[ExM Users] Can not diagnose Swift/T script failure

Michael Wilde wilde at anl.gov
Wed Oct 15 16:31:50 CDT 2014


Status as of Monday (but I know Justin did a few commits since...)

-------- Forwarded Message --------
Subject: 	Re: Latest on the Swift/T Titan problem?
Date: 	Mon, 13 Oct 2014 16:11:31 -0500
From: 	Justin M Wozniak <wozniak at mcs.anl.gov>
To: 	Michael Wilde <wilde at anl.gov>



Pure shell: works
Pure Tcl: works
ADLB batcher: works
<------->
Swift/T: fails

Something in Turbine breaks the OS.  I have been bisecting over a
stripped-down Turbine to identify exactly what this is.  If I strip down
the Turbine link step, I can get it to work.  So I think it is something
in our use of Cray modules.  (I try to not use modules/softenv but it is
affecting the link.)

This may fix the Swift/NAMD issue, which would be a complementary
positive outcome.

On 10/13/2014 12:10 PM, Michael Wilde wrote:
> One avenue I was wondering about for debugging:  What PrgEnv is the
> Swift/T on Titan built for?
> Im assuming gnu, but if so, how does that interact with the
> PrgEnv-intel needed for the DISCUS app?
>
> I would think to handle such cases we need to run Turbine under the
> PrgEnv it needs, and then run any app( ) calls that needs a different
> PrgEnv under a wrapper shell script that provides it, just for that
> process.
>
> Which may in turn raise the issue of whether "module load
> PrgEnv-intel" for example is scalable to hundredes to tens of
> thousands of concurrent invocations.
>
> - Mike
>
> On 10/13/14 11:38 AM, Justin M Wozniak wrote:
>>
>> Yes, I am testing launching tasks with a plain shell script and a
>> plain Tcl script.  I will then move to the ADLB batcher program.
>>
>> On 10/13/2014 11:34 AM, Michael Wilde wrote:
>>> Hi Justin,
>>>
>>> I did not get to spend any time on this over the weekend.
>>>
>>> What is your latest assessment of the issue?
>>>
>>> I did some runs Fri night that were disturbing: two runs of 256
>>> sleeps on 128 cores, each yielding a different number of output
>>> files (but missing ~ 5 in one case and ~ 10 in the other).
>>>
>>> I was going to do a test of a simple MPI app forking tasks with
>>> system( ), to see if same failure pattern is seen or not, as #CPUs
>>> increases from 128 to 1024.
>>>
...

