[Stellar-discuss] n processors
Bill Paxton
paxton at kitp.ucsb.edu
Sun Dec 4 17:02:35 GMT 2005
On Dec 3, 2005, at 7:14 PM, Frank Timmes wrote:
> one item that hasn't been discussed is if this
> new generation code should be serial or parallel.
> if parallel, do we aim for shared memory or distributed memory
> architectures (e.g., openmp versus mpi or both)?
Hi Frank,
Good point -- and another area where I'm way out of touch with
current state of the art tools. Any suggestions for background
reading for me? And just to add to the pot, what about the issues of
dual core processors that share a single cache memory? Is this
included in the usual shared memory approaches or does it require
something special (at the thread level I presume).
Perhaps a reasonable guideline for decisions such as this is what to
we need in order to make good use of a typical laptop computer --
both now and in the next few years. In other words, we're not
designing for 3D simulations on a super computer with massive
parallelism and huge data sharing. But even our target laptop will
have soon multiple processors, at least at the "multi core" level, so
some form of parallelism is certainly attractive. And the stellar
evolution problem certainly presents lots of opportunities with
things like EOS and burning being done in parallel on different shells.
I guess I'm talking myself into a shared memory, multi-threaded
design. Are there good tools to support that in Fortran90? And can
we do it while still honoring the "trumps-all" principle of KISS?
--Bill
More information about the stellar-discuss
mailing list