Cooperative Computing Lab
CCL | Software | Download | Manuals | Forum | Papers
CCL Home

Research

Software Community Operations

Work Queue Tutorial

This tutorial will have you install CCTools and will take you through some distributed computation examples using Work Queue.

  1. Getting Started
    1. Tutorial Setup
    2. Set Environment Variables
  2. Work Queue Example
    1. Setup
    2. Analysis
    3. Running
  3. Exercise
    1. Using Project Names and Catalog Server
  4. Homework
    1. Chaining Tasks

Getting Started

Tutorial Setup

The setup for this tutorial follows the setup described and installed in the Makeflow tutorial.

Login in to the below head node where you will install CCTools and run this tutorial.

ssh USERNAME@ec2-23-23-23-169.compute-1.amazonaws.com

Please replace USERNAME with the username in the sheet of paper handed out to you. Use the password given there as well.

Set Environment Variables

For this tutorial, you will need to add your CCTools directory to your $PATH:

export PATH=~/cctools/bin:${PATH}

We will use both Python and Perl in this tutorial. They need to be able to find our installed packages. So set these environment variables for running the Python and Perl examples:

export PYTHONPATH=${PYTHONPATH}:~/cctools/lib/python2.6/site-packages export PERL5LIB=${PERL5LIB}:~/cctools/lib/perl5/site_perl

Work Queue program running the Simulation Executable

It is simple to write a program in C/Perl/Python (or any language with appropriate Work Queue bindings) which can generate tasks to be queued for Work Queue. In this example, we will create 20 simulation tasks using simulation.py.

Setup

mkdir ~/cctools-tutorial/wq cd ~/cctools-tutorial/wq wget http://www.nd.edu/~ccl/software/tutorials/ccgrid13/wq/simulation.py wget http://www.nd.edu/~ccl/software/tutorials/ccgrid13/wq/input.txt wget http://www.nd.edu/~ccl/software/tutorials/ccgrid13/wq/wq.py

Analysis

We will analyze the code in wq.py to study a workflow implemented using the Python bindings for the Work Queue API.

$ cat wq.py #!/usr/bin/env python from work_queue import * import sys try: Q = WorkQueue(port = 0) except: print "could not instantiate Work Queue master" sys.exit(1) print "Listening on port %d." % Q.port print "Submitting 20 simulation tasks..." for i in range(0, 20): infile = "input.txt" outfile = "file.%0.2x" % i command = "python simulation.py %d < %s > %s" % (i, infile, outfile) T = Task(command) T.specify_file("simulation.py", "simulation.py", WORK_QUEUE_INPUT, cache = True) T.specify_file(infile, infile, WORK_QUEUE_INPUT, cache = False) T.specify_file(outfile, outfile, WORK_QUEUE_OUTPUT, cache = False) taskid = Q.submit(T) print "done." print "Waiting for tasks to complete..." while not Q.empty(): T = Q.wait(5) if T: print "task (id# %d) complete: %s (return code %d)" % (T.id, T.command, T.return_status) print "done."

Here we load the work_queue Python binding. Python will look in PYTHONPATH which is setup in your environment.

This instantiates a Work Queue master to which you may submit work. Setting the port to 0 instructs Work Queue to pick an arbitrary port to bind on.

We create a task which takes a shell command argument. The first task created in this workflow will have the command:

T = Task("python simulation.py 0 < input.txt > file.00");

Each task usually depends on a number of files to run. These include the executable and any input files. Here we specify the simulation.py executable and its input infile. Notice that we specify both simulation.py and infile twice when calling specify_file. The first argument is the name of the file on the master and the second argument is the name of the file we want created on the worker. Usually these filenames will be the same as in this example.

We specify the output file, outfile, which we want transferred back to the master.

At this point we have finished the description of our task and it is ready to be submitted for execution on the Work Queue workers. Q.submit submits this task.

At this point we wish to wait for all submitted tasks to complete. So long as the queue is not empty, we continue to call Q.wait waiting for the result of a task we submitted.

Here we call Q.wait(5) which takes a timeout argument. The call to wait will return a finished task which allows us to analyze the return_status or output. In this example, we set the timeout to 5 seconds which allows our application to do other things if a task is taking an inordinate amount of time to complete. We could have used the constant WORK_QUEUE_WAITFORTASK to wait indefinitely until a task completes.

Perl equivalent

The Perl code of the above program is in wq.pl and is shown here:

$ cat wq.pl #!/usr/bin/perl use work_queue; my $q = work_queue_create(0); if (not defined($q)) { print "could not instantiate Work Queue master\n"; exit 1; } $port = work_queue_port($q); print "Listening on port $port.\n"; print "Submitting 20 simulation tasks..."; for (my $i = 0; $i < 20; $i++) { my $infile = "input.txt"; my $outfile = sprintf("file.%0.2x", $i); my $command = "python simulation.py $i < $infile > $outfile"; my $t = work_queue_task_create($command); work_queue_task_specify_file($t,"simulation.py","simulation.py",$WORK_QUEUE_INPUT,$WORK_QUEUE_CACHE); work_queue_task_specify_file($t,$infile,$infile,$WORK_QUEUE_INPUT,$WORK_QUEUE_CACHE); work_queue_task_specify_file($t,$outfile,$outfile,$WORK_QUEUE_OUTPUT,$WORK_QUEUE_NOCACHE); my $taskid = work_queue_submit($q, $t); } print "done." print "Waiting for tasks to complete...\n"; while (not work_queue_empty($q)) { my $t = work_queue_wait($q, 5); if (defined($t)) { print "task (id#$t->{taskid}) complete:$t->{command_line} (return code $t->{return_status})\n"; work_queue_task_delete($t); } } print "done.\n"; work_queue_delete($q); exit 0;