-- 
Justin M Wozniak


On 10/15/14 4:29 PM, Michael Wilde wrote:
> Justin, can you report on your latest findings on this?
>
> (I will fwd Justin's latest notes from Monday...)
>
> - Mike
>
> On 10/15/14 4:27 PM, Tim Armstrong wrote:
>> Did you reach a resolution on this?  I'm back in town now and could 
>> look at adding the ability to retrieve the exit code.
>>
>> - Tim
>>
>> On Fri, Oct 3, 2014 at 12:07 AM, Michael Wilde <wilde at anl.gov 
>> <mailto:wilde at anl.gov>> wrote:
>>
>>     Also: identical code seems to run fine at this setting:
>>       PPN=16 turbine-cray-run.zsh -s titan.settings -n 64 $tcl  # (4
>>     nodes)
>>     but fails at this setting:
>>       PPN=16 turbine-cray-run.zsh -s titan.settings -n 128 $tcl # (8
>>     nodes)
>>
>>     Changing the app that calls DISCUS directly back to a shell
>>     wrapper around DISCUS shows that when the job fails the app
>>     (shell wrapper) is indeed executed, and logs a message to the app
>>     stdout file which I see. But I dont see a message after DISCUS.
>>
>>     To test if perhaps many concurrent DISCUS invocations (127) are
>>     causing a problem which is fatal to the whole job, I added a
>>     random sleep before DISCUS.  This had no effect on the problem.
>>
>>     Next I eliminated the actual DISCUS call, and just did a random
>>     sleep from 0 to 9 seconds. Here I discovered that all the jobs
>>     with sleep > 0 failed in the same manner as with discus, while
>>     the sleep 0 cases all printed all their echos and exited.
>>
>>     So there is something more deeply wrong going on, having nothing
>>     to do with the DISCUS app. I'll try to narrow this down, and see
>>     if perhaps its happening on other non-Cray systems like Midway.
>>
>>     - Mike
>>
>>
>>     On 10/2/14 10:52 PM, Michael Wilde wrote:
>>>     Thanks, Tim. But, stderr and stdout were both redirected to
>>>     files, and all .err and .out files are empty.
>>>
>>>     I'll double-check that these redirects are working, but in prior
>>>     runs, I did indeed get app error and output in those files.
>>>
>>>     It looks more to me like the app is failing to launch.
>>>
>>>     - Mike
>>>
>>>     On 10/2/14 10:22 PM, Tim Armstrong wrote:
>>>>     The app stderr/stdout should go to Swift/T stderr/stdout unless
>>>>     redirected. The problem is most likely discus returning a
>>>>     non-zero error code.
>>>>
>>>>     - Tim
>>>>
>>>>     On Thu, Oct 2, 2014 at 8:42 PM, Michael Wilde <wilde at anl.gov
>>>>     <mailto:wilde at anl.gov>> wrote:
>>>>
>>>>
>>>>         Im getting a failure from a Swift/T run on Titan that I
>>>>         cant diagnose.
>>>>
>>>>         The script is running 16K DISCUS simulations, about 30 secs
>>>>         each, 16 per node, on 8 nodes. 127 workers, 1 Server. It
>>>>         seems to run well at smaller scale (eg 1K simulations).
>>>>
>>>>         My PBS output file shows that some of the initial round of
>>>>         simulations fail.  It seems that when this happens, the
>>>>         Swift script exits within the first round of simulations,
>>>>         and none seem to even start.
>>>>
>>>>         Some of the logs are below.  I'll continue to experiment
>>>>         and add more logging to try to isolate the cause.
>>>>
>>>>         I was initially suspicious of an OOM problem, but I dont
>>>>         see any sign of that.
>>>>
>>>>         In the line below "Swift: external command failed:
>>>>         /lustre/atlas/proj-shared/mat049/scu/build/DiffuseCode/titan/discus/prog/discus
>>>>         --macro ../stacking-faults3.mac 0.004 0.264 1 0.75
>>>>         ../Si.cll 0.004_0.264.inte 0.004_0.264.stru" - is there a
>>>>         way to tell what Swift things the exit status of these
>>>>         failing apps are?  What is causing the whole run to abort?
>>>>
>>>>         It almost looks like one app, or a small number of them,
>>>>         are generating a SIGABRT, but thats not clear, and I have
>>>>         no stdout/err info from the app processes.
>>>>
>>>>         I'll look for some more detailed debug options, and/or add
>>>>         some wrappers around the apps.
>>>>
>>>>         Thanks,
>>>>
>>>>         - Mike
>>>>
