LWR

This project is a Python server application that allows a Galaxy server to run jobs on remote systems (including Windows) without requiring a shared mounted file systems. Unlike traditional Galaxy job runners - input files, scripts, and config files may be transferred to the remote system, the job is executed, and the result downloaded back to the Galaxy server.

Full documentation for the project can be found on Read The Docs.

Configuring Galaxy

Galaxy job runners are configured in Galaxy’s job_conf.xml file. Some small examples of how to configure this can be found here, but be sure to checkout job_conf.xml.sample_advanced in your Galaxy code base or on Bitbucket for complete information.

Downloading LWR

The LWR server application is distributed as a Python project and can be obtained via mercurial from bitbucket.org using the following command:

hg clone http://bitbucket.org/jmchilton/lwr

LWR Dependencies

Several Python packages must be installed to run the LWR server. These can either be installed into a Python virtualenv or into your system wide Python environment using easy_install. Instructions for both are outlined below. Additionally, if DRMAA is going to be used to communicate with a cluster, this dependency must be installed as well - again see note below.

virtualenv

The script setup_venv.sh distributed with the LWR server is a short-cut for *nix machines to setup a Python environment (including the installation of virtualenv). Full details for installation suitable for *nix are as follows. These instructions can work for Windows as well but generally the easy_install instructions below are more robust for Window’s environments.

  1. Install virtualenv (if not available):

    pip install virtualenv
    
  2. Create a new Python environment:

    virtualenv .venv
    
  3. Activate environment (varies by OS).

From a Linux or MacOS terminal:

. .venv/bin/activate

From a Windows terminal:

.venv\Scripts\activate
  1. Install required dependencies into this virtual environment:

    pip install -r requirements.txt
    

easy_install

Install python setuptools for your platform, more details on how to do this can be found here.

The easy_install command line application will be installed as part of setuptools. Use the following command to install the needed packages via easy_install:

easy_install paste wsgiutils PasteScript PasteDeploy webob six psutil

DRMAA

If your LWR instance is going to communicate with a cluster via DRMAA, in addition to the above dependencies, a DRMAA library will need to be installed and the python dependency drmaa will need to be installed as well.:

. .venv/bin/activate; pip install drmaa

or:

easy_install drmaa

Running the LWR Server Application

*nix Instructions

The LWR can be started and stopped via the run.sh script distributed with the LWR.:

./run.sh --daemon
./run.sh --stop-daemon

These commands will start and stop the WSGI web server in daemon mode. In this mode, logs are writtin to paster.log.

If uWSGI, circus and/or chassuette, are available, more sophisticated web servers can be launched via this run.sh command. See the script for more details.

Alternative Cross Platform Instructions (Windows and *nix)

The paster command line application will be installed as part of the previous dependency installation process. This application can be used to start and stop a paste web server running the LWR. This can be done by executing the following command:

The server may be ran as a daemon via the command:

paster serve server.ini --daemon

When running as daemon, the server may be stopped with the following command:

paster serve server.ini --stop-daemon

If you setup a virtual environment for the LWR you will need to activate this before executing these commands.

Configuring the LWR Server Application

Rename the server.ini.sample file distributed with LWR to server.ini, and edit the values therein to configure the server application. Default values are specified for all configuration options that will work if LWR is running on the same host as Galaxy. However, the parameter “host” must be specified for remote submissions to the LWR server to run properly. The server.ini file contains documentation for many configuration parameters you may want to modify.

Some advanced configuration topics are discussed below.

Security

Out of the box the LWR essentially allows anyone with network access to the LWR server to execute arbitrary code and read and write any files the web server can. Hence, in most settings steps should be taken to secure the LWR server.

LWR Web Server

The LWR web server can be configured to use SSL and to require the client (i.e. Galaxy) to pass along a private token authorizing use.

pyOpenSSL is required to configure an LWR web server to server content via HTTPS/SSL. This dependency can be difficult to install and seems to be getting more difficult. Under Linux you will want to ensure the needed dependencies to compile pyOpenSSL are available - for instance in a fresh Ubuntu image you will likely need:

sudo apt-get install libffi-dev python-dev libssl-dev

Then pyOpenSSL can be installed with the following command (be sure to source your virtualenv if setup above):

pip install pyOpenSSL

Under Windows only older versions for pyOpenSSL are installable via pre- compiled binaries (i.e. using easy_install) so it might be good to use non- standard sources such as eGenix.

