[mvapich-discuss] librdmacm warning when I run MPI programs

[Jonathan Perkins]

On Wed, Sep 26, 2012 at 02:11:10PM +0530, Uday K Bondhugula wrote:
> On 09/24/2012 08:05 PM, Jonathan Perkins wrote:
> >On Mon, Sep 24, 2012 at 04:16:27PM +0530, Vaivaswatha N wrote:
> >>Hi,
> >>
> >>I'm trying to use MVAPICH to run an MPI program in my college cluster. My
> >>program is working correctly, but the following warnings are emitted.
> >>
> >>librdmacm: couldn't read ABI version.
> >>librdmacm: assuming: 4
> >>
> >>Searching on Google seems to hint that this is related to the InifniBand
> >>not being used for the communication. Can someone comment if this is true
> >>and what could be causing the issue?
> >
> >This warning usually indicates a problem with the OFED installation
> >which could prevent the RDMA connection manager from functioning
> >properly.  You may want to bring this up to your Admin.
> 
> The problem didn't exist earlier, but showed up after the kernel was
> updated from 2.6.18-128.el5 to 2.6.18-308.13.1.el5 (latest one in
> CentOS 5.3 repo now). The packages librdmacm and ofed-scripts
> themselves are up-to-date. No other configuration was changed as far
> as I can remember. This is with
> 
> mvapich2-1.8a1
> librdmacm-1.0.10-1.el5
> ofed-scripts-1.5-0

Thanks for the information.  It looks like the kernel upgrade may
require a reinstall of OFED (or yum reinstall of the packages).
Otherwise you may need to downgrade to the previous kernel.

You mentioned that you're using the alpha release of 1.8.  May I suggest
that you take this opportunity to upgrade to the latest tarball from the
1.8 branch.  We have made the general release as well as some additional
bug fixes since 1.8a1 was released.

You may also want to take a look at our 1.9a release which contains all
the bug fixes available in the 1.8 branch as well as some additional
performance enhancements and other features.

-- 
Jonathan Perkins
http://www.cse.ohio-state.edu/~perkinjo