Besides the goals of classical intertool PIF implementations featuring object-orientedness (PIF/GESTALT, [Boni91]) or suitability for TCAD environments (BPIF, [Wong91]) the VISTA implementation stresses efficiency in terms of run-time performance and database compactness.
For comparison purposes a test for writing and reading
10000 three-dimensional points of a PIF pointList
was performed on a DECstation 3100 and compared to the two intertool PIF
implementations above. The test yields 0.51 s write and 0.66 s read time, and
the database created is 36 kB in size using the compressed array storage
implemented in the PAI
.
The object-oriented implementation [Boni91] needs 1066 s write and 171 s read time for this test, and creates a database of 854 kB size (this test was performed on a Sun-3/260). The TCAD-oriented PIF implementation [Wong91] needs 4 s to write and 8 s to read the same number of points, and creates a database of 200 kB size (this test was performed on a DECstation 3100). It is questionable if the database size of the last test is indeed comparable, since three double precision values would consume 24 bytes per point, but the last implementation only uses 20 bytes per point.
As can be seen from this evaluation, linking a TCAD application to the VISTA environment is not a performance issue. In contrast to a client-server approach, the administrative and communication overhead is negligible for any application consuming a few seconds of CPU time - the commonly used argument, that PIF is not practical because of its low run-time performance does not hold true.
Considering realistic three-dimensional grids, these contain definitely more
than 10000 points, thus the same test was performed with
1000000 points on a DEC Alpha 3000/600 with a
64-bit architecture. This machine needs 7.9 s write time and 5.2 s read time,
which is even more negligible compared to typical execution times of
three-dimensional simulators. A database of only 2 MB in size is created due
to the effective compression algorithm implemented in the PAI's compression
layer.
Besides these raw point read and write times, many other aspects of PAI performance have to be considered. However, since there are no other performance data available of the other intertool PIF implementations, and since these aspects relate to practical usage of the PAI in large TCAD engineering problems, a PIF example as a result from a complete SFC process flow simulation is evaluated in Section 8.7.