Scalability of CHARMM on various architectures

[Alan Grossfield]

:We are designing a cluster for which the most important code will be the
:computational chemistry program, CHARMM.  In our preliminary tests on
:an existing cluster, we have confirmed our expectation that interthread 
:communications will be our first bottleneck.  A test run on a typical 
:problem did not scale beyond 4 nodes on a cluster composed of P2-450
:processors and fast ethernet interconnections.  In contrast, the same
:code scales almost perfectly up to at least 12 processors on an SGI-Unix 
:SMP machine.  CHARMM uses PVM.

    Actually, all modern versions of charmm use mpi.  The scaling behavior
    is more or less well-known, though, and quite frustrating.  You can do
    a bit better if you use Josip Loncharic's TCP patches (8 processors is
    maybe 6x faster than 1 processor for MD using PME with ~10K atoms), but
    that's about it.

: * If we go with pentium II/III processors, how far is Myrinet likely to
:   permit us to scale these calculations?

    It should be better -- I know Bernie Brooks' lab is using gigabit
    ethernet to connect LoBoS (they're one of the primary sites for CHARMM
    development), but I haven't checked their benchmarks recently.

:
: * With Myrinet, would alpha nodes tend to scale any better than pentium nodes
-:?
:

    Probably, because of the better bus speeds, but I haven't seen data on
    this (maybe the paralogic guys can comment, if they've done benchmarks).


Alan Grossfield
-------------------------------------
| New email:  alan at dasher.wustl.edu |
| Update accordingly                |
-------------------------------------