-dispatcher.marathon.mesos:7077 \
  --deploy-mode cluster \
  --kill driver-20160122035057-0001

The SparkPiTestApp driver program is running: alt text

The output of SparkPiTestApp alt text