You can download this program using:

wget http://www.nd.edu/~ccl/software/tutorials/ccgrid13/wq/wq.pl

Running

We are now going to run the Work Queue program shown above for running the simulation tasks. Pick one of the programs to run:

To run the Python program (wq.py) do: To run the Perl program (wq.pl) do:
python wq.py perl wq.pl

If you encounter an error when running this step, be sure you did not forget to setup your environment.

When a Work Queue program is run, it prints the port on which it is listening for connections from the workers. For example:

$ python wq.py Listening on port XXXX. ... $ perl wq.pl Listening on port XXXX. ...

We have successfully started a master and found the port and hostname on which it is listening. The master now is waiting for connections from workers so that it can dispatch the submitted tasks for execution.

To start a work_queue_worker for this master, open another terminal window and login into the EC2 head node with your given username:

ssh USERNAME@ec2-23-23-23-169.compute-1.amazonaws.com

As before, replace USERNAME with the username provided to you (and its password).

Add the CCTools directory to our $PATH as before:

export PATH=~/cctools/bin:${PATH}

In this newly opened terminal, start a work_queue_worker for the master by giving it the port and hostname of the master. Also, enable the debugging output (-d all).

work_queue_worker -t 10 -d all localhost XXXX ...

replacing XXXX with the port the Work Queue master program is listening on.

The debug output for the worker will appear on your terminal. When the tasks are finished, the worker will quit after the 10 second timeout.

Running Work Queue workers on EC2

We can also submit workers to run on dedicated EC2 instances using the ec2_submit_workers script.

Start one of the Work Queue master programs again:

$ python wq.py Listening on port XXXX. ... $ perl wq.pl Listening on port XXXX. ...

Now, goto the second terminal and start 5 workers on the cluster with a 10 second timeout as:

$ bash ec2_submit_workers -t 10 ec2-23-23-23-169.compute-1.amazonaws.com XXXX ccgrid-tutorial ~/.ssh/ccgrid-tutorial.pem 5

where XXXX is the port of Work Queue master program.

Note: Similar to ec2_submit_workers, we also have condor_submit_workers, sge_submit_workers, torque_submit_workers, pbs_submit_workers for submitting workers to Condor, SGE, Torque, and PBS respectively.

Exercise

Using Project Names and Catalog Server

In this exercise, you will use the catalog server and project name feature to have the workers automatically find and establish connection with the Work Queue master in wq.py (or wq.pl) without being provided with the master's hostname and port.

To do this, you will first modify the master program to use the default catalog server (running at catalog.cse.nd.edu) by enabling the catalog mode. Then, you will provide the master with a project name that will be advertised to the default catalog server.

For Work Queue programs written in Python, you will use the specify_master_mode() API to set the catalog mode. Similarly, you will modify the program to specify a project name using the specify_name() API. An example of their usage is given below:

try: Q = WorkQueue(port = 0) except: sys.exit(1) Q.specify_master_mode(WORK_QUEUE_MASTER_MODE_CATALOG) Q.specify_name("MYPROJECT")

For Work Queue programs in C/Perl, you will use the work_queue_specify_master_mode() API to set the catalog mode. To specify a project name, you will use work_queue_specify_name(). For example, a Work Queue program written in Perl will look like:

my $q = work_queue_create(0); if (not defined($q)) { print "could not instantiate Work Queue master\n"; exit 1; } work_queue_specify_master_mode($q, WORK_QUEUE_MASTER_MODE_CATALOG) work_queue_specify_name($q, "MYPROJECT")

NOTE: Pick your own distinct name for MYPROJECT.

You will then start workers for your Work Queue master by specifying the option to use the default catalog server (-a option) and the project name of your master (-N option). Example:

$ work_queue_worker -d all -a -N MYPROJECT

(Replace MYPROJECT with the distinct name you chose for the Work Queue master).

Homework

Chaining Tasks

The goal of this exercise is to change the workflow to chain the executions of simulation.py so that the output of one simulation is the input of another. For this exercise, the workflow should look like:

chaining simulation

Because our simulation.py is sophisticated and runs on average for 5 seconds, in this example we will only do 5 instances of the simulation (instead of 100) so it takes about 25 seconds.

For this exercise, remember that when you run Q.submit(T), it finalizes the task and allows it to be sent to a worker for execution. You will need to wait for the output from a worker to come back before sending out the next one. As before, you can wait for the completion of a task using Q.wait(5).

Once you have finished implementing your version of the chained simulation, you may compare with this python solution or perl solution.