DRP: Dynamic Resource Provisioning
Ioan Raicu, Ian T Foster, Carl Kesselman, Kate Keahey, Keith M. Jackson, David E. Konerding, Tim Freeman, Ravi Madduri · 2006
by the Scalable Services breakout session where the following people were present: Ian Foster, Carl Kesselman, Kate Keahey, Keith Jackson, David Konerding, Tim Freeman, Ravi Madduri, and Ioan Raicu (if I missed anyone, please feel free to add them). After discussing three different systems (AstroPortal [1], GridFTP [2], and Condor Brick [3]) that all needed to do dynamic resource provisioning, we came to the conclusion that there are several things in common across all these three systems, and that other feature are desirable: • Typical system components and other related components: o User(s) to produce work o Work Dispatcher o Resource Monitor (observes state information at the work dispatcher) o Resource Manager (resource provisioner based on information from Resource Monitor) o Worker Starter (used to bootstrap worker code by the Resource Manager) • Coarse grained resource provisioning (i.e. GRAM, etc); typical resources allocated for hours to days • Fine grained work dispatching (i.e. WS calls, Condor glidein, proprietary protocol based on UDP/TCP/IP, etc); typical work completion in seconds to minutes