Workflow Concept of WS-PGRADE/gUSE - Science Gateways for Distributed Computing Infrastructures - page 36

Last but not least, the class of data-based routing collects several cases when the
existence or the value of the data affects the workflow s further interpretation such

as task pre- and post-condition considering the existence and value of the data, or
data-based routing. The task precondition

'

data value pattern (Fig. 3.1 ) describes
the case when the job can be run, or can be blocked depending on the value of the
incoming data. gUSE precisely covers this situation by introducing the concept of
the

—

“ port-dependent condition ” detailed in Sect. 3.6.4 .

3.4 Levels of Parallelism

Depending on the execution of the jobs and workflow settings, four levels of
parallelism can be identi ed in a WS-PGRADE/gUSE workflow. The lowest level,

or node-level parallelism denoted as circled in Fig. 3.2 , is where the appli-

cation itself is prepared to utilize the bene

“ J1 ”

ts of multicore processors or cluster
systems. In multicore environments these applications are usually designed as
multithread applications like GPU programs. In the case of cluster systems, the
applications use speci

cations
(such as OpenMPI). Besides this option, gUSE supports parallel execution of dif-
ferent jobs placed at different parallel branches of the workflow graph as the most
intuitive and simple concurrent execution. It is denoted by J1 and J2 circled
together in Fig. 3.2 and is called branch-level parallelism. A third level of paral-
lelism covers the situation when one algorithm should be executed on a large
parameter

c programming libraries that implement MPI speci

field, generally called parameter study or parameter sweep (PS) execu-
tion. This scenario is illustrated in the middle part of Fig. 3.2 and called PS par-
allelism, and the node that can expose such a feature is called PS node. Various

Fig. 3.2 Levels of parallelism