Re: High Performance Linpack
On Wed, Mar 21, 2007 at 01:40:57PM +0000, Jo Shields wrote:
> On Wed, 2007-03-21 at 13:59 +0100, Steffen Grunewald wrote:
> > On Wed, Mar 21, 2007 at 12:59:36PM +0100, Steffen Grunewald wrote:
> > > Hi,
> > >
> > > I'm looking for a recent version of HighPerformanceLinpack (you know, the
> > > one that is used to build the top-500 list) and instructions how to
> > > compile and run it under Debian/Etch on amd64 architecture.
> > > I have (C)BLAS and LAM-MPI already installed, but keep getting compile
> > > errors (from deep inside) :-(
> >
> > Replying to myself:
> >
> > I had to install libmpich1.0-dev (and BLAS/ATLAS dev packages), and heavily
> > twaek the Make.arch file, to get a xhpl binary. But I can only run it on
> > MPICH, not LAM-MPI. Suggestions?
>
> Is there a specific reason for using LAM?
>
> Performance on a cluster using Gigabit Ethernet and GNU compilers will
> be diabolical - and I don't think any of the high performance
> interconnects use LAM as a base
>
you should replace blas/atlas with libgoto (hand optimized assembly blas
routines) performance with gnu compilers is the same as commercial
compilers with the libgoto library for blas operations (for HPL).
(also mvapich which is an mpi implemenation over infiniband uses mpich
as a base, im also not sure if any high performance interconnections use
LAM as a base or not)
Jimmy.
--
Jimmy Tang
Trinity Centre for High Performance Computing,
Lloyd Building, Trinity College Dublin, Dublin 2, Ireland.
http://www.tchpc.tcd.ie/ | http://www.tchpc.tcd.ie/~jtang
Reply to: