Connecting Galaxy to a compute cluster

Contributors

Authors: Nate Coraor Björn Grüning Nicola Soranzo Helena Rasche

Questions

• How to connect Galaxy to a compute cluster?

• How can I configure job dependent resources, like cores, memory for my DRM?

Objectives

• Understand all components of the Galaxy job running stack

• Understand how the job_conf.xml file controls Galaxy’s jobs subsystem

• Know how to map tools to job destinations

• The various ways in which tools can be mapped to destinations, both statically and dynamically

Galaxy Job Configuration

• Configured in config/job_conf.xml
• XML format with macro support
• Major components:
  • Plugins: distributed resource manager (DRM) modules to load
  • Handlers: job handler processes managing the lifecycle of jobs
  • Destinations: where to send jobs, and what parameters to run those jobs with
  • Tool to destination/handler mappings
  • Resource selection mappings: give users job execution options on the tool form
  • Limits: job runtime limits, e.g. the max number of concurrent jobs

Speaker Notes

• The job_conf file is a very powerul galaxy configuration piece critical to smooth cluster operation.
• Written in XML it connects your server with the available cluster resources.
• You can configure it in myriad ways.
• Study the advanced sample provided with codebase once you get a basic understanding.
• There are several major components of the job conf file.
• Plugins, handlers, destinations, tools, resources, and limits.
• We’ll go into detail on each of these in the tutorial.

Why cluster?

Running jobs on the Galaxy server negatively impacts Galaxy UI performance

Even adding one other host helps

Can restart Galaxy without interrupting jobs

Speaker Notes

• Galaxy itself is not resource hungry, but the jobs often are.
• Offloading the jobs to different machines is a more sustainable and reliable setup.
• This can prevent user jobs from making Galaxy unresponsive.

Plugins

Correspond to job runner plugins in lib/galaxy/jobs/runners

.left[Plugins for:]

• local
• Slurm (DRMAA subclass)
• DRMAA: SGE, PBS Pro, LSF, Torque
• HTCondor
• Torque: Using the pbs_python library
• Pulsar: Galaxy’s own remote job management system
• Command Line Interface (CLI) via SSH
• Kubernetes
• Go-Docker
• Chronos

Speaker Notes

• Galaxy supports plugins for various job runners covering most of the popular DRMs.
• The Galaxy community also maintains its own job management system called Pulsar.
• If the scheduler you use is missing, talk to us!

Cluster library stack (DRMAA)

Speaker Notes

• The cluster library stack we use in this tutorial will use DRMAA.
• DRMAA is an interface that many distributed resource managers provide.
• Galaxy can use DRMAA to interact with these in an agnostic manner.
• However, there are more underlying technologies that you are going to depend on.
• You don’t need to have an in-depth understanding to run cluster deployment correctly.

Handlers

Define which Galaxy processes are job handlers

• id attribute should match the --server-name param value of a process
• Dedicated handlers can be reserved, e.g. for small, high throughput jobs
• The list of plugins that are loaded by a job handler can be limited using <plugin> subelements (e.g. when the DRMAA plugin needs to be loaded with different library paths)
• Not defined in job_conf.xml when using job handler mules
• Defines how jobs are assigned to individual processes (use db-skip-locked !)

Speaker Notes

• Handlers are the Galaxy processes which interact with the cluster.
• You can define dedicated handlers for different types of jobs, or to interact with different clusters.
• Additionally, handlers definition in the job configuration controls how jobs are assigned to individual processes.
• There are many options for the assignment process, all are discussed in the advanced sample job configuration.
• db-skip-locked is the best choice for most cases, it enables handlers to grab multiple jobs to work on at once.

Destinations

Define how jobs should be run

• Which plugin? (Slurm, Condor, Pulsar, etc?)
• In a Docker container? Which one?
• DRM params (queue, cores, memory, walltime)?
• Environment (variables e.g. $PATH, source an env file, run a command)?

