[Darshan-users] Darshan crashing compute node

Thu Apr 2 11:29:12 CDT 2020

FYI, I opened a GitLab issue to track this problem: https://xgitlab.cels.anl.gov/darshan/darshan/issues/270

The crashes are a new surprise to us, so would like to have a look before our next release. I'll look to see if I can get a newer Lustre version working in a container or something to test against and will report back what I find.

Thanks,
--Shane
________________________________
From: Darshan-users <darshan-users-bounces at lists.mcs.anl.gov> on behalf of Carns, Philip H. <carns at mcs.anl.gov>
Sent: Thursday, April 2, 2020 10:27 AM
To: André R. Carneiro <andre.es at gmail.com>; darshan-users at lists.mcs.anl.gov <darshan-users at lists.mcs.anl.gov>
Subject: Re: [Darshan-users] Darshan crashing compute node

Thanks for the bug reports André!  You are describing two different problems:

1) incompatibility between the Darshan Lustre instrumentation and recent Lustre releases: we'll definitely look into this.  As you found in the mailing list archives there has been a known problem here, but this is a new permutation that it causes a node crash now.  That's disappointing, to say the least.  If anyone happens to have a pointer to how to stand up a Lustre 2.11 instance in docker or a VM they can share, that would be greatly appreciated and would help us track compatibility better.  The Lustre systems we have access to are not new enough. In the mean time you are correct to disable that component at build time.

2) floating point or divide by zero when writing log to Lustre: This one is odd because at the point of the problem, Darshan is not instrumenting anything: it is simply an application writing data to Lustre.  We don't generally see problems here.  Darshan is doing a collective write here as the last step of aggregating instrumentation from all ranks into a single log file.  The only slightly unusual things that Darshan could be doing are a) setting MPI-IO hints or b) using particular datatypes to organize the writes.  Neither should be a problem, but it looks like they might have triggered a bug somewhere.

If you don't mind trying another test case, could you repeat one of the experiments that crashes with a floating point or divide by zero bug down in MPI_File_write_at_all() with the following environment variable set?  (please keep disabling the Lustre module at build time, we have no other fix for that ready yet)

export DARSHAN_LOGHINTS=

(in other words, set the DARSHAN_LOGHINTS environment variable to an empty string)

I believe that this will clear the MPI-IO hints that Darshan would normally specify when opening the output log file.  If that makes your example work cleanly, then we can narrow down which hint is the problem and maybe get some help outside of Darshan for the root cause.  If it still fails, then we need to look elsewhere for the problem.

thanks,
-Phil
________________________________
From: Darshan-users <darshan-users-bounces at lists.mcs.anl.gov> on behalf of André R. Carneiro <andre.es at gmail.com>
Sent: Wednesday, April 1, 2020 1:39 PM
To: darshan-users at lists.mcs.anl.gov <darshan-users at lists.mcs.anl.gov>
Subject: [Darshan-users] Darshan crashing compute node

Hi,

I'm traying to use darshan (tested all 3.1.X versions) on a HPC cluster running with Red Hat 7.6 (kernel 3.10.0-957.el7.x86_64) and using Lustre Filesystem version 2.11.0.300_cray_102_g3dbace1, but when the application starts, the compute node crashes with a kernel panic message on the console.

While searching on the Darshan mailing list I came across this message: https://lists.mcs.anl.gov/mailman/htdig/darshan-users/2019-October/000542.html

I disabled the Lustre module and recompiled darshan. The crash isn't happening anymore, but now it's unable to write the darshan log file on the lustre filesystem, with the following error:

*Using OpenMPI 3.1.5 and GCC 7

Program received signal SIGFPE: Floating-point exception - erroneous arithmetic operation.
Backtrace for this error:
#0  0x7f4f6759f27f in ???
#1  0x7f4f687ababe in ???
#2  0x7f4f687add06 in ???
#3  0x7f4f687db6c0 in ???
#4  0x7f4f687dbddb in ???
#5  0x7f4f6879d6f1 in ???
#6  0x7f4f6871892b in ???
#7  0x7f4f691d0ae1 in MPI_File_write_at_all
at lib/darshan-mpiio.c:536
#8  0x7f4f691bea7f in darshan_log_append_all
at lib/darshan-core.c:1800
#9  0x7f4f691c1907 in darshan_log_write_name_record_hash
at lib/darshan-core.c:1761
#10  0x7f4f691c1907 in darshan_core_shutdown
at lib/darshan-core.c:546
#11  0x7f4f691be402 in MPI_Finalize
at lib/darshan-core-init-finalize.c:82
#12  0x7f4f68b6a798 in ???
#13  0x4023bb in ???
#14  0x401ae6 in ???
#15  0x7f4f6758b3d4 in ???
#16  0x401b16 in ???
#17  0xffffffffffffffff in ???
--------------------------------------------------------------------------

*Using Intel PSXE 2018 with Intel MPI

forrtl: severe (71): integer divide by zero
Image              PC                Routine            Line        Source
exec.exe           000000000045282E  Unknown               Unknown  Unknown
libpthread-2.17.s  00002B8B5A5FE5D0  Unknown               Unknown  Unknown
libmpi_lustre.so.  00002B8B659D4FDF  ADIOI_LUSTRE_Get_     Unknown  Unknown
libmpi_lustre.so.  00002B8B659CFFD9  ADIOI_LUSTRE_Writ     Unknown  Unknown
libmpi.so.12.0     00002B8B59A4C15C  Unknown               Unknown  Unknown
libmpi.so.12       00002B8B59A4D1D5  PMPI_File_write_a     Unknown  Unknown
libdarshan.so      00002B8B58F90312  MPI_File_write_at     Unknown  Unknown
libdarshan.so      00002B8B58F7E63A  Unknown               Unknown  Unknown
libdarshan.so      00002B8B58F815B0  darshan_core_shut     Unknown  Unknown
libdarshan.so      00002B8B58F7DFF3  MPI_Finalize          Unknown  Unknown
libmpifort.so.12.  00002B8B592414DA  pmpi_finalize__       Unknown  Unknown
exec.exe           00000000004490A5  Unknown               Unknown  Unknown
exec.exe           00000000004032DE  Unknown               Unknown  Unknown
libc-2.17.so<http://libc-2.17.so>       00002B8B5AB2F3D5  __libc_start_main     Unknown  Unknown
exec.exe           00000000004031E9  Unknown               Unknown  Unknown

If I configure darshan to write on a different filesystem (local /tmp of the first compute node), it works out fine, but then I'm restricted to use only one compute node, since the output dir have to be shared among all nodes (MPI tasks).

Is there a workaround for this? At the moment, my cluster only have the lustre filesystem as a shared filesystem among all compute nodes.

Best regards.

--
Abraços³,
André Ramos Carneiro.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/darshan-users/attachments/20200402/019db368/attachment.html>