[GE users] Question about MPICH and SGE under Rocks

Reuti reuti at staff.uni-marburg.de
Tue Mar 8 01:23:57 GMT 2005


    [ The following text is in the "ISO-8859-1" character set. ]
    [ Your display is set for the "ISO-8859-10" character set.  ]
    [ Some special characters may be displayed incorrectly. ]

Quoting Ron Chen <ron_chen_123 at yahoo.com>:

> --- Rayson Ho <raysonho at eseenet.com> wrote:
> > And if you don't want to recompile MPICH, then you
> > can see whether you like
> > Reuti's "proof of concept" or not...
> > 
> > http://gridengine.sunsource.net/servlets/ReadMsg?
> > msgId=24717&listName=users
> > 
> > (He sent it out almost a month ago, and that thread
> > was started by you...
> > we can't offer any help if we don't get any
> > response.)
> 
> To be fair, the "proof of copcept" is not ready for
> production use:
> 
> 1) if more than 1 job runs on the node, then the log
> file would be overriden. So we need to add the job ID
> as part of the name of the logfile.
> 
> 2) I am not sure if it is fine or not to dlopen the
> same shared library many times. May be it's better to
> do the init part once.

If there is really a demand for such a lib, I could follow the thing on to get 
a complete lib and a Howto page to use it with SGE.
 
> Also, if we can get Reuti's code into MPICH, then we
> don't need to do the LD_PRELOAD thing. And it would
> also help Torque/PBS/LSF/Slrum/Condor to keep track of
> shared mem segments and clean up as well.

The only problem with a modified MPICH would be, that it's only working if you 
have the source of the programs, as MPICH is generating .a files by default and 
so they are already linked into the application. Some licensed software we got 
as binary only have it built in this way (but they didn't used 
--with-comm=shared, so we are not facing the shared-memory problem) [although 
it's possible to compile MPICH to get .so files with --enable-sharedlib - 
rarely used IMO].

On the other hand: the LD_PRELOAD thing would be generic for MPICH, MPICH2,... 
as long as the apps are dynamically linked against libc.

Cheers - Reuti

---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
For additional commands, e-mail: users-help at gridengine.sunsource.net




More information about the gridengine-users mailing list