[Beowulf] mpich vs hp mpi performance

[Joshua Baker-LePain]

On Fri, 26 Sep 2008 at 6:28pm, Linus Harling wrote

> Thank you for the input! The system is a Dell T605 2x4core Opteron 2354
> with 32GB RAM. With a second one like it being added (probably using
> ethernet) if need arises. Which means most/all of the communication will
> be by shared memory to begin with.
>
> The app is LS-DYNA, a FEM-suite from LSTC: http://www.ls-dyna.com/
>
> I'm just a sysadmin tasked with installing the machine, and have limited
> knowledge in math, but my guess is that the code is quite
> communication-heavy considering the specs of the machine and an
> assumption that FEM is essentially SIMD. Am I correct in my assumptions
> and does anyone have any experience of different mpi-implementations on
> such codes?

DYNA has 2 main solver types -- explicit (iteration based) and implicit 
(matrix inversion, essentially).  If you're only using one system and the 
explicit solver, then there's no need for any MPI, as ls-dyna is 
multi-threaded.  Last I knew, however, the implicit solver is only 
parallelized in mpp-dyna.

I've found the explicit solver to not be overly communications heavy. 
I.e., I've seen decent scaling using plain-jane GigE and very modest node 
counts.  Unfortunately, I haven't done any benchmarks of the various MPI 
versions available.  HTH.

-- 
Joshua Baker-LePain
QB3 Shared Cluster Sysadmin
UCSF