Running Jobs on Remote Resources with Pulsar

Contributors

Authors: Nate Coraor Simon Gladman Marius van den Beek Helena Rasche

Questions

• How does pulsar work?

• How can I deploy it?

Objectives

• Have an understanding of what Pulsar is and how it works

• Install and configure a Pulsar server on a remote linux machine

• Be able to get Galaxy to send jobs to a remote Pulsar server

Requirements

• slides Slides: Ansible
• tutorial Hands-on: Ansible
• slides Slides: Galaxy Installation with Ansible
• tutorial Hands-on: Galaxy Installation with Ansible
• A server/VM on which to deploy Pulsar

What are heterogenous compute resources?

Differences in:

• Operating system or version
• Users/groups
• Data accessibility
• Administrative control
• Physical Location (i.e. Cities)

Galaxy expects:

• One OS, version (dependencies)
• Shared filesystem w/ fixed paths

Example - Australia

Partial solution - CLI job runner

SSH to remote, submit jobs with CLI sbatch, qsub, etc.

Still depends on shared FS

Pulsar

Galaxy’s remote job management system

• Can run jobs on any(?) OS including Windows

• Multiple modes of operation for every environment

Pulsar - Architecture

• Pulsar server runs on remote resource (e.g. cluster head node)

• Galaxy Pulsar job runner is Pulsar client

• Communication is via HTTP or AMQP, language is JSON

• File transport is dependent on communication method

Pulsar - Architecture

Pulsar Transports - RESTful

Pulsar server listens over HTTP(S)

Pulsar client (Galaxy) initiates connections to Pulsar server

Good for:

• Environments where firewall, open ports are not concerns
• No external dependencies (AMQP server)

Pulsar Transports - AMQP

Pulsar server and client connect to AMQP server

Good for:

• Firewalled/NATted remote compute
• Networks w/ bad connectivity

Pulsar Transports - Embedded

Galaxy runs Pulsar server internally

Good for:

• Manipulating paths
• Copying input datasets from non-shared filesystem

Pulsar - Job file staging

Pulsar can be configured to push or pull when using RESTful:

• Push
  • Galaxy sends job inputs, metadata to Pulsar over HTTP
  • Upon completion signal from Pulsar, Galaxy pulls from Pulsar over HTTP
• Pull
  • Upon setup signal, Pulsar pulls job inputs, metadata from Galaxy over HTTP
  • Upon completion, Pulsar pushes to Galaxy over HTTP

Pulsar can use libcurl for more robust transfers with resume capability

AMQP is pull-only because Pulsar does not run HTTP server

Pulsar - Dependency management

Pulsar does not provide Tool Shed tool dependency management. But:

• It has a similar dependency resolver config to Galaxy
• It can auto-install conda dependencies
• It can use containers too!

Pulsar - Job management

Pulsar “managers” provide job running interfaces:

• queued_python: Run locally on the Pulsar server
• queued_drmaa: Run on a cluster with DRMAA
• queued_cli: Run on a cluster with local qsub, sbatch, etc.
• queued_condor: Run on HTCondor

Pulsar Australia

Resources

• Pulsar Read-the-docs
  • https://pulsar.readthedocs.io/en/latest/index.html
• Pulsar on galaxyproject.org
  • https://galaxyproject.org/admin/config/pulsar/
• Pulsar Github
  • https://github.com/galaxyproject/pulsar
• Pulsar Ansible
  • https://github.com/galaxyproject/ansible-pulsar

Key Points

• Pulsar allows you to easily add geographically distributed compute resources into your Galaxy instance
• It also works well in situations where the compute resources cannot share storage pools.

Thank you!