>>>>         ---
>>>>
>>>>         T$ cat output.txt.2115044.out
>>>>
>>>>         Swift: external command failed:
>>>>         /lustre/atlas/proj-shared/mat049/scu/build/DiffuseCode/titan/discus/prog/discus
>>>>         --macro ../stacking-faults3.mac 0.004 0.264 1 0.75
>>>>         ../Si.cll 0.004_0.264.inte 0.004_0.264.stru
>>>>
>>>>         Swift: killing MPI job...
>>>>         ADLB_Abort(1)
>>>>         MPI_Abort(1)
>>>>
>>>>         Swift: external command failed:
>>>>         /lustre/atlas/proj-shared/mat049/scu/build/DiffuseCode/titan/discus/prog/discus
>>>>         --macro ../stacking-faults3.mac 0.004 0.272 1 0.75
>>>>         ../Si.cll 0.004_0.272.inte 0.004_0.272.stru
>>>>
>>>>         Swift: killing MPI job...
>>>>         ADLB_Abort(1)
>>>>         MPI_Abort(1)
>>>>
>>>>         Swift: external command failed:
>>>>         /lustre/atlas/proj-shared/mat049/scu/build/DiffuseCode/titan/discus/prog/discus
>>>>         --macro ../stacking-faults3.mac 0.26 0.052000000000000005 1
>>>>         0.75 ../Si.cll 0.260_0.052.inte 0.260_0.052.stru
>>>>
>>>>         Swift: killing MPI job...
>>>>         ADLB_Abort(1)
>>>>         MPI_Abort(1)
>>>>
>>>>         Swift: external command failed:
>>>>         /lustre/atlas/proj-shared/mat049/scu/build/DiffuseCode/titan/discus/prog/discus
>>>>         --macro ../stacking-faults3.mac 0.004 0.084 1 0.75
>>>>         ../Si.cll 0.004_0.084.inte 0.004_0.084.stru
>>>>
>>>>         Swift: killing MPI job...
>>>>         ADLB_Abort(1)
>>>>         MPI_Abort(1)
>>>>
>>>>         Swift: external command failed:
>>>>         /lustre/atlas/proj-shared/mat049/scu/build/DiffuseCode/titan/discus/prog/discus
>>>>         --macro ../stacking-faults3.mac 0.26 0.024 1 0.75 ../Si.cll
>>>>         0.260_0.024.inte 0.260_0.024.stru
>>>>
>>>>         Swift: killing MPI job...
>>>>         ADLB_Abort(1)
>>>>         MPI_Abort(1)
>>>>
>>>>         Swift: external command failed:
>>>>         /lustre/atlas/proj-shared/mat049/scu/build/DiffuseCode/titan/discus/prog/discus
>>>>         --macro ../stacking-faults3.mac 0.004 0.056 1 0.75
>>>>         ../Si.cll 0.004_0.056.inte 0.004_0.056.stru
>>>>
>>>>         Swift: killing MPI job...
>>>>         ADLB_Abort(1)
>>>>         MPI_Abort(1)
>>>>
>>>>         Swift: external command failed:
>>>>         /lustre/atlas/proj-shared/mat049/scu/build/DiffuseCode/titan/discus/prog/discus
>>>>         --macro ../stacking-faults3.mac 0.004 0.364 1 0.75
>>>>         ../Si.cll 0.004_0.364.inte 0.004_0.364.stru
>>>>
>>>>         Swift: killing MPI job...
>>>>         ADLB_Abort(1)
>>>>         MPI_Abort(1)
>>>>
>>>>         Swift: external command failed:
>>>>         /lustre/atlas/proj-shared/mat049/scu/build/DiffuseCode/titan/discus/prog/discus
>>>>         --macro ../stacking-faults3.mac 0.004 0.07200000000000001 1
>>>>         0.75 ../Si.cll 0.004_0.072.inte 0.004_0.072.stru
>>>>
>>>>         Swift: killing MPI job...
>>>>         ADLB_Abort(1)
>>>>         MPI_Abort(1)
>>>>
>>>>         Swift: external command failed:
>>>>         /lustre/atlas/proj-shared/mat049/scu/build/DiffuseCode/titan/discus/prog/discus
>>>>         --macro ../stacking-faults3.mac 0.26 0.044 1 0.75 ../Si.cll
>>>>         0.260_0.044.inte 0.260_0.044.stru
>>>>
>>>>         Swift: killing MPI job...
>>>>         ADLB_Abort(1)
>>>>         MPI_Abort(1)
>>>>         Application 7706378 exit codes: 134
>>>>         Application 7706378 exit signals: Killed
>>>>         Application 7706378 resources: utime ~5s, stime ~19s, Rss
>>>>         ~11016, inblocks ~48986, outblocks ~205
>>>>         T$
>>>>
>>>>         ---
>>>>
>>>>         T$ cat output.txt
>>>>         bash_profile: loading modules
>>>>         DB /ccs/home/wildemj/.modules: loading modules...
>>>>         Turbine: turbine-aprun.sh
>>>>         10/02/2014 09:15PM
>>>>
>>>>         TURBINE_HOME:
>>>>         /lustre/atlas2/mat049/proj-shared/sfw/compute/turbine
>>>>         SCRIPT:
>>>>         /lustre/atlas2/mat049/proj-shared/wildemj/swift-discus/stacking-faults3.tcl
>>>>         PROCS:        128
>>>>         NODES:        8
>>>>         PPN:          16
>>>>         WALLTIME:     00:15:00
>>>>
>>>>         TURBINE_WORKERS: 127