Speaker Notes

• The destination section of the job configuration file is a map that defines which jobs go where.
• Jobs from any destination, can be processed by any plugin.
• Every job will find a path through this configuration and eventually get dispatched to the matching runner.
• These destinations can specify things like environment variables or resource requirements.

The default job configuration

.left[config/job_conf.xml.sample_basic:]

<?xml version="1.0"?>
<job_conf>
    <plugins>
        <plugin id="local" type="runner" load="galaxy.jobs.runners.local:LocalJobRunner" workers="4"/>
    </plugins>
    <destinations>
        <destination id="local" runner="local"/>
    </destinations>
</job_conf>

Speaker Notes

• This is the default job configuration.
• It uses a local runner with 4 workers, or processes to process jobs.
• As a result if you restart Galaxy, jobs will be lost.

Job Config - Tags

Both destinations and handlers can be grouped by tags

• Allows random selection from multiple resources
• Allows concurrency limits at the destination group level

Speaker Notes

• Tags can be applied to both destinations and handlers.
• This permits selecting randomly amongst the handlers or destinations.
• Tags can help the load distribution.

Job Environment

<env> tag in destinations: configure the job exec environment

Source and command execution will be handled on the remote destination, don’t need to work on the Galaxy server

Speaker Notes

• You can specify environment variables on the destination.
• Galaxy will ensure these are executed in the same environment and ahead of the job.

Limits

Available limits

• Walltime (if not available with your DRM)
• Output size (if any tool output grows larger than this limit)
• Concurrency: Number of “active” (queued or running) jobs

Speaker Notes

• Configuration of job limits is best acommplished using both the DRM provided limits and the ones from Galaxy.
• Walltime is best set in your DRM, while output size is only possible through Galaxy.
• We recommend you set these at the DRM level which is better equipped to terminate misbehaving jobs.
• The most important limit however is usually concurrency.

Concurrency Limits

Available limits

• Number of active jobs per registered user
• Number of active jobs per unregistered user
• Number of active jobs per registered user in a specified destination or destination tag
• Number of total active jobs in a specified destination or destination tag

Speaker Notes

• Using concurrency limits lets you ensure quality of service for everyone.
• By limiting jobs per user, you can prevent a single user from overwhelming the server, and ensure everyone can run jobs.
• Additionally with concurrency limits you can balance your instance between internal and external users.

Shared Filesystem

Most job plugins require a shared filesystem between the Galaxy server and compute.

The exception is Pulsar. More on this in Using heterogeneous compute resources

Speaker Notes

• Most DRMs require a shared filesystem to ensure datasets are available to the jobs.
• Galaxy’s Pulsar does not, and can be used in situations where no shared filesystem is available.

Shared Filesystem

Our simple example works because of two important principles:

1. Some things are located at the same path on Galaxy server and node(s)
   • Galaxy application (/srv/galaxy/server)
   • Tool dependencies
2. Some things are the same on Galaxy server and node(s)
   • Job working directory
   • Input and output datasets

The first can be worked around with symlinks or Pulsar embedded (later)

The second can be worked around with Pulsar REST/MQ (with a performance/throughput penalty)

Speaker Notes

• For the DRMs which require a shared filesystem there are additional requirements.
• First, Galaxy and the tool dependencies are at the same location on the head and compute nodes.
• Job directories must be in a shared location on both head and compute nodes.
• This is mentioned in more detail in the tutorial.

Multiprocessing

Some tools can greatly improve performance by using multiple cores

Galaxy automatically sets $GALAXY_SLOTS to the CPU/core count you specify when submitting, for example, 4:

• Slurm: sbatch --ntasks=4
• SGE: qsub -pe threads 4
• Torque/PBS Pro: qsub -l nodes=1:ppn=4
• Condor: request_cpus: 4

Tool configs: Consume \${GALAXY_SLOTS:-4}

Speaker Notes

• For multiprocessing to be available both the tool and the Galaxy tool wrapper need to support it.
• You need to understand what tools are being run and set destinations for them with the appropriate specification.
• You’ll need to check for presence of GALAXY_SLOTS in the tool wrappers and tool macros to see if the tool supports multiple threads.

