[Wien] Intel MKL 7.2

[Jeff Spirko]

Hello, Wien2k users,

MKL 7.2 has an interesting feature that it will automatically
multi-thread the matrix diagonalizations on multi-processor
machines.  You can see if this is happening by pressing capital 'H'
in "top" while lapw1 is running.  MKL multi-threading can be
controlled with the environment variables OMP_NUM_THREADS and/or
MKL_SERIAL.  (Search in the MKL documentation for these strings for
more info.)  Setting OMP_NUM_THREADS=2 on a single-processor
hyper-threaded machine is possible, but I don't know if it will
help.

A word of caution: I have a case for which lapw1 will sometimes fail
to find eigenvalues for some k-points if this multi-threading is
allowed to proceed.  The regular single-processor version can be
forced by placing the setting "MKL_SERIAL=yes" at the BEGINNING of
the command line (or set as an environment variable in general).  If
I do this (by typing "MKL_SERIAL=yes runsp_lapw") then lapw1 works
normally.

Is anyone successfully using MKL 7.2 in Linux?  Has anybody else
seen this?


By the way, if no eigenvalues are found for a K-point in my case,
the only error was a strange read failure in lapw2 (at line 1360 in
fermi.f, reading from unit 30).  This should be made less obscure.

Best Regards,
-- 
Jeff Spirko   spirko at lehigh.edu   spirko at yahoo.com   WD3V   |=>

The study of non-linear physics is like the study of non-elephant biology.

All theoretical chemistry is really physics;
and all theoretical chemists know it. -- Richard P. Feynman