Re: [myrinet] RE: Re: Re: mpich PBS etc. and gm_ports not closing
Douglas Johnson <[email protected]>
| Newsgroups | gmane.network.myrinet.general |
|---|---|
| Message-ID | <[email protected]> |
Yes, there is signal handling in the mpirun script. But it's not completely reliable. SIGINT is transmitted through the rsh to the remote processes, no other signals are transmitted. Look at the BUGS section of the rsh man page. There needs to be more intelligent signal handling at the MPICH layer, there are many reasons why not rely on the shell script and rsh. The original problem was C-c not killing all the MPI processes the way a user would expect, but like I said, I wouldn't completely trust the perl script to handle all things that users could come up with. Doug For example, for whatever reason a node that is running an mpi process is shutdown. When linux is shutting down it will first try to send a TERM signal, then sleeps and then a kill signal to clean up any processes that are left. The MPI process on that node needs to catch the TERM signal and call MPI_Abort. On Wed, 20 Sep 2000, Greg Lindahl wrote: > > Here's a patch that catches the C-c and calls MPID_Abort. > > I don't understand. Normally mpirun (perl script) receives the ^C from a > user. Does the standard one not have a signal handler? > > -- g > -- Doug Johnson Systems Developer/Engineer The Ohio Supercomputer Center [email protected]