>>>>         ADLB_SERVERS:    1
>>>>
>>>>         TCLSH:
>>>>         /lustre/atlas2/mat049/proj-shared/sfw/tcl-8.6.2/bin/tclsh8.6
>>>>
>>>>         JOB OUTPUT:
>>>>
>>>>         Rank 2 [Thu Oct  2 21:15:32 2014] [c12-0c1s1n3] application
>>>>         called MPI_Abort(MPI_COMM_WORLD, 1) - process 2
>>>>         Rank 4 [Thu Oct  2 21:15:32 2014] [c14-0c1s1n3] application
>>>>         called MPI_Abort(MPI_COMM_WORLD, 1) - process 4
>>>>         Rank 6 [Thu Oct  2 21:15:32 2014] [c14-0c1s1n1] application
>>>>         called MPI_Abort(MPI_COMM_WORLD, 1) - process 6
>>>>         Rank 3 [Thu Oct  2 21:15:32 2014] [c14-0c1s1n2] application
>>>>         called MPI_Abort(MPI_COMM_WORLD, 1) - process 3
>>>>         Rank 5 [Thu Oct  2 21:15:32 2014] [c14-0c1s1n0] application
>>>>         called MPI_Abort(MPI_COMM_WORLD, 1) - process 5
>>>>         Rank 1 [Thu Oct  2 21:15:32 2014] [c12-0c1s1n2] application
>>>>         called MPI_Abort(MPI_COMM_WORLD, 1) - process 1
>>>>         Rank 7 [Thu Oct  2 21:15:32 2014] [c14-0c1s0n0] application
>>>>         called MPI_Abort(MPI_COMM_WORLD, 1) - process 7
>>>>         _pmiu_daemon(SIGCHLD): [NID 06114] [c14-0c1s1n0] [Thu Oct 
>>>>         2 21:15:32 2014] PE RANK 5 exit signal Aborted
>>>>         _pmiu_daemon(SIGCHLD): [NID 06076] [c14-0c1s1n2] [Thu Oct 
>>>>         2 21:15:32 2014] PE RANK 3 exit signal Aborted
>>>>         _pmiu_daemon(SIGCHLD): [NID 06077] [c14-0c1s1n3] [Thu Oct 
>>>>         2 21:15:32 2014] PE RANK 4 exit signal Aborted
>>>>         _pmiu_daemon(SIGCHLD): [NID 04675] [c12-0c1s1n3] [Thu Oct 
>>>>         2 21:15:32 2014] PE RANK 2 exit signal Aborted
>>>>         [NID 04675] 2014-10-02 21:15:32 Apid 7706378: initiated
>>>>         application termination
>>>>         _pmiu_daemon(SIGCHLD): [NID 06115] [c14-0c1s1n1] [Thu Oct 
>>>>         2 21:15:32 2014] PE RANK 6 exit signal Aborted
>>>>         T$
>>>>         T$
>>>>         T$
>>>>         T$ cat turbine.log
>>>>         JOB:               2115044
>>>>         COMMAND:           stacking-faults3.tcl
>>>>         HOSTNAME: ccs.ornl.gov <http://ccs.ornl.gov>
>>>>         SUBMITTED:         10/02/2014 09:14PM
>>>>         PROCS:             128
>>>>         PPN:               16
>>>>         NODES:             8
>>>>         TURBINE_WORKERS:
>>>>         ADLB_SERVERS:
>>>>         WALLTIME:          00:15:00
>>>>         ADLB_EXHAUST_TIME:
>>>>         T$
>>>>
>>>>
>>>>
>>>>
>>>>         -- 
>>>>         Michael Wilde
>>>>         Mathematics and Computer Science     Computation Institute
>>>>         Argonne National Laboratory    The University of Chicago
>>>>
>>>>         _______________________________________________
>>>>         ExM-user mailing list
>>>>         ExM-user at lists.mcs.anl.gov <mailto:ExM-user at lists.mcs.anl.gov>
>>>>         https://lists.mcs.anl.gov/mailman/listinfo/exm-user
>>>>
>>>>
>>>>
>>>>
>>>>     _______________________________________________
>>>>     ExM-user mailing list
>>>>     ExM-user at lists.mcs.anl.gov  <mailto:ExM-user at lists.mcs.anl.gov>
>>>>     https://lists.mcs.anl.gov/mailman/listinfo/exm-user
>>>
>>>     -- 
>>>     Michael Wilde
>>>     Mathematics and Computer Science          Computation Institute
>>>     Argonne National Laboratory               The University of Chicago
>>
>>     -- 
>>     Michael Wilde
>>     Mathematics and Computer Science          Computation Institute
>>     Argonne National Laboratory               The University of Chicago
>>
>>
>>     _______________________________________________
>>     ExM-user mailing list
>>     ExM-user at lists.mcs.anl.gov <mailto:ExM-user at lists.mcs.anl.gov>
>>     https://lists.mcs.anl.gov/mailman/listinfo/exm-user
>>
>>
>
> -- 
> Michael Wilde
> Mathematics and Computer Science          Computation Institute
> Argonne National Laboratory               The University of Chicago

-- 
Michael Wilde
Mathematics and Computer Science          Computation Institute
Argonne National Laboratory               The University of Chicago

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/exm-user/attachments/20141015/26f5b6af/attachment-0001.html>


More information about the ExM-user mailing list