[MPICH] Running on root's MPD as either root or another user

[Ralph Butler]

The key is to make sure the mpd files are installed as setuid-root as  
described in the manual.
Here's the relevant blurb:

5.1.6 Running MPD as Root
MPD can run as root to support multiple users simultaneously. To do  
this,
it easiest to simply do the “make install” in the mpd sub-directory  
as root.
This will cause the mpdroot program to be installed in the bin  
directory with
setuid-root permissions.

If things are already installed, you can alter mpdroot's permissions  
with the chmod
command, e.g.:
     chmod +s mpdroot

--ralph

On MonOct 1, at Mon Oct 1 3:03PM, Matthew Chambers wrote:

> Hi all,
>
> The install guide seems unclear about how to do this or if it's  
> even possible.  I have my user's .mpd.conf as:
> MPD_SECRETWORD=foo
> MPD_USE_ROOT_MPD=1
>
> And my /etc/mpd.conf is:
> MPD_SECRETWORD=foo
>
> Permissions on the user's conf are: -rwx------
>
> When permissions on the /etc/mpd.conf are: -rwx------
> I can run mpdtrace and MPI jobs as root just fine, but when I try  
> either of those as the user, I get "open failed for root's mpd conf  
> filempdtrace: mpd_uncaught_except_tb handling"
>
> So when I set permissions on the /etc/mpd.conf are: -rwx---r--
> I can run mpdtrace as the user, but when I try to run an MPI job  
> (e.g. mpiexec -n 16 hostname ; not an MPI program, but still works  
> fine with MPICH2), it hangs.  When I break it with CTRL-C I get the  
> message "(mpiexec 413): mpiexec: failed to obtain sock from  
> manager".  I can't run mpdtrace or MPI jobs as root because of:  
> "configuration file /etc/mpd.conf is accessible by others \ change  
> permissions to allow read and write access only by you".  This  
> seems a bit of a catch-22.
>
> I'm mostly interested in getting it working for the user.  I don't  
> have the MPD binaries installed in the standard location, so I'm  
> not sure how to set the permissions.  I tried setting everything  
> to: -rwxr-sr-x
> But that has the same error.  The user only has an account on the  
> head node, is that the problem?
>
> Thanks,
> Matt Chambers
>