wiki:DesktopGrid

Version 16 (modified by -TJ, 14 years ago) (diff)

--

Grid computing with BOINC

Grid versus volunteer computing

Grid computing is a form of distributed computing in which an organization (business, university, etc.) uses its existing computers (desktop and/or cluster nodes) to handle its own long-running computational tasks. This differs from volunteer computing in several ways:

  • The computing resources can be trusted; i.e. one can assume that the PCs don't return results that are intentionally wrong, and that they don't falsify credit. Hence there is typically no need for replication.
  • There is no need for screensaver graphics; in fact it may be desirable to have the computation be completely invisible and out of the control of the PC user.
  • Client deployment is typically automated.

Using BOINC as a grid platform

Although it was originally designed for volunteer computing, BOINC works very well for grid computing. The steps in creating a BOINC-based grid are:

To ensure that outside hosts can't participate in your project or access its files, configure your firewall to prevent HTTP access to your BOINC server.

Some resources

  • The SZTAKI desktop grid project has developed software allowing hierarchical organizations to share resources in a way that reflects the hierarchy.

Integrating BOINC with other grid platforms

  • Run BOINC using Condor's backfill mechanism
  • Researchers at CERN have set up a system where submitted jobs are sent either to a BOINC project or to a GRAM job manager. They developed two utilities, kill_wu and poll_wu, to support this. They are in the boinc/tools directory. Contact Christian Søttrup (chrulle at fatbat.dk) for more info.
  • The Lattice project from the University of Maryland has developed a Grid system that integrates Globus, BOINC, and several other software components.
  • The SuperLink project from Technion University developed a system that dynamically assigns jobs to a central server, a local cluster, the EGEE grid, or a BOINC project, based on their estimated runtime.
  • The MindModeling@Home project has incorporated two United States Air Force Department of Defence LSF clusters and an additional Beowulf cluster in their beta project. Tools to facilitate this included: LSF submission scripts, control scripts to dynamically clean up potentially reschedule clients upon node time out and a multiple BOINC instance installer.