next up previous index
Next: 6.2.1 Data Sharing Across Up: 6 Run Control Previous: 6.1.1 Multiple Projects

6.2 Experimental Splits

Process flow split pointsgif       define positions in the flow where variations in parameter values or process steps occur. Recomputing new experiments only from split points on is a powerful strategy to minimize the total computational effort and to yield results more quickly. Split points are generated by comparing runs to be submitted against the split tree of all existing runs of a given process and determining the most similar one that contains valid output data. Comparison is based on the complete tool parameter set actually used to invoke the tool rather than a potentially incomplete - with respect to the computation of output data - specification of process steps. Figure 6.2 shows three runs with parameter modification at the second and the third step. As the specifications - and thus the results - of the first steps of all runs and of the second steps of runs two and three are identical, they need not be computed. Split points are defined to reflect the flow of data taking place.

 figure1632
Figure 6.2:   Reduction of total computation by splitting. Splitting avoids recomputation of results that remain unchanged across different runs.

When comparing step parameter sets, the numerical accuracy of the comparison has to be limited to an appropriate relative difference, i.e., two floating point numbers are treated as equal when their relative difference is smaller that a given number. In VISTA/SFC, this epsilon  is adjustable by the user (see Figure 6.1), a value of tex2html_wrap_inline8211 has proven to be well-suited for most applications.





Christoph Pichler
Thu Mar 13 14:30:47 MET 1997