Running REEF examples

Persistent Evaluators / Distributed Shell demo

I suggest to use the persistent evaluators demo instead of the distributed shell - it has more functionality, and is interactive, which makes a great demo.

There's a script REEF/bin/persistent_eval.sh that can run PE demo in interactive or batch mode. Here are some examples:

• ./persistent_eval.sh - run interactively in local mode
• ./persistent_eval.sh -local false - run interactively on YARN
• ./persistent_eval.sh -cmd date - run command date once on all evaluators in local mode (i.e. act as a distributed shell)
• ./persistent_eval.sh -cmd date -local false - same as above, on YARN
• ./persistent_eval.sh -cmd date -local false -num_runs 10 - run command date on YARN ten times, reusing the evaluators.

Try it out on vnectar-21

P.S. few more things:

• press ^D or type exit to quit the interactive shell.
• the script runs on Linux only;
• on Windows, mvn -PPersistentEval in reef-examples will start the local interactive shell.
• I output results to the log (i.e. stderr) and to stdout. If you send stderr to /dev/null, the demo looks quite pretty, e.g. try mvn -PPersistentEval 2> /dev/null
• use rlwrap ./persistent_eval.sh (rlwrap is installed on vnectar-21) to have command line history and other readline functionality in interactive mode

Suspend/Resume demo

Terminal 1:

#
./suspend.sh -local false -cycles 120 -delay 1 -port 7008
#

Terminal 2:

#
./suspend-control.sh -port 7008 -cmd suspend -task <your task ID here>
#
./suspend-control.sh -port 7008 -cmd resume -task <your task ID here>
#

This demo starts a bunch of Tasks that do nothing but cycle from 0 to N and sleep for 1 second on each iteration. Tasks can be suspended and resumed; on resume, task restores its state and starts its idle cycle from the state where it had been suspended before.

To run the job:

• ./suspend.sh -cycles 600 -delay 1 -port 7008 - run in local mode, and cycle from 0 to 600 with 1 second delay between iterations; it will listen on localhost:7008 for suspend/resume commands.
• ./suspend.sh -local false -cycles 600 -delay 1 -port 7008 - same as above, on YARN.

Look for task IDs in the log - you will need them to suspend/resume the tasks.

To suspend/resume the tasks:

• run ./suspend-control.sh -port 7008 -cmd suspend -task <your task ID here> to suspend the task
• ./suspend-control.sh -port 7008 -cmd resume -task <your task ID here> to resume

you should see suspend/resume messages in the job logs.

Few things to notice:

• ./suspend-control.sh connects to the localhost only;
• shell scripts run only on Linux;
• to run suspend demo on Windows, do mvn -PSuspendDemo - this will run the tasks in local mode for 10 cycles without suspending/resuming them;
• if the task ID is not found, the entire job fails;
• because of issue #235, currently there are no update messages from the tasks in the client log; however, there are messages when tasks complete or get suspended or resumed.

So the general strategy for the suspend/resume demo would be:

• open two windows with ssh to vnectar-21;
• in one, run ./suspend.sh -cycles 120 (i.e. roughly for 2 minutes);
• look for the task id in the job logs;
• does not work: watch for updates from running tasks
• in other window, run ./suspend-control.sh to suspend that task
• see the message in the job logs about task being suspended
• does not work: watch for updates from all tasks except the suspended one
• run ./suspend-control.sh again to resume the task
• see the resumed task message in the job logs
• does not work: see updates from all tasks again
• wait to see some tasks complete
• does not work: see the resumed task catching up
• wait more to see the resumed task complete

Batch Gradient Descent

On vnectar: To run, execute bin/bgd.sh. There isn't much to look at at this point beyond nice log files flying by.

Locally: cd reef-examples followed by mvn -PBGD

PageRank

To run, execute bin/pagerank.sh.

grep "vertex\[" the evaluator logs to see the final un-normalized ranks of all the vertices. grep "update counts" the evaluator logs to see the number of updates performed per vertex.

Run pagerank on other graphs with

bin/pagerank.sh -graphfile="/grid/0/rsync/t-branm/p2p-Gnutella08.txt" 
-sourceStr="com.microsoft.reef.examples.pagerank.FileGraphPartitionSource"

You should change to .level=INFO in logging.properties and rebuild reef-common, or else this will be very slow.

cp /grid/0/rsync/t-branm/logging.properties /grid/0/rsync/reef-sailfish/REEF/reef-common/src/main/resources/com/microsoft/reef/logging.properties
cd reef-common
mvn clean install -DskipTests

# undo the above
git checkout reef-common/src/main/resources/com/microsoft/reef/logging.properties
cd reef-common
mvn clean install -DskipTests

Run on YARN on vnectar with

bin/pagerank.sh -local false -graphfile="/grid/0/rsync/t-branm/p2p-Gnutella08.txt" 
-sourceStr="com.microsoft.reef.examples.pagerank.FileGraphPartitionSource"

#
bin/pagerank.sh -local false -graphfile="/grid/0/rsync/t-branm/p2p-Gnutella08.txt" -sourceStr="com.microsoft.reef.examples.pagerank.FileGraphPartitionSource" 2>&1
#

Run pez MR examples

cd /grid/0/rsync/reef-sailfish/reef-applications/pez/bin
./hadoop-examples.sh randomtextwriter test.txt
./hadoop-examples.sh pi 10 100 # 3.148
./hadoop-examples.sh wordcount
./hadoop-examples.sh grep

Killall Reef Launchers:

 kill -9 `jps | grep Launcher | cut -f1,1 -d' '`