Re: [ESPResSo-users] mpi and compressed block files

[Axel Arnold]

Hi Martin!

On 09/07/2012 01:43 PM, Martin Lindén wrote:
> Instead, I think the problem occurs when trying to close the channel
> associated with the gzipped file:
> > > error waiting for process to exit: child process lost (is SIGCHLD
> > > ignored or trapped?)
> > >       while executing
> > > "close $innnn"
> > >       (file "blockread3.tcl" line 14)
Ok, that error is something else, caused by OpenMPI that abuses signals 
to communicate between the nodes. With up to three nodes, everything is 
ok, as it uses SIGUSR1 and SIGUSR2, but for the 4th node, it uses 
SIGPIPE/SIGCHLD. This makes it impossible for the program to decide 
whether a spawned child process has terminated, and therefore you get 
the error above. To avoid the error, you can put a catch around the 
close statement. You will just not be able to figure out whether gzip 
had problems, but otherwise, this works.

Now, as a warning, there are also MPI implementations, e.g. on BlueGene, 
which just return close any tasks immediately. On platforms like this, 
reading gzipped files through the Tcl popen is currently impossible. 
From tcl8.6b1 on, there seems to be a native zlib command 
(http://wiki.tcl.tk/4610), but that is fairly new stuff and you would 
have to compile your own Tcl.

Cheers,
Axel

--
JP Dr. Axel Arnold           Tel: +49 711 685 67609
ICP, Universität Stuttgart   Email: address@hidden foun
Pfaffenwaldring 27
70569 Stuttgart, Germany