32 bit vs 64 bit computer ?
Many of your questions may have already been answered in earlier discussions or in the FAQ. The search results page will indicate current discussions as well as past list serves, articles, and papers.
James Cownie jcownie at etnus.comFri Nov 23 04:27:12 PST 2001
- Previous message: 32 bit vs 64 bit computer ?
- Next message: Fortran compilers for Linux/mpich
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
> Before MPI (unfortunately) came to dominate message-passing, PVM was > the standard library used. PVM is designed for heterogeneous > systems. For example I have a code that uses MPI internally on both a > Cray T3E and also a Fujitsu Vector Processor but which uses PVM to > communicate between the two big machines. Despite the implication above that MPI is inferior to PVM in its support of heterogeneous systems, the MPI standard _was_ designed for heterogeneous systems. A conforming MPI program provides enough information on both send and receive to allow the MPI implementation to translate data between machine formats (without requiring a function call per data element to achieve it as PVM used to do!). The issue which is likely preventing you from exploiting this is that of _starting_ MPI processes on these two different machines and exploiting the vendor optimised MPI on both of them. Since CRAY has no incentive to make their MPI handle a Fujitsu VPP, and Fujitsu has no incentive to make their MPI handle a Cray T3E interoperability of _vendor optimised_ MPIs is small. (Though, of course, your Quadrics' MPI will work in an optimised fashion with the Fujitsu VPP and T3E, I expect :-) However, if you're prepared to use a portable MPI such as MPICH, then you can easily handle heterogeneous machines inside a single program. (See the Globus/MPI work, for instance). I have also seen work which used the MPI profiling interface to wrap a vendor MPI so that it would inter-operate with a portable MPI. So, in summary 1) The MPI specification fully supports heterogeneity. 2) There are MPI implementations which support heterogeneity. 3) You're living in another universe if you think that vendors will spend any time making their MPI implementations inter-operate off-box with their competitors, rather than tweaking their on-box performance in the hope of wiping out said competitors ! -- Jim James Cownie <jcownie at etnus.com> Etnus, LLC. +44 117 9071438 http://www.etnus.com
- Previous message: 32 bit vs 64 bit computer ?
- Next message: Fortran compilers for Linux/mpich
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
More information about the Beowulf mailing list
