|
IBM that simply require access to grid-managed data, or
they might be complex parallel applications that use myriad heterogeneous
computational or data resources.
Unlike first-generation grid codes that were large and
monolithic, these component modules are small and dynamic. And unlike early
codes, which tended to be "pleasantly parallel," these modern codes are
heterogeneously parallel, often requiring more than one precursor component to
be completed before computation can begin. This parallel heterogeneity, mixed
with resource heterogeneity, requires sophisticated workflow planning and
execution tools to abstractly plan and execute the workflow. As these
requirements are shared across many scientific communities, several tools exist.
The Planning for Execution in Grids (Pegasus) environment is
the workflow planning tool primarily used within the Telescience Project.
Pegasus is a framework that maps complex scientific workflows onto distributed
resources, such as the grid. Pegasus maps an abstract workflow description to
its executable form, and Condor Directed Acyclic Graph Manager (DAGMan) executes
the jobs specified in the executable workflow. Pegasus and DAGMan map and
execute workflows on Condor pools, clusters managed by LSF or PBS, TeraGrid
hosts, and individual hosts.
Pegasus operates on abstract workflow descriptions where the
analysis is described in terms of application components and the data that the
components use. The workflow is abstract because it does not identify the
resources necessary for execution. Pegasus takes this abstract workflow
description and produces an executable workflow that identifies the compute
resources needed and includes data management nodes that stage the data in and
out of the computations. Additional workflow nodes are added to register the
newly derived data products so they can be located at a later time.
|
