[Bug 171993] Review Request: mpich2 - An implementation of MPI

[bugzilla at redhat.com]

Please do not reply directly to this email. All additional
comments should be made in the comments box of this bug report.

Summary: Review Request: mpich2 -  An implementation of MPI


https://bugzilla.redhat.com/show_bug.cgi?id=171993





------- Additional Comments From orion at cora.nwra.com  2008-04-16 17:03 EST -------
(In reply to comment #66)
> I should note that I've made significant changes to lam/openmpi that aren't
> reflected in Fedora yet.  The most important of which is dumping the
> alternatives usage entirely.  It has proven to be unwieldy at best, a nightmare
> at worst.  Instead, my current packages use mpi-selector.

Ah, great, yet another tool for manipulating environment variables.  And one
with a domain specific name.  I'm all for dropping alternatives, it's definitely
crap for situations like this.  But have you thought about using Modules
(environment-modules in Fedora)?  It's a general purpose environment
manipulation tool.

> Also of note is that, in line with the comments in comment #7 and comment #8, my
> current packages use %{_libdir}/%{name} as the prefix for all files in the lam
> package, and %{_libdir}/%{name}/%{version}-%{opt_cc} as the prefix for all files
> in openmpi (the difference being because I've never been given any requests to
> have more than one version or compiler of lam installed at a time, but I have
> been given requests to have multiple versions and multipler compilers of openmpi
> at the same time).

That's good.  However for such a scheme makes it into Fedora, I think it's going
to need to go through the packaging committee.  You run into all kinds of FHS
violations as enforced by rpmlint like:

openmpi.i386: E: file-in-usr-marked-as-conffile
/usr/lib/openmpi/1.2.5-gcc/etc/openmpi-default-hostfile
openmpi-libs.i386: E: postin-without-ldconfig
/usr/lib/openmpi/1.2.5-gcc/lib/libmpi_cxx.so.0.0.0
openmpi-libs.i386: E: library-without-ldconfig-postun
/usr/lib/openmpi/1.2.5-gcc/lib/libmpi_cxx.so.0.0.0

> Another item of note is that I have outstanding requests to also include
> mvapich/mvapich2 in the distro for use over infiniband.  
 
> If these packages are redundant, then would it not be better to use mvapich
> (which I think supports more of the high speed interconnects natively, but I
> could be wrong)?  

Well the FAQ says:

What is MVAPICH/MVAPICH2?

MVAPICH is a high performance implementation of MPI-1 over InfiniBand based on
MPICH1. MVAPICH2 is a high performance MPI-2 implementation based on MPICH2.
High performance and scalable support over the Verbs Level Interface (VAPI) are
provided by these packages. These packages also support other underlying
transport interfaces (such as the emerging uDAPL, OpenIB/Gen2 and the standard
TCP/IP) for portability across multiple networks. 

And the download pages says:

MVAPICH2 1.0.2p1 is available as a single integrated package (with MPICH2
1.0.5p4) for download. 

Now, whether and an installed MVAPICH2 provides a pure superset of MPICH2
capabilities remains to be seen.

-- 
Configure bugmail: https://bugzilla.redhat.com/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the QA contact for the bug, or are watching the QA contact.