Once installed, you will need to set the option ssl_pem in server.ini. This parameter should reference an OpenSSL certificate file for use by the Python paste server. This parameter can be set to * to automatically generate such a certificate. Such a certificate can manually be generated by the following method:

$ openssl genrsa 1024 > host.key
$ chmod 400 host.key
$ openssl req -new -x509 -nodes -sha1 -days 365  \
          -key host.key > host.cert
$ cat host.cert host.key > host.pem
$ chmod 400 host.pem

More information can be found in the paste httpserver documentation.

Finally, in order to force Galaxy to authorize itself, you will want to specify a private token - by simply setting private_key to some long random string in server.ini.

Once SSL has been enabled and a private token configured, Galaxy job destinations should include a private_token parameter to authenticate these jobs.

LWR Message Queue

If LWR is processing Galaxy requests via a message queue instead of a web server the underlying security mechanisms of the message queue should be used to secure the LWR communication - configuring SSL with the LWR and a private_token above are not required.

This will likely consist of setting some combination of amqp_connect_ssl_ca_certs, amqp_connect_ssl_keyfile, amqp_connect_ssl_certfile, amqp_connect_ssl_cert_reqs, in LWR’s server.ini file. See server.ini.sample for more details and the Kombo documentation for even more information.

Customizing the LWR Environment

In more sophisticated deployments, the LWR’s environment will need to be tweaked - for instance to define a DRMAA_LIBRARY_PATH environment variable for the drmaa Python module or to define the location to a find a location of Galaxy (via GALAXY_HOME) if certain Galaxy tools require it or if Galaxy metadata is being set by the LWR. The recommend way to do this is to copy local_env.sh.sample to local_env.sh and customize it.

This file of deployment specific environment tweaks will be source by run.sh if it exists as well as by other LWR scripts in more advanced usage scenarios.

Job Managers (Queues)

By default the LWR will maintain its own queue of jobs. While ideal for simple deployments such as those targetting a single Windows instance, if the LWR is going to be used on more sophisticate clusters, it can be configured to maintain multiple such queues with different properties or to delegate to external job queues (via DRMAA, qsub/qstat CLI commands, or Condor).

For more information on configured external job managers, see the job managers documentation.

Warning: If you are using DRMAA, be sure to define DRMAA_LIBRARY_PATH in local_env.sh defined above.

Galaxy Tools

Some Galaxy tool wrappers require a copy of the Galaxy codebase itself to run. Such tools will not run under Windows, but on *nix hosts the LWR can be configured to add the required Galaxy code a jobs PYTHON_PATH by setting GALAXY_HOME environment variable in the LWR’s local_env.sh file (described above).

Caching (Experimental)

LWR and its clients can be configured to cache job input files. For some workflows this can result in a significant decrease in data transfer and greater throughput. On the LWR side - the property file_cache_dir in server.ini must be set. See Galaxy’s job_conf.xml for information on configuring the client.

More discussion on this can be found in this galaxy-dev mailing list thread and future plans and progress can be tracked on this Trello card.

Message Queue (Experimental)

Galaxy and the LWR can be configured to communicate via a message queue instead of an LWR web server. In this mode, the LWR will download files from and upload files to Galaxy instead of the inverse - this may be very advantageous if the LWR needs to be deployed behind a firewall or if the Galaxy server is already setup (via proxy web server) for large file transfers.

To bind the LWR server to a message queue, one needs to first ensure the kombu Python dependency is installed (pip install kombu). Once this available, simply set the message_queue_url property in server.ini to the correct URL of your configured AMQP endpoint.

Configuring your AMQP compatible message queue is beyond the scope of this document - see RabbitMQ for instance for more details (other MQs should work also).

Testing

https://travis-ci.org/jmchilton/lwr.png?branch=master https://coveralls.io/repos/jmchilton/lwr/badge.png?branch=master

A simple sanity test can be run against a running LWR server by executing the following command (replace the URL command with the URL of your running LWR application):

python run_client_tests.py --url=http://localhost:8913

Development

This project is distributed with unit and integration tests (many of which will not run under Windows), the following command will install the needed python components to run these tests.:

pip install -r dev-requirements.txt

The following command will then run these tests:

nosetests

The following command will then produce a coverage report corresponding to this test and place it in the coverage_html_report subdirectory of this project.:

coverage html

Job Managers

By default the LWR will maintain its own queue of jobs. Alternatively, the LWR can be configured to maintain multiple such queues with different properties or to delegate to external job queues (via DRMAA, qsub/qstat CLI commands, or Condor).

To change the default configuration, rename the file job_managers.ini.sample distributed with the LWR to job_managers.ini and modify it to reflect your desired configuration, and finally uncomment the line #job_managers_config = job_managers.ini in server.ini.

