Re: MPICH-GM and the Intel 7.0 F95 compiler

Hi,

Just to let you know that the combination of the new 1.2.5 mpich-gm and Intel
Fortran 7.0 does not work with the FLASH code from Univ. Chicago, on Linux
RedHat 7.3 with updates, except kernel 2.4.18-5, on a bi-Xeon smp cluster.

Everything compiles fine, but the runs rapidly produce garbage numbers,
even when the compiler optimizations are turned off.

Sorry not to be able to be more specific, FLASH is a big code and for the moment
I don't see how to begin tracking the problem.

	Daniel Pfenniger


Susan Blackford wrote:
> Dear Myrinet users,
> 
> Several customers have asked for details of why the Intel 7.0 F90 compiler
> was not supported in MPICH-GM 1.2.4..8a and earlier.
> 
> Here are the specifics.
> 
> For various reasons, in MPICH-GM we must catch when memory pages are
> given back to the kernel, via munmap() or sbrk(). We accomplished this by
> providing our own malloc (actually the same code from the LibC but with
> some stubs). It works fine and it's portable as long as
> applications/libs/compilers use malloc/free.
> 
> The Intel 7.0 compiler implements the F90 statements "allocate" and
> "deallocate" using mmap/munmap/sbrk directly instead of malloc/free. In
> this case, registered pages may not be deregistered when given back to
> the kernel and it may lead to (silent) memory mapping corruption.
> 
> Intel informed us that only mmap/munmap were used to manage memory
> allocation, so catching the munmap system call on Linux is a safe
> work-around. Intel 8.X will come back to a malloc/free implementation.
> 
> The latest release of MPICH-GM, MPICH-GM 1.2.5..9, supports the 
> Intel 7.0 compilers.
>   http://www.myri.com/scs/
> 
> Feel free to contact us at [email protected] if you have further questions.
> 
> Susan