Memory requirements

For Slurm only, Galaxy will set $GALAXY_MEMORY_MB and $GALAXY_MEMORY_MB_PER_SLOT as integers.

Other DRMs: Please PR the appropriate code.

For Java tools, be sure to set -Xmx, e.g.:

<destination id="foo" ...>
    <env id="_JAVA_OPTIONS">-Xmx4096m</env>
</destination>

Speaker Notes

• Memory requirements can be set as well.
• For some tools, you’ll need to additionally provide environment variables to specify memory limits.
• This is different per DRM.

Run jobs as the “real” user

If your Galaxy users == System users:

• Submit jobs to cluster as the actual user
• Configurable callout scripts before/after job to change ownership
• Probably requires limited sudo for Galaxy user

See: Cluster documentation

Speaker Notes

• If you galaxy users map to the system users you can have Galaxy run the jobs with the account of those users.
• This allows proper resource accounting, but comes at some additional configuration complexities.

Job Config - Mapping Tools to Destinations

Problem: Tool A uses single core, Tool B uses multiple

• Both submit to the same cluster
• Need different submit parameters (--ntasks=1 vs. --ntasks=4 in Slurm)

Speaker Notes

• Mapping tools to destinations is the heart of the job configuration.
• This permits you to define which tools go to which destinations, and what resources they need.

Job Config - Mapping Tools to Destinations

Solution:

    <destinations default="single">
        <destination id="single" runner="slurm" />
        <destination id="multi" runner="slurm">
            <param id="nativeSpecification">--ntasks=4</param>
        </destination>
    </destinations>
    <tools>
        <tool id="hisat2" destination="multi"/>
    </tools>

Speaker Notes

• Here is an example mapping the hisat2 tool to a definition named multi.
• The multi destination specifies that 4 cores should be allocated for each job, and uses the slurm plugin.

The Dynamic Job Runner

For when basic tool-to-destination mapping isn’t enough

Speaker Notes

• However this static mapping sometimes isn’t sufficient.
• Here a dynamic mapping can be used instead.
• Galaxy has several different methods for accomplishing this.

The Dynamic Job Runner

A special built-in job runner plugin

Map jobs to destinations on more than just tool IDs

.left[Two types:]

• Dynamic Tool Destinations
• Python function

See: Dynamic Destination Mapping

Speaker Notes

• There are two built in ways to do this: dynamic tool destinations, and custom Python functions.
• We will cover both of these in the tutorial.

Dynamic Tool Destinations

.left[Configurable mappings without programming:]

• YAML format config file tool_destinations.yml
• Map based on tool ID plus:
  • Input dataset size(s)
  • Input dataset number of records
  • User
• Maps to static destinations defined in job config

Speaker Notes

• The Dynamic Tool Destinations are written as a yaml file.
• You can easily write rules based on file input sizes or number of inputs or user information.
• This can be used to determine memory and cpu allocations.

Arbitrary Python Functions

.left[Programmable mappings:]

• Written as Python function in lib/galaxy/jobs/rules/
• Map based on:
  • Tool ID
  • User email or username
  • Inputs
  • Tool Parameters
  • Defined “helper” functions based on DB contents
  • Anything else discoverable
    • Cluster queue depth?
    • …?
• Can dynamically modify destinations in job config (i.e. sbatch params)

Speaker Notes

• If Dynamic Tool Destinations are insufficiently flexible, then custom Python functions can be written.
• These can use any arbitrary information you want.
• They have full access to submitter information, job parameters, and any other resource you might want.
• They can dynamically modify destination parameters during runtime.
• If you need flexibility, these are what you want.

Key Points

• Galaxy supports a variety of different DRMs.
• Dynamic Tool Destinations are a convenient way to map
• Job resource parameters can allow you to give your users control over job resource requirements, if they are knowledgeable about the tools and compute resources available to them.

Thank you!