I'm using the methods in the paper https://www.mcs.anl.gov/~lusk/papers/fault-tolerance.pdf to make my Monte Carlo jobs fault tolerant (using inter-communicators rather than intra-communicators). The processes have a manager/worker arrangement, with the workers being single instances of a simulation, each with unique randomized inputs. These workers are created using MPI_Comm_spawn(). When a worker finishes, it must exit, and up to now I've been calling MPI_Finalize in the worker before exiting.

From what you've said, this is incorrect because MPI_Finalize is collective over all of the processes in the job? Obviously my managers do not exit when the workers exit. Can you suggest any way to do cleanup before a worker exists, so that the job can continue?

Thanks, Kurt -----Original Message----- From: Joachim Protze <[email protected]> Sent: Tuesday, April 7, 2020 3:37 AM To: [email protected]; Mccall, Kurt E. (MSFC-EV41) <[email protected]> Cc: Zhou, Hui <[email protected]> Subject: [EXTERNAL] Re: [mpich-discuss] Abnormal termination on Linux As long as you use a standard MPI implementation, there is no resilience and the MPI job will not survive a segfault or any other MPI error. Probably you should target your question towards "Why does my application segfault in the first place?" If it's only one process, that segfaults probably use the core file for debugging. By default I assume corefile size is limited on most clusters, so increase it in your batch job: ulimit -c unlimited gdb a.out -c corefile.hostname.timestamp backtrace quit Am 07.04.20 um 00:23 schrieb Zhou, Hui via discuss:

I see. MPICH doesn’t trap any signal as I checked 😊. So back to your issue, let’s assume your trap did work, but after the signal handler, which after it called `MPI_Finalize`, the code is still going to run the same code that causes segfault in the first place, right? Wouldn’t that endup in infinite signal loop or maybe the kernel smart out and by-passed your handler? Just a thought experiment.

Calling MPI_Finalize would not help, because MPI_Finalize is a collective call and will not finish before all other processes have called MPI_Finalize. Furthermore, in such error situation you won't have any guarantees, that the process does not have any open blocking or non-blocking communication. Such communication would need to be completed before calling MPI_Finalize. - Joachim

Beside, MPICH is thread-safe but not interrupt safe. So you should not really call MPI functions inside signal handlers.

-- Hui Zhou

From: "Mccall, Kurt E. (MSFC-EV41)" <[email protected]> Date: Monday, April 6, 2020 at 4:50 PM To: "Zhou, Hui" <[email protected]>, "[email protected]" <[email protected]> Subject: RE: [mpich-discuss] Abnormal termination on Linux

Hui,

I’d like to trap segfaults so that the process that raised them can be shut down gracefully/finalized without taking down the whole MPI job. Maybe you are right and I shouldn’t trap them.

However, when I install my signal handlers, I check if MPI already has a signal handler installed for each of them. It isn’t reporting that MPI has done so. Perhaps my code is incorrect?

setUpOneHandler(const int signum) { struct sigaction sa_new, sa_old;

sa_new.sa_handler = mpiSignalHandler; sa_new.sa_flags = 0; sigemptyset(&sa_new.sa_mask);

if (sigaction(signum, NULL, &sa_old) < 0) { // ERROR: could not query old handler

} else if (sa_old.sa_handler == SIG_IGN || sa_old.sa_handler == SIG_DFL) { // MPI hasn't set its own handler for this signal, so we // will install our own.

if (sigaction(signum, &sa_new, NULL) < 0) { // ERROR: could not set new handler } } else { // MPI already has a handler installed for this signal. Do nothing } }

From: Zhou, Hui <[email protected]> Sent: Monday, April 6, 2020 2:08 PM To: Mccall, Kurt E. (MSFC-EV41) <[email protected]>; [email protected] Subject: [EXTERNAL] Re: [mpich-discuss] Abnormal termination on Linux

Thanks, Kurt. I think the reason your signal trap didn’t work is because `mpiexec` is trapping it first. Segfault is code error. Why would you want to trap it?

-- Hui Zhou

From: "Mccall, Kurt E. (MSFC-EV41)" <[email protected]<mailto:[email protected]>> Date: Monday, April 6, 2020 at 1:52 PM To: "Zhou, Hui" <[email protected]<mailto:[email protected]>>, "[email protected]<mailto:[email protected]>" <[email protected]<mailto:[email protected]>> Subject: RE: [mpich-discuss] Abnormal termination on Linux

Hui,

Sorry for not mentioning that. MPICH 3.3.2 compiled with pgc++ 19.5.

Kurt

From: Zhou, Hui <[email protected]<mailto:[email protected]>> Sent: Monday, April 6, 2020 12:58 PM To: [email protected]<mailto:[email protected]> Cc: Mccall, Kurt E. (MSFC-EV41) <[email protected]<mailto:[email protected]>> Subject: [EXTERNAL] Re: [mpich-discuss] Abnormal termination on Linux

Which version of MPICH were you running?

-- Hui Zhou

From: "Mccall, Kurt E. (MSFC-EV41) via discuss" <[email protected]<mailto:[email protected]>> Reply-To: "[email protected]<mailto:[email protected]>" <[email protected]<mailto:[email protected]>> Date: Monday, April 6, 2020 at 12:45 PM To: "[email protected]<mailto:[email protected]>" <[email protected]<mailto:[email protected]>> Cc: "Mccall, Kurt E. (MSFC-EV41)" <[email protected]<mailto:[email protected]>> Subject: Re: [mpich-discuss] Abnormal termination on Linux

I should mention that I am unable to predict in which node or process the abnormal termination occurs, so I can’t practically attach a debugger and try to intercept the error.

Kurt

From: Mccall, Kurt E. (MSFC-EV41) <[email protected]<mailto:[email protected]>> Sent: Monday, April 6, 2020 11:50 AM To: [email protected]<mailto:[email protected]> Cc: Mccall, Kurt E. (MSFC-EV41) <[email protected]<mailto:[email protected]>> Subject: Abnormal termination on Linux

I have a couple of questions about abnormal termination. The EXIT CODE below is 11, which could be signal SIGSEGV, or is it something defined by MPICH? If it is SIGSEGV, it is strange because my signal handler isn’t catching it and cleaning up properly (the signal handler calls MPI_Finalize()). Is there any way to get more information about the location of the error?

= BAD TERMINATION OF ONE OF YOUR APPLICATION PROCESSES = PID 14385 RUNNING AT n020.cluster.com = EXIT CODE: 11 = CLEANING UP REMAINING PROCESSES = YOU CAN IGNORE THE BELOW CLEANUP MESSAGES

Thanks, Kurt

_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

-- Dipl.-Inf. Joachim Protze IT Center Group: High Performance Computing Division: Computational Science and Engineering RWTH Aachen University Seffenter Weg 23 D 52074 Aachen (Germany) Tel: +49 241 80- 24765 Fax: +49 241 80-624765 [email protected] www.itc.rwth-aachen.de