Quoting Jeff Squyres <jsquyres_at_[hidden]>:
> On Thu, 17 Jul 2003, Matthew Bettencourt wrote:
>
> > lam-7.0 issues
> > WHen I unpack the latest version and run configure I get
> >
> > checking size of FORTRAN INTEGER... unknown
> > configure: WARNING: *** Problem running configure test!
> > [snipped]
> >
> > I looked back about 100 lines and saw nothing I understood? has anyone
> > built this for aix 5.1??
>
> Yes. I can build properly on AIX 5.1 with the native AIX compilers
> (xlc/xlc_r, xlC/xlC_r, xlf/xlf_r). I have not tested the GNU compilers on
> AIX 5.1 because our AIX 5.1 machine doesn't have them installed. :-)
>
> Generally, this error indicates that configure was not able to compile/run
> a simple fortran program. Can you verify that your fortran compiler
> works?
Yes, f77 code works fine with this. Even the lam-6.x compiles just fine as
well. I am not sure how much the configur escript has changed since then, I did
notice more options.
>
> > Performance issue. On these SP systems I am noticing a latency of a
> > MPI_Send/Recv of about 1.2 to 1.5ms. Is this typical. With the native
> > MPI I am getting someing like .3ms or a factor of 4 better. I am trying
> > using the IBM's Colony switch for communication (binding to that IP vs
> > the ethernet IP on lamboot)
>
> Yes, I would expect this kind of behavior from LAM on IBM hardware. Keep
> in mind that IBM's MPI is implemented on LAPI (and probably shared memory
> -- I'm not too familiar with IBM's MPI), and LAM will either use shared
> memory or TCP. LAM's shared memory RPI *may* get close results to IBM's
> MPI, but I suspect that IBM has tweaked the heck out of their MPI to get
> absolutely fantastic performance on their hardware. That, and the fact
> that native LAPI performance is significantly better than native TCP
> performance pretty much guarantees that IBM's MPI will outperform LAM on
> their hardware.
This is run on three nodes so it is going across the switch for that. I do know
that the IBM does have a highly tuned MPI, but it lacks one key feature (well,
the POE actually) the ability to run multiple MPI jobs within a single batch
script at the sametime.
SO, I don't know what hope I have with this, but thanks for your input. I wish
that IBM would make their product more general in terms of things like
hardwiring ports into there POE so you can only run one instance on a node.
Allowing for better subscheduling in their scheduler like compaq's prun has or,
actually my personal favorite is lamexec/mpirun TBH which is why I am trying to
fix this right now.
If you have any suggestions of things that I may wish to set that would be great
to get some more performance out of this thing
Thanks
Matt
>
> It is not outside the realm of possibility that LAM will have a LAPI RPI
> to get performance in the same order of magnitude as IBM's MPI, but we
> don't have the resources to do that at the present time.
>
> Anyone want to volunteer? :-)
>
> --
> {+} Jeff Squyres
> {+} jsquyres_at_[hidden]
> {+} http://www.lam-mpi.org/
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
>
-------------------------------------------------
This mail sent through IMP: http://horde.org/imp/
|