Likely the cleanest way to interface with an external queueing system is going to be DRMAA. In this case, one should likely copy local_env.sh.sample to local_env.sh and update it to set DRMAA_LIBRARY_PATH to point to the correct libdrmaa.so file. Also, the Python drmaa module must be installed (see more information about drmaa dependency <https://lwr.readthedocs.org/#job-managers>).

Sample Configuration

## Default job manager is queued and runs 1 concurrent job.
[manager:_default_]
type = queued_python
max_concurrent_jobs=1

## Create a named queued (example) and run as many concurrent jobs as
## server has cores. The Galaxy LWR url should have /managers/example 
## appended to it to use a named manager such as this.
#[manager:example]
#type=queued_python
#max_concurrent_jobs=*

## DRMAA backed manager (vanilla).
## Be sure drmaa Python module install and DRMAA_LIBRARY_PATH points
## to a valid DRMAA shared library file. You may also need to adjust 
## LD_LIBRARY_PATH.
#[manager:_default_]
#type=queued_drmaa
#native_specification=-P bignodes -R y -pe threads 8

## Condor backed manager.
#[manager:_default_]
#type=queued_condor
## Optionally, additional condor submission parameters can be 
## set as follows:
#submit_universe=vanilla
#submit_request_memory=32
#submit_requirements=OpSys == "LINUX" && Arch =="INTEL"
#submit_rank=Memory >= 64
## These would set universe, request_memory, requirements, and rank
## in the condor submission file to the specified values. For
## more information on condor submission files see the following link:
## http://research.cs.wisc.edu/htcondor/quick-start.html.

## CLI Manager Locally
## Manage jobs via command-line execution of qsub, qdel, stat.
#[manager:_default_]
#type=queued_cli
#job_plugin=Torque

## CLI Manager via Remote Shell
## Manage jobs via qsub, qdel, qstat on remote host `queuemanager` as
## Unix user `queueuser`.
#[manager:_default_]
#type=queued_cli
#job_plugin=Torque
#shell_plugin=SecureShell
#shell_hostname=queuemanager
#shell_username=queueuser

## DRMAA (via external users) manager.  
## This variant of the DRMAA manager will run jobs as the supplied user.
#[manager:_default_]
#type=queued_external_drmaa
#production=true
#chown_working_directory_script=scripts/chown_working_directory.bash
#drmaa_kill_script=scripts/drmaa_kill.bash
#drmaa_launch_script=scripts/drmaa_launch.bash


## NOT YET IMPLEMENTED. PBS backed manager.
#[manager:_default_]
#type=queued_pbs

## Disable server-side LWR queuing (suitable for older style LWR use
## when queues were maintained in Galaxy.) Deprecated, will be removed
## at some point soon.
#[manager:_default_]
#type=unqueued


## MQ-Options:
## If using a message queue the LWR will actively monitor status of jobs
## in order to issue status update messages. The following options are
## then available to any managers.

## Minimum seconds between polling intervals (increase to reduce resources
## consumed by the LWR).
#min_polling_interval = 0.5 

Running Jobs As External User

TODO: Fill out this section with information from this thread <http://dev.list.galaxyproject.org/Managing-Data-Locality-tp4662438.html>.

Galaxy Configuration

Examples

The most complete and updated documentation for configuring Galaxy job destinations is Galaxy’s job_conf.xml.sample_advanced file (check it out on Bitbucket). These examples just provide a different LWR-centric perspective on some of the documentation in that file.

Simple Windows LWR Web Server

The following Galaxy job_conf.xml assumes you have deployed a simple LWR web server to the Windows host windowshost.examle.com on the default port (8913) with a private_key (defined in server.ini) of 123456789changeme. Most Galaxy jobs will just route use Galaxy’s local job runner but msconvert and proteinpilot will be sent to the LWR server on windowshost.examle.com. Sophisticated tool dependency resolution is not available for Windows-based LWR servers so ensure the underlying application are on the LWR’s path.

