We use the full PGI-CDK to fuel our clusters with compilers and the mpich library.
We have to support SuSE 8.1 Xeon, SLES8 and SuSE 9.0 Opteron systems.
PGI-CDK 5.2 supports all these platforms smoothly although it was installed
remotely for the SuSE 8.1 Xeon cluster and from one node of the SLES8 Opteron
cluster for all Opteron systems.
The same thing for PGI-CDK 6.0 has turned into a nightmare. None of the mpich
bindings works, partly due to very tight bindings of certain routines to a specific
glibc version, partly due to unresolved symbols in the mpich library for C that
are not located in any glibc but in the f2c library. Another problem is the remote
execution method which needs to be determined at installlation rather than at
I had hoped that with version 5.2 the stand-alone behavior of the installation process
had come to an end, but the behavior of 6.0 is the worst ever. I would like to urge
the developers to avoid too specific bindings of symbols to glibc subversions and
mixing symbols from the f2c library into the pure C mpich library. Altogether a
substantially improved support for the installation of the PGI-CDK package on a
network drive which has to support more than one platform would be highly desirable.
For the time beeing we had to return to version 5.2.