[FLASH-USERS] Notes to All about FLASH on the IBM BlueGene

[Aaron Jackson]

Hi FLASH users,

 

I'm currently running FLASH at the New York Blue BlueGene/L and P machines.
I ran into a couple of issues that users of the BlueGene machines should be
aware of.  The first issue is more of a question.

 

Being relatively new to FLASH, I was using Makefile.h files from other sites
directories for BlueGene machines to base my own Makefile off of.  In
addition, I was also reading the man pages for IBM's XL compilers. I found
an inconsistency with one of the compiler options.  In most of the sites
directories using BlueGene machines, the compiler option qcache is set as
-qcache=auto. However, reading through the man pages of blrts_xlf90, I found
that the -qcache option's default changes depending on what is set by the
qarch option.  Furthermore, according to the man pages, -qcache=auto
optimizes the cache for the compiling machine, i.e. the login-node, not the
compute-nodes.  Is there any reason not to trust the man pages?

 

Also, while I'm discussing XL compilers, what other compiler options have
other "FLASH on BlueGene" users found useful / successful at optimization?
I'm currently only able to utilize -O2 optimization.

 

The second issue, which is more of a notice/warning, is that I've been
trying  to track down a bug in my code used within FLASH.  I was putting the
result of a linear interpolation into a power, i.e. 10.0**<result>.  I was
getting inconsistent results on BG/L even with the same executable.  It
turns out that making 10.0 into 10.0e0 fixes this.  I subsequently found a
compiler option which promotes all real constants to double precision for
the XL compilers, -qdpc=e.  I suggest using this along with setting the real
default to be 8 bytes.

 

Aaron Jackson

Department of Physics & Astronomy

Stony Brook University

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://flash.rochester.edu/pipermail/flash-users/attachments/20090618/35eaf970/attachment.htm>