<?xml version="1.0"?>
<job_conf>
    <plugins>
        <plugin id="local" type="runner" load="galaxy.jobs.runners.local:LocalJobRunner"/>
        <plugin id="lwr" type="runner" load="galaxy.jobs.runners.lwr:LwrJobRunner"/>
    </plugins>
    <handlers>
        <handler id="main"/>        
    </handlers>
    <destinations default="local">
        <destination id="local" runner="local"/>
        <destination id="win_lwr" runner="lwr">
            <param id="url">https://windowshost.examle.com:8913/</param>
            <param id="private_token">123456789changeme</param>
        </destination>
    </destinations>
    <tools>
        <tool id="msconvert" destination="win_lwr" />
        <tool id="proteinpilot" destination="win_lwr" />
	</tools>    
</job_conf>

Targeting a Linux Cluster (LWR Web Server)

The following Galaxy job_conf.xml assumes you have a very typical Galaxy setup - there is a local, smaller cluster that mounts all of Galaxy’s data (so no need for the LWR) and a bigger shared resource that cannot mount Galaxy’s files requiring the use of the LWR. This variant routes some larger assembly jobs to the remote cluster - namely the trinity and abyss tools. Be sure the underlying applications required by the trinity and abyss tools are the LWR path or set tool_dependency_dir in server.ini and setup Galaxy env.sh-style packages definitions for these applications).

<?xml version="1.0"?>
<job_conf>
    <plugins>
        <plugin id="drmaa" type="runner" load="galaxy.jobs.runners.drmaa:DRMAAJobRunner">
        <plugin id="lwr" type="runner" load="galaxy.jobs.runners.lwr:LwrJobRunner"/>
    </plugins>
    <handlers>
        <handler id="main"/>        
    </handlers>
    <destinations default="drmaa">
        <destination id="local_cluster" runner="drmaa">
            <param id="native_specification">-P littlenodes -R y -pe threads 4</param>            
        </destination>
        <destination id="remote_cluster" runner="lwr">
            <param id="url">http://remotelogin:8913/</param>
            <param id="submit_native_specification">-P bignodes -R y -pe threads 16</param>
            <!-- Look for trinity package at remote location - define tool_dependency_dir
            in the LWR server.ini file.
            -->
            <param id="dependency_resolution">remote</params>
            <!-- Use more correct parameter generation for *nix. Needs testing on Windows
                 servers before this becomes default. -->
            <param id="rewrite_parameters">True</params>
        </destination>
    </destinations>
    <tools>
        <tool id="trinity" destination="remote_cluster" />
        <tool id="abyss" destination="remote_cluster" />
	</tools>
</job_conf>

For this configuration, on the LWR side be sure to set a DRMAA_LIBRARY_PATH in local_env.sh, install the Python drmaa module, and configure a DRMAA job manager (example job_managers.ini follows).

[manager:_default_]
type=queued_drmaa

Targeting a Linux Cluster (LWR over Message Queue)

For LWR instances sitting behind a firewall a web server may be impossible. If the same LWR configuration discussed above is additionally configured with a message_queue_url of amqp://rabbituser:rabb8pa8sw0d@mqserver:5672// in server.ini the following Galaxy configuration will cause this message queue to be used for communication. This is also likely better for large file transfers since typically your production Galaxy server will be sitting behind a high-performance proxy but not the LWR.

<?xml version="1.0"?>
<job_conf>
    <plugins>
        <plugin id="drmaa" type="runner" load="galaxy.jobs.runners.drmaa:DRMAAJobRunner">
        <plugin id="lwr" type="runner" load="galaxy.jobs.runners.lwr:LwrJobRunner">
            <!-- Must tell LWR where to send files. -->
            <param id="galaxy_url">https://galaxyserver</param>
            <!-- Message Queue Connection (should match message_queue_url in LWR's server.ini)
            -->
            <param id="url">amqp://rabbituser:rabb8pa8sw0d@mqserver:5672//</param>
        </plugin>
    </plugins>
    <handlers>
        <handler id="main"/>        
    </handlers>
    <destinations default="drmaa">
        <destination id="local_cluster" runner="drmaa">
            <param id="native_specification">-P littlenodes -R y -pe threads 4</param>            
        </destination>
        <destination id="remote_cluster" runner="lwr">
            <!-- Tell Galaxy where files are being store on remote system, no
                 web server it can simply ask for this information.
            -->
            <param id="jobs_directory">/path/to/remote/lwr/lwr_staging/</param>
            <!-- Invert file transfers - have LWR initiate downloads during preprocessing
                 and uploads during postprocessing. -->
            <param id="default_file_action">remote_transfer</param>

            <!-- Remaining parameters same as previous example -->
            <param id="submit_native_specification">-P bignodes -R y -pe threads 16</param>
            <param id="dependency_resolution">remote</params>
            <param id="rewrite_parameters">True</params>            
        </destination>
    </destinations>
    <tools>
        <tool id="trinity" destination="remote_cluster" />
        <tool id="abyss" destination="remote_cluster" />
	</tools>
</job_conf>

Targeting Apache Mesos (Prototype)

See commit message for initial work on this and this post on galaxy-dev.

