Segfault for ODD+Pythia8+Geant4

[andiwand]

@benjaminhuth pointed out that ODD+Pythia8+Geant4 will segfault in full chain

I just verified this. See attached files for more information.

segfault.txt full_chain.txt

[paulgessinger]

Hm, could be the two interfere?

[asalzburger]

Is this executed in single thread ?

[asalzburger]

No good numThreads=-1 - this would need Geant4MT.

[benjaminhuth]

For me this crashes with one thread as well

[asalzburger]

Is that with the same error / segfault ?

[benjaminhuth]

Hmm it looks a bit different to be honest, but couldn't check in detail for no. I attache my gdb backtrace, maybe this gives a hint.

backtrace.txt

[benjaminhuth]

@andiwand could you maybe also run it in gdb to see if it is the same fault (its not entirely clear to me from the error message)

[andiwand]

Hm, could be the two interfere?

I think so yes. Pythia8 only works and Geant4 only works.

@andiwand could you maybe also run it in gdb to see if it is the same fault (its not entirely clear to me from the error message)

sure will do

[Corentin-Allaire]

I can confirm I just encountered the same issue using the geant4.py example : ### CAUGHT SIGNAL: 11 ### address: 0x7f9417827000, signal = SIGSEGV, value = 11, description = segmentation violation. Address not mapped to object. I tried updating to the latest G4 version (11.0.3) but it didn't change anything

[Corentin-Allaire]

As we discussed during today's meeting I tried to replace the ODD by the GDLM implementation of Alice_v3 and it ran through with just a few warning. So the issue is either with the ODD itself or the DDG4DetectorConstruction...

[stale[bot]]

This issue/PR has been automatically marked as stale because it has not had recent activity. The stale label will be removed if any interaction occurs.

[Corentin-Allaire]

I have had an other look at this and I just notice something. If I start removing the support from the ODD xml the segfault happen much later so maybe there is something bad with the support surface definition ?

[stale[bot]]

This issue/PR has been automatically marked as stale because it has not had recent activity. The stale label will be removed if any interaction occurs.

[Corentin-Allaire]

I was checking back this issue out of curiosity and it is still there. Maybe we should try to investigate this again at some point ?

[paulgessinger]

For sure this is something that we need to fix. Do we have a script to reproduce this?

[benjaminhuth]

Okay, I have investigated this a bit and some new infos:

First of all, I enabled some logging facilities in Geant4, which gave me the result that this is caused by photons quite far away from the center in z direction (z is around 1e4):

This is reproducible in pythia also with different seeds. Then I also could reproduce the crash with the ParticleGun:

addParticleGun(
    s,
    MomentumConfig(0.1 * u.GeV, 2.0 * u.GeV, transverse=True),
    EtaConfig(-4.0, 4.0, uniform=True),
    ParticleConfig(2, acts.PdgParticle.eGamma),
    vtxGen=acts.examples.GaussianVertexGenerator(
        stddev=acts.Vector4(10 * u.mm, 10 * u.mm, 10 * u.mm, 0.0 * u.ns),
        mean=acts.Vector4(18, 3.78, 1.09e4, 0),
    ),
    multiplicity=100,
    rnd=rnd,
)

I'm not totally sure what to do with these information, but maybe someone has an idea :)

[paulgessinger]

So it's G4 breaking in a specific region of the detector?

[Corentin-Allaire]

Wait the energy goes to 0 in the second step. Could it be that G4 doesn't handle photon stopping in some volumes ?

[benjaminhuth]

Wait the energy goes to 0 in the second step. Could it be that G4 doesn't handle photon stopping in some volumes ?

No I think with the electron in the pixel endcap is everythin fine, the photon below is the problem. There it only loggs the 0th step and then segfaults.

I could imagine that a problem is that it starts already outside of the detector (in the world_volume_1)?

Could it be that the world volume is to small or something like that?

[Corentin-Allaire]

Oh yeah I was looking at the wrong line... But you are right, the world volume size is 10m along z so this photon is outside the DD4Hep detector.

[Corentin-Allaire]

Unfortunately, I don't think this is the only issue :( I tried to edit the particle selector to remove all particle with x, y or z larger than 5m (in abs) and it still crashes with ttbar. How did you get those extra log Benjamin ?

[benjaminhuth]

Allready merged: https://github.com/acts-project/acts/pull/1790 With a new build from main branch you should be able to enable it via setting the logLevel to VERBOSE in the addGeant4 function.

[Corentin-Allaire]

Oh perfect I will have a look next week in more detail then !

[benjaminhuth]

Unfortunately, I don't think this is the only issue :( I tried to edit the particle selector to remove all particle with x, y or z larger than 5m (in abs) and it still crashes with ttbar. How did you get those extra log Benjamin ?

Actually I was able to run one event in the pythia8+geant4+ODD combination without segfault by increasing the world volumen manually from 10m to 100m in the ODD xml files...

I'm not sure if something like that would be a reasonable fix? Has this any other implications @asalzburger ?

I will try to run more events now, however, they take quite a long time (around 30 minutes per event)

[benjaminhuth]

I will try to run more events now, however, they take quite a long time (around 30 minutes per event)

Okay, actually it does not resolve the issue, I still get the segfault in a later event. Maybe it has just changed the random numbers a bit so that 1 event went through.

[Corentin-Allaire]

A bit unrelated but there is a bug in 'addGeant4' in 'simulation.py'. Line 597 it uses particles_input for the G4 input (instead of particles_selected) ignoring the particle selector. I can open a quick MR to fix this

[Corentin-Allaire]

A bit unrelated but there is a bug in 'addGeant4' in 'simulation.py'. Line 597 it uses particles_input for the G4 input (instead of particles_selected) ignoring the particle selector. I can open a quick MR to fix this

If someone wants to have a look : https://github.com/acts-project/acts/pull/1792

[Corentin-Allaire]

With this you can cut the particle outside the detector by adding preselectParticles = ParticleSelectorConfig(eta=(-3.0, 3.0),absZ=(0, 1e4), pt=(150 * u.MeV, None), removeNeutral=True), to the addGeant4. Doesn't solve the segfault in the ttbar case (but solve the photon issue).

[Corentin-Allaire]

Actually the code seem to be running on my side and doesn't segfault anymore... Can someone else confirm ?

[andiwand]

@Corentin-Allaire are you using the chain from above? otherwise if you could share the script I can try to verify

[Corentin-Allaire]

@andiwand here is the chain I use : full_chain_odd.txt

[andiwand]

this is still segfaulting for me on 7a3761d2b3f35c802bc03622d3b55fc9d463e426

segfault.txt

[Corentin-Allaire]

I might have changed something else by accident let me check (maybe you can also run it with verbose log of G4 ?)

[benjaminhuth]

Actually, for me it worked now at least for two events without segfault (I only applied the z-selection, not the pt or eta ones). Thats quite good news.

However, I got the following interesting warning:

Maybe I should add I had a timestamp-based seed, not the usual 42.

[andiwand]

did you run geant4 in verbose mode @benjaminhuth ? somehow it runs now for a couple of minutes without crashing

[benjaminhuth]

nope, at least that last one not. But I think we must consider the verbose mode to be extreeemly slow due to this huge printouts...

[andiwand]

hm for me it is still crashing even with 2 events

this is my geant version:

**************************************************************
 Geant4 version Name: geant4-11-00-patch-01 [MT]   (8-March-2022)
                       Copyright : Geant4 Collaboration
                      References : NIM A 506 (2003), 250-303
                                 : IEEE-TNS 53 (2006), 270-278
                                 : NIM A 835 (2016), 186-225
                             WWW : http://geant4.org/
**************************************************************

but feel free to close the ticket since it works for both of you now

[Corentin-Allaire]

@andiwand did you keep the number of thread to 1 ?

[Corentin-Allaire]

I realised I modified the preselection to cut on x,y and z so I removed that change but it still works on 5 events for now. I do have a slightly more recent G4 version :

 Geant4 version Name: geant4-11-00-patch-03 [MT]   (16-September-2022)
                       Copyright : Geant4 Collaboration
                      References : NIM A 506 (2003), 250-303
                                 : IEEE-TNS 53 (2006), 270-278
                                 : NIM A 835 (2016), 186-225
                             WWW : http://geant4.org/
**************************************************************```

[andiwand]

yeah exactly I did't modify the script. just executed the one you sent here https://github.com/acts-project/acts/issues/1578#issuecomment-1401645240

[benjaminhuth]

So I have geant4-11-00-patch-03 as well, and for me it also works now with 5 events. Maybe its indeed the geant version?

[andiwand]

let me update and check again

[Corentin-Allaire]

I remember reading there was some issue with patch 01 witch is why I updated the first time I run into the odd segfault

[Corentin-Allaire]

Let me also open and MR to solve to let people use G4 with the Odd

[benjaminhuth]

Actually I think this is a good workaround, though not 100% satisfying...

I just wonder if the change in the full_chain_odd.py is enough to kind of document this for others?

[Corentin-Allaire]

In my opinion this is a pythia issue and not an acts one

[Corentin-Allaire]

I have opened a PR : https://github.com/acts-project/acts/pull/1794. I am mentioning this issue in the comment of the code if people want to understand the full story

[andiwand]

I recompiled G4 with 11.1.0 and it seems not to crash anymore

[Corentin-Allaire]

I am in the process of running on 100 events just to check that we don't get particles with X or Y > 10m

[Corentin-Allaire]

I had the simulation run for 100 ttbar events and no issue occurred. I think we can close this one then ! As a summary in case someone comes back here :

• Pythia sometime generate particles outside the detector world volume (in particular in Z), using a ParticleSelector can resolve the issue. If the issue was to come back, we might need to extend the ParticleSelector to also cut in X and Y.
• G4 simulation cannot be run in multithreaded mode in Acts, doing so will result in crashes.
• G4 version anterior to geant4-11-00-patch-03 might also result is segfault when running the simulation (has been shown for version geant4-11-00-patch-01 at least). In case of crashed in G4 try to upgrade to the latest version.