Etc...

There are many more options for configuring what paths get staging/unstaged how, how Galaxy metadata is generated, running jobs as the real user, defining multiple job managers on the LWR side, etc.... If you ever have any questions please don’t hesistate to ask John Chilton (jmchilton@gmail.com).

File Actions

Most of the parameters settable in Galaxy’s job configuration file job_conf.xml are straight forward - but specifing how Galaxy and the LWR stage various files may benefit from more explaination.

As demonstrated in the above default_file_action describes how inputs, outputs, etc... are staged. The default transfer has Galaxy initiate HTTP transfers. This makes little sense in the contxt of message queues so this should be overridden and set to remote_transfer which causes the LWR to initiate the file transfers. Additional options are available including none, copy, and remote_copy.

In addition to this default - paths may be overridden based on various patterns to allow optimization of file transfers in real production infrastructures where various systems mount different file stores and file stores with different paths on different systems.

To do this, the LWR destination in job_conf.xml may specify a parameter named file_action_config. This needs to be some config file path (if relative, relative to Galaxy’s root) like lwr_actions.yaml (can be YAML or JSON - but older Galaxy’s only supported JSON).

paths: 
  # Use transfer (or remote_transfer) if only Galaxy mounts a directory.
  - path: /galaxy/files/store/1
    action: transfer

  # Use copy (or remote_copy) if remote LWR server also mounts the directory
  # but the actual compute servers do not.
  - path: /galaxy/files/store/2
    action: copy

  # If Galaxy, the LWR, and the compute nodes all mount the same directory
  # staging can be disabled altogether for given paths.
  - path: /galaxy/files/store/3
    action: none

  # Following block demonstrates specifying paths by globs as well as rewriting
  # unstructured data in .loc files.
  - path: /mnt/indices/**/bwa/**/*.fa
    match_type: glob
    path_types: unstructured  # Set to *any* to apply to defaults & unstructured paths.
    action: transfer
    depth: 1  # Stage whole directory with job and just file.

  # Following block demonstrates rewriting paths without staging. Useful for
  # instance if Galaxy's data indices are mounted on both servers but with
  # different paths.
  - path: /galaxy/data
    path_types: unstructured
    action: rewrite
    source_directory: /galaxy/data
    destination_directory: /work/galaxy/data

Configuring a Public LWR Server

An LWR server can be pointed at a Galaxy toolbox XML file and opened to the world. By default, an LWR is allowed to run anything Galaxy (or other client) sends it. The toolbox and referenced tool files are used to restrict what what the LWR will run.

This can be sort of thought of as web services defined by Galaxy tool files - with all the advantages (dead simple configuration for clients, ability to hide details related date and computation) and disadvantages (lack of reproducibility if the LWR server goes away, potential lack of transparency).

Securing a Public LWR

The following options should be set in server.ini to configure a public LWR server.

  • assign_ids=uuid - By default the LWR will just the ids Galaxy instances. Setting this setting to uuid will result in each job being assigned a UUID, ensuring different clients will not and cannot interfer with each other.
  • tool_config_files=/path/to/tools.xml - As noted above, this is used to restrict what tools clients can run. All tools on public LWR servers should have validators for commands (and optionally for configfiles) defined. The syntax for these elements can be found in the ValidatorTest test case.

Writing Secure Tools

Validating in this fashion is complicated and potentially error prone, so it is advisable to keep command-lines as simple as possible. configfiles and reorganizing parameter handling in wrappers scripts can assist in this.

Consider the following simple example:

tool.xml:

<tool>
   <command interpreter="python">wrapper.py --input1 'Text' --input2 'Text2' --input3 4.5</command>
   ...

wrapper.py:

def main():
     parser = OptionParser()
     parser.add_option("--input1")
     parser.add_option("--input2")
     parser.add_option("--input3")
     (options, args) = parser.parse_args()

Even this simple example is easier to validate and secure if it is reworked as so:

tool.xml:

<tool>
  <configfiles>
    <configfile name="args">--input1 'Text' --input2 'Text2' --input3 4.5</configfile>
  </configfiles>
  <command interpreter="python">wrapper.py $args</command>
  ...

wrapper.py:

import sys, shlex

def main():
    args_config = sys.argv[1]
    args_string = open(args_config, "r").read()

    parser = OptionParser()
    parser.add_option("--input1")
    parser.add_option("--input2")
    parser.add_option("--input3")
    (options, args) = parser.parse_args(shlex.split(args_string))